Paper deep dive
The AI Transformation Gap Index (AITG): An Empirical Framework for Measuring AI Transformation Opportunity, Disruption Risk, and Value Creation at the Industry and Firm Level
Dean Barr
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%
Last extracted: 7/20/2026, 7:21:24 AM
Summary
The paper introduces the AI Transformation Gap Index (AITG), an empirical framework for measuring the distance between a firm's current AI deployment and a time-varying, industry-constrained capability frontier. The framework consists of five modules: Industry AI Susceptibility Score (IASS), AI Frontier Coefficient (AFC), Implementation Feasibility Score (IFS), Value Creation Bridge (VCB), and AI Disruption Risk Index (ADRI). It maps these gaps to financial outcomes, execution feasibility, and disruption risk, calibrated for 22 industry verticals and applied to 14 public companies.
Entities (11)
Relation Signals (9)
Dean Barr → authored → AI Transformation Gap Index
confidence 99% · The AI Transformation Gap Index (AITG) ... Dean Barr Applied AI Researcher
AI Transformation Gap Index → includesmodule → AI Frontier Coefficient
confidence 95% · Five linked modules address this gap: ... a dynamic capability ceiling that evolves with frontier capabilities (AFC)
AI Transformation Gap Index → includesmodule → Implementation Feasibility Score
confidence 95% · Five linked modules address this gap: ... trajectory based firm scoring with integrated execution risk (IFS)
AI Transformation Gap Index → includesmodule → Value Creation Bridge
confidence 95% · Five linked modules address this gap: ... a CES bottleneck value decomposition mapping gap scores to enterprise value (VCB)
AI Transformation Gap Index → includesmodule → AI Disruption Risk Index
confidence 95% · Five linked modules address this gap: ... and a competitive hazard measure for inaction (ADRI).
AI Transformation Gap Index → includesmodule → Industry AI Susceptibility Score
confidence 95% · Five linked modules address this gap: cross industry normalization (IASS)
JPMorgan Chase → analyzedby → AI Transformation Gap Index
confidence 90% · Two companies, JPMorgan Chase and Zions Bancorporation, serve as primary depth cases illustrating within industry contrasts
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Despite the scale of capital being deployed toward AI initiatives, no empirical framework currently exists for benchmarking where a firm stands relative to competitors in AI readiness and deployment, or for translating that position into auditable financial outcomes. In practice, private equity deal teams, management consultants, and corporate strategists have relied on qualitative judgment and ad-hoc maturity labels; tools that are neither comparable across industries nor grounded in observable economic data. This paper introduces the AI Transformation Gap Index (AITG), a composite empirical framework that measures the distance between a firm's current AI deployment and a time varying, industry constrained capability frontier, then maps that distance to dollar denominated value creation, execution feasibility under uncertainty, and competitive disruption risk. Five linked modules address this gap: cross industry normalization (IASS), a dynamic capability ceiling that evolves with frontier capabilities (AFC), trajectory based firm scoring with integrated execution risk (IFS), a CES bottleneck value decomposition mapping gap scores to enterprise value (VCB), and a competitive hazard measure for inaction (ADRI). I calibrate the framework for 22 industry verticals and apply it to 14 public companies using public filings. A retrospective construct validity exercise correlating AITG scores with observed EBITDA margin expansion yields Spearman rho_s = 0.818 (n = 10), directionally consistent with predictions though insufficient for causal identification. A counterintuitive result emerges: the largest AI transformation gaps do not produce the highest value density, because implementation friction, CES bottlenecks, and timing lags erode the theoretical upside of wide gaps.
Tags
Links
- Source: https://arxiv.org/abs/2603.13278v1
- Canonical: https://arxiv.org/abs/2603.13278v1
Trouble viewing inline? Open PDF directly →
Full Text
268,526 characters extracted from source content.
Expand or collapse full text
The AI Transformation Gap Index (AITG) An Empirical Framework for Measuring AI Transformation Opportunity, Disruption Risk, and Value Creation at the Industry and Firm Level Dean Barr Applied AI Researcher dean@dsconsult.ai The AI Transformation Gap Index (AITG), Industry AI Susceptibility Score (IASS), Value Creation Bridge (VCB), Implementation Feasibility Score (IFS), AI Disruption Risk Index (ADRI), and AITG Value Density are original constructs introduced in this paper. (February 2026) Abstract Despite the scale of capital being deployed toward AI initiatives, no empirical framework currently exists for benchmarking where a firm stands relative to competitors in AI readiness and deployment, or for translating that position into auditable financial outcomes. In practice, private equity deal teams, management consultants, and corporate strategists have relied on qualitative judgment and ad hoc maturity labels; tools that are neither comparable across industries nor grounded in observable economic data. This paper introduces the AI Transformation Gap Index (AITG), a composite empirical framework that measures the distance between a firm’s current AI deployment and a time varying, industry constrained capability frontier, then maps that distance to dollar denominated value creation, execution feasibility under uncertainty, and competitive disruption risk. Five linked modules address this gap: cross industry normalization (IASS), a dynamic capability ceiling that evolves with frontier capabilities (AFC), trajectory based firm scoring with integrated execution risk (IFS), a CES bottleneck value decomposition mapping gap scores to enterprise value (VCB), and a competitive hazard measure for inaction (ADRI). I calibrate the framework for 22 industry verticals and apply it to 14 public companies using public filings. A retrospective construct validity exercise correlating AITG scores with observed EBITDA margin expansion yields Spearman ρs=0.818 _s=0.818 (n=10n=10), directionally consistent with predictions though insufficient for causal identification. A counterintuitive result emerges: the largest AI transformation gaps do not produce the highest value density, because implementation friction, CES bottlenecks, and timing lags erode the theoretical upside of wide gaps. Keywords: AI capability frontier; decision support under uncertainty; AI deployment planning; composite indicators; firm level productivity; AI adoption; task based framework; disruption risk modeling; general purpose technology. JEL Codes: O33, G12, L10, J24, C43, L25, D81. A consolidated notation reference and parameter calibration register appear in Appendix A (Tables 15 and 16). 1 Introduction 1.1 Motivation Artificial intelligence is widely recognized as a general purpose technology with large but uneven firm level effects David (1990); Brynjolfsson et al. (2021). Yet the prevailing tools used to assess “AI maturity” remain difficult to interpret economically, and in practice consultant practitioners, private equity deal teams, and corporate strategists lack a rigorous, comparable method for estimating how much value a company or industry can unlock through AI transformation. Ordinal maturity labels are rarely comparable across industries and are not mapped to auditable changes in productivity, margins, or enterprise value. Two sources of measurement error are especially consequential. Cross sectionally, a maturity score assigned to a bank is not commensurate with the same score assigned to a construction firm because the feasible automation frontier differs structurally by industry. Temporally, the set of feasible AI capabilities changes rapidly, so a static ceiling assumption can become stale within an investment horizon. Together these limitations create predictable allocation errors. Investors capitalize narratives weakly grounded in measurable operational change while underweighting firms incurring near term complementary costs whose benefits arrive with delay, the productivity J curve Brynjolfsson et al. (2021). Brynjolfsson et al. (2021) demonstrate that GPTs require substantial complementary intangible investments before measured productivity rises; Brynjolfsson and Hitt (1996, 2000) documented the identical pattern for information technology. Meanwhile, Acemoglu (2024) offers a carefully bounded estimate of no more than 0.66 % TFP growth over ten years, a constraint I engage with directly in Section 11. I introduce the AI Transformation Gap Index (AITG) to address this measurement gap. Its five linked modules (Sections 3–8) jointly provide cross industry normalization, a time evolving capability ceiling, a dollar denominated value creation bridge, and a quantified competitive hazard score. All parameters are formally specified and anchored to publicly verifiable data sources. 1.2 Contributions I make five conceptual contributions and provide an illustrative empirical calibration. The framework combines three input classes: (i) publicly observed data (BLS O*NET/OEWS, EDGAR, Lightcast, Census BTOS, Stanford HAI AI Index), (i) rubric based author scored constructs whose interrater reliability is not yet established (Section 11.5), and (i) assumption driven parameters with sensitivity ranges reported in ESM Tables S3–S5. The 14 company application constitutes an illustrative calibration, not a statistically powered validation study. 1. Industry normalized ceiling (IASS). I construct the IASS from O*NET task structure, OEWS occupational weights, regulatory friction, and competitive diffusion dynamics O*NET (2024); BLS OEWS (2023). Unlike additive checklist frameworks (McKinsey, 2025; Gartner, 2023; Westerman et al., 2014), the IASS uses geometric aggregation Mazziotta and Pareto (2013); OECD (2008) with a Regulatory Friction Factor hard floor ψ(RFF)ψ(RFF), enforcing noncompensability. Calibrations for 22 industry verticals are precomputed from BLS, Lightcast, and SEC filings. 2. Time varying frontier (AFC). The AI Frontier Coefficient is a formally specified, time indexed multiplier that recalibrates the IASS ceiling as model capabilities advance (MMLU from 43% to 93% between 2020–2024 Stanford HAI (2025); METR task horizon doubling every seven months Kwa et al. (2025)). I anchor θi _i exclusively to O*NET Automatable Task Density, eliminating the circular specification that arises when frontier update rules conflate capability expansion with adoption breadth. 3. Trajectory based firm scoring with endogenous execution risk. Firm state is decomposed into six observable dimensions and mapped to a cascading three wave logistic trajectory Brynjolfsson et al. (2021); Rogers (2003). Organizational capacity and data readiness are endogenized into wave steepness and inflection timing Brynjolfsson and Hitt (2000); Gibbons and Henderson (2012), recovering the correct dynamic compounding structure. The formal inverse mapping uses a mandatory piecewise specification guarding against NaN crashes. Data nonrivalry (Jones and Tonetti, 2020) is operationalized as a Firm Scale Factor Φf _f. 4. Value decomposition with enabling constraints (VCB). The VCB decomposes the gap into seven value pools, gates each on enabling infrastructure via a low elasticity CES bottleneck aggregator (ρ=5ρ=5, σ=1/6σ=1/6), and applies a nonlinear capture function. I identify and correct four compounding errors in standard AI to EV conversions with correction magnitudes ranging from 2×2× to 25×25×. 5. Competitive hazard from delay (ADRI). The ADRI measures the competitive hazard from not acting, operationalized as an instantaneous hazard intensity. A firm with a wide gap and high ADRI faces an increasing hazard rate as diffusion compresses margins and organizational debt compounds. Illustrative calibration. I apply the framework to 14 public companies across eight industries using a tiered evidence protocol based on public filings (Section 9). These applications demonstrate internal consistency and measurement mechanics, not causal claims. Two companies, JPMorgan Chase and Zions Bancorporation, serve as primary depth cases illustrating within industry contrasts under identical ceiling conditions. A retrospective construct validity exercise using 2021 AITG scores and 2021–2023 realized EBITDA margin changes yields a Spearman ρs=0.818 _s=0.818 (n=10n=10 nonfinancial firms), directionally consistent with predictions but insufficient for causal identification given sample size and omitted variable concerns. ESM Part IV supplies structured application guidance and a companion Excel workbook implementing all modules for self assessment. Full reproducibility code, including Monte Carlo sensitivity analysis, stress tests, a 25 question survey scoring pipeline, and the Excel workbook builder, is available at https://github.com/deanbrr/aitg-framework. 1.3 Related Work The AITG integrates three literatures not previously connected in a single operational framework. The task based automation literature Autor et al. (2003); Autor and Dorn (2013); Acemoglu and Restrepo (2018, 2019a) provides the microeconomic foundation for structural differences in AI susceptibility; I operationalize it as the Cognitive Task Density dimension of the IASS. The composite indicator methodology OECD (2008); Nardo and Saisana (2005) provides the normalization and sensitivity analysis framework required for institutional credibility. The IT productivity and organizational complementarity literature Brynjolfsson and Hitt (1996, 2000); Brynjolfsson and McElheran (2016); Brynjolfsson et al. (2021); Gibbons and Henderson (2012) provides the theoretical basis for why AI investment to returns relationships are delayed, nonlinear, and conditioned on organizational readiness. Unlike Bass (1969) single event diffusion models, the AITG three wave cascading logistic decomposes adoption by sequential technology wave rather than by innovation versus imitation. An alternative Cox (1972) proportional hazards formulation would complement the ADRI hazard intensity (Eq. 20) by providing a semiparametric baseline hazard; I chose cascading logistics for wave specific modularity, since each wave has independent kwk_w and t0,wt_0,w parameters updatable as deployment evidence accumulates. Unlike deployment aware optimization models such as ML Compass Digalakis et al. (2025), which use KKT based constrained optimization at the system level, the AITG operates at the firm and industry level, addressing cross industry normalization, time evolving ceilings, and competitive disruption risk. The KKT intuition of binding constraints is operationalized through the CES bottleneck aggregator: value capture is gated by the minimum weighted harmonic of enabling dimensions, not their average. Data marketplace pricing models (Jones and Tonetti, 2020) provide a theoretical foundation for the VCB Data Monetization pool: the nonrivalry of data implies marginal value increasing with reuse, captured through the concave function η(g)η(g). Empirical evaluation of frontier AI agents on autonomous task benchmarks Kinniment et al. (2024) documents conservative production level autonomy, motivating the Wave 3 steepness discount (kw=3k_w=3 set 30 % below Wave 2) and the IFS δOCC _OCC penalty for firms lacking agentic readiness. Section 2 establishes theoretical foundations; Sections 3–8 develop the five modules; Section 9 provides empirical illustrations; Section 10 reports sensitivity analysis; Section 11 addresses limitations; and Section 12 concludes. 2 Theoretical Foundations 2.1 The AI Transformed Frontier as an Industry Constrained Optimum The central construct is the frontier: the best operating configuration a firm in a given industry could achieve with current AI capabilities. Definition 2.1 (AI Transformed Frontier). For firm f in industry i at time t, define the AI Transformed frontier state Si,t∗S^*_i,t as the operating configuration that maximizes risk adjusted enterprise value subject to: (i) process and physical constraints inherent to industry i; (i) binding regulatory requirements applicable at time t; (i) the data generation processes naturally available in industry i; and (iv) the AI capability set tA_t available at time t. Three deliberate choices are embedded here. First, the frontier is dynamic: as tA_t expands, Si,t∗S^*_i,t shifts, opening new gap even for firms that have not changed. Second, it is industry constrained: a construction firm at its frontier looks categorically different from a financial services firm at its frontier. Third, the objective is risk adjusted enterprise value, anchoring the framework to economics rather than engineering and avoiding the methodological mistake of defining perfection as “uses AI everywhere.” The distance between a firm’s current state and this frontier captures the investment opportunity. Definition 2.2 (AI Transformation Gap). The AI Transformation Gap for firm f at time t is: Δf,t≡d(Sf,t,Si,t∗) _f,t\;≡\;d\! (S_f,t,\;S^*_i,t ) (1) where d(⋅)d(\,·\,·\,) is a distance in a standardized capability space and Sf,tS_f,t is the firm’s current implementation state. Δf,t _f,t is a value creation signal, not a technology adoption signal, because Si,t∗S^*_i,t is anchored to economically material levers (cost, margin, working capital, cycle time). A large gap can be bullish, indicating a large addressable opportunity, or bearish, indicating structural inability to close it. I separate gap measurement from closure probability; frameworks that blend the two sacrifice interpretability. 2.2 AI Elasticity by Industry Industries differ dramatically in how much enterprise value responds to AI investment. I formalize this heterogeneity as follows. Definition 2.3 (AI Elasticity by Industry). AI Elasticity by Industry EiE_i is the marginal enterprise value response to AI transformation investment, holding scale constant: Ei≡∂(EV/Revenue)i∂(AI Capability Index)iE_i\;≡\; ∂\! (EV/Revenue )_i∂\! (AI Capability Index )_i (2) EiE_i is continuous, not binary. The task based literature Autor and Dorn (2013); Acemoglu and Restrepo (2018) establishes that automation suitability varies substantially within jobs and across tasks; industry level susceptibility is a distribution over task types, not a categorical label. The IASS operationalizes EiE_i through six weighted dimensions. 2.3 Connection to the Acemoglu–Restrepo Task Framework The AITG’s conceptual foundation maps directly onto the canonical task based model of Acemoglu and Restrepo (2018). Production uses tasks z∈[N−1,N]z∈[N-1,N]; tasks below threshold I are performed by capital (automated), tasks above I by labor. The production function is: Y=Π(I,N)[Γ(I,N)1/σ(ALL)(σ−1)/σ+(1−Γ)1/σ(AKK)(σ−1)/σ]σ/(σ−1)Y= (I,N)\! [ (I,N)^1/σ(A_LL)^(σ-1)/σ+(1- )^1/σ(A_KK)^(σ-1)/σ ]^σ/(σ-1) (3) where Γ(I,N) (I,N) is the labor task content of production. Automation (increasing I) always reduces Γ and labor share; new task creation (increasing N) always raises both. The displacement reinstatement decomposition Acemoglu and Restrepo (2019a) is: dlnΓ=−displacement effect⏟automation (↑I)+reinstatement effect⏟new tasks (↑N)d = -displacement effect_automation\;($ I$)+ reinstatement effect_new tasks\;($ N$) (4) The AITG’s Cognitive Task Density dimension (Eq. 38) operationalizes Γ(I,N) (I,N) empirically at the industry level. A firm’s AITG score corresponds to If/Ii∗I_f/I^*_i, the ratio of its current automation frontier to the industry’s achievable frontier. Moving this ratio toward 1.0 is the investment thesis. 2.4 The Productivity J Curve and Measurement Implications AITG scores do not immediately predict observable productivity gains. Brynjolfsson et al. (2021) demonstrate that general purpose technologies require complementary intangible investments: process redesign, workforce retraining, and organizational restructuring, all expensed rather than capitalized. Measured productivity declines initially; benefits emerge once the complementary capital is in place. Adjusting for unmeasured intangibles, the authors find true TFP 15.9 % higher than official measures by 2017. The lag can persist 20–30 years for major GPTs, consistent with David (1990) on the electrification delay. I calibrate the AITG Value Creation Bridge directly to this J curve logic: costs load in the first 12–18 months and value emerges in a logistic pattern over 24–42 months (Section 7.6). This also provides the definitive answer to “where is the ROI?”: it is there, but delayed by the complement accumulation process Brynjolfsson and Hitt (2000). 2.5 Winner Take Most Dynamics and Urgency Autor et al. (2020) document that across the U.S. economy, industry sales are concentrating and labor’s share is falling, driven by “superstar firms” that scale efficiently with technology. Their key empirical prediction is confirmed across all seven tests: concentrating industries show the largest labor share declines, fastest productivity growth, and highest markups. The mechanism is “scale without mass”: the most productive firms serve large markets with fewer workers. Furthermore, the nonrivalry of data Jones and Tonetti (2020) means that data rich AI adopters improve their models continuously, while laggards face a cold start disadvantage. This creates a compounding dynamic. I design the ADRI (Section 6) to score precisely this effect. The empirical regularity from Syverson (2011), a 90th/10th percentile TFP ratio of 1.92 within 4 digit SIC manufacturing industries, with low productivity firms exiting at higher rates, provides the empirical baseline for competitive displacement risk. 3 The Industry AI Susceptibility Score (IASS) What Is the IASS? Conceptual Summary The IASS answers one question before looking at any individual company: how much of this industry’s value creation is structurally capturable by AI? Think of it as the ceiling on the building before you invest in the top floor. A construction firm and a financial services firm can both have “great AI initiatives,” but they are playing in fundamentally different games. Construction workers build in unstructured physical environments with high variability and regulatory human sign off requirements. A commercial banker processes structured data, executes repeatable cognitive workflows, and operates in a digitized record environment. The AI ceiling for financial services is approximately 2×2× higher than for construction, not because financial firms are better managed, but because the task structure of the industry makes it structurally more susceptible to AI automation. Example. Investment Banking (IASS=∗10.39^*=10.39) vs. Construction (IASS =4.26=4.26). Both have competent management teams. Both can hire AI talent. But a dollar of AI investment in investment banking works against a structurally rich task environment: document heavy workflows, large structured transaction databases, high cognitive density decision processes. The same dollar in construction works against primarily physical, variable environment tasks that current AI cannot reliably automate. The IASS captures this structural difference before any company specific score is computed. Who should use this? An investment committee uses it to eliminate sectors below IASS 5.0 from the transformation investment universe. A corporate CAIO uses it to benchmark where their industry sits relative to peers and understand what the structural ceiling on their AI program actually is, not what their consultants promise. 3.1 Design Principles I construct the IASS as a composite indicator following OECD best practice methodology OECD (2008); Nardo and Saisana (2005). Three nonnegotiable design principles govern its construction: 1. Reproducibility. Any analyst following the documented procedures with the same source data must arrive within ±0.5± 0.5 points on the IASS composite. 2. Theoretical grounding. Each dimension must have an established causal mechanism linking it to AI enabled value creation. 3. Sensitivity transparency. Weights represent tradeoffs among dimensions, not “importance” coefficients OECD (2008). All weight choices are subjected to Monte Carlo robustness analysis (Section 10). 3.2 Dimensional Structure and Scoring Six observable dimensions constitute the IASS (Table 1), each scorable from public data without inside access. I weight DRSA and PRI more heavily given their documented roles as deployment bottlenecks (Standish Group, 2015), consistent with AI value creation complementarity (Autor, 2024; Acemoglu et al., 2022). Geometric aggregation (Eq. 5) enforces noncompensability: no substitutability across dimensions. Let wdw_d be the normalised weight of dimension d (summing to 1), and s~d,i s_d,i the min max normalised score for dimension d in industry i (Eq. 40). The base IASS is: IASSi=exp(∑d=1Dwdlns~d,i)IASS_i= \! ( _d=1^Dw_d s_d,i ) (5) I apply the Regulatory Friction Factor (ψ) as a hard floor multiplier outside the geometric product, ensuring that absolute regulatory prohibitions act as a noncompensable ceiling rather than a compensable dimension. Formally: ψi=min(1,∏jdfloor,j,i),IASSi=ψi⋅exp(∑dwdlns~d,i) _i= \! (1,\; _jd_floor,j,i ), _i= _i· \! ( _dw_d s_d,i ) (6) where dfloor,j,i∈(0,1]d_floor,j,i∈(0,1] are regulatory floor coefficients for each binding prohibition j in industry i. When ψi=1 _i=1 the industry faces no binding regulatory ceiling; when ψi<1 _i<1 the effective ceiling is suppressed proportionally. Healthcare Services (ψ=0.743ψ=0.743) and Life Sciences (ψ=0.663ψ=0.663) are the two industries with binding constraints in the current calibration. For single dimension implementation (e.g., the Excel Companion), the scalar approximation ψi=min(1,(RFFi5)1.5) _i= \! (1,\; ( RFF_i5 )^1.5 ) (7) where RFFi∈[0,10]RFF_i∈[0,10] is the raw Regulatory Friction Factor score, matches the multifloor product to within ±0.02± 0.02 for the current 22 industry calibration. Table 1: IASS: Six Dimensions, Sources, and Weights Dimension Definition and Mechanism Primary Source Weight Cognitive Task Density (CTD) Share of industry labor hours in cognitively automatable task Frey and Osborne (2017) s (routine cognitive + bounded fraction of nonroutine cognitive). Direct operationalization of Γ(I,N) (I,N) Acemoglu and Restrepo (2018); Acemoglu and Autor (2011). O*NET task statements O*NET (2024); Felten et al. (2021); Webb (2020) + BLS OEWS BLS OEWS (2023) Bonney et al. (2024) 0.25 Data Richness & Structural Availability (DRSA) Degree to which the industry generates structured, machine readable data as operational exhaust vs. requiring heavy instrumentation to create it. SEC EDGAR iXBRL SEC EDGAR (2024); tech stack surveys 0.20 Process Repeatability Index (PRI) Share of workflows that are high frequency and low variance (standardized, repeatable). Low task entropy industries are more automatable. Lightcast job posting natural language processing (NLP) Lightcast (2024) 0.20 Regulatory Friction Factor (RFF) Degree to which regulation constrains AI deployment. Splits into ceiling compression (prohibited uses) and time/cost friction (compliance overhead). Scored so 10 = no friction. EU AI Act EU AI Act (2024); HIPAA; FINRA 24-09; FDA GMLP 0.15 Competitive AI Diffusion Rate (CADR) Speed of AI diffusion into the competitive ecosystem. Measures urgency: high CADR shortens the window for undisturbed transformation. Lightcast AI/ML skill CAGR by NAICS Lightcast (2024); AlphaSense earnings NLP AlphaSense (2024) 0.10 Capital–Labor Substitutability (CLSR) Labor cost as fraction of total cost times AI substitutable fraction. Sets the dollar scale of the labor productivity value pool. BLS labor cost series; BEA capital accounts BEA (2023); Compustat Compustat (2024) 0.10 3.2.1 Sub Dimension Construction The six IASS dimensions are constructed from O*NET task statements, BLS occupational weights, Lightcast job posting frequencies, and regulatory source documents. Cognitive Task Density (CTD) uses employment weighted automatable task shares (Eq. 38); the Process Repeatability Index (PRI) combines task entropy with standardization language frequency. All dimensions are winsorized at the 5th/95th percentile and min max normalized to [0,10][0,10] (Eq. 40). Full formulas, intermediate definitions, and normalization remarks appear in Appendix B. 3.3 Sensitivity to Weighting Monte Carlo perturbation (M=10,000M=10,000 draws, ±5%± 5\,\% weight variation) confirms that no pair of anchor industries exchanges rank under any draw; the mean absolute rank shift is R¯s=0.19 R_s=0.19 positions. Full robustness bounds appear in Appendix C (Table 19). 4 The AI Frontier Coefficient (AFC) 4.1 Motivation: Nonstationarity of the AI Capability Ceiling Any assessment framework built against a static capability definition becomes systematically stale within 18–24 months. Table 2 documents representative capability trajectories for the seven benchmark domains that constitute CtC_t. Table 2: AI Capability Benchmark Progression (Selected Milestones) Benchmark Domain 2020 2024 2025–26 Source MMLU-Pro General knowledge 43.9 % 76.2 % 87.0 % Stanford HAI 2025 SWE-Bench Verified Software engineering 4.4 % 49.0 % 80.0 % OpenAI 2025b SWE-Bench Pro Multilang. coding n/a n/a 55.6 % OpenAI 2025b AIME 2025 Competition math n/a 71.0 % 100.0 % OpenAI 2025b FrontierMath (T1–3) Research math n/a n/a 40.3 % OpenAI 2025b GPQA Diamond Graduate science n/a 63.0 % 93.2 % OpenAI 2025b ARC-AGI-2 Abstract reasoning n/a 4.0 % 52.9 % OpenAI 2025b MMMU Multimodal understanding n/a 56.8 % 84.2 % OpenAI 2025a METR task horizon Agent autonomy 1 min 15 min ≥ 60 min Kwa et al. 2025 Implied CtC_t (EWMA composite) 1.00 1.62 1.90 2025–26 column reports GPT-5.2 Thinking (December 2025) scores except MMMU and METR (GPT-5, August 2025). n/a = benchmark not yet published at that date. Full CtC_t weights and methodology in ESM Part I. The AI Capability Index CtC_t has risen from C2020=1.00C_2020=1.00 (GPT-3 baseline) to C2024=1.62C_2024=1.62 (GPT-4o/o1) to C2025=1.74C_2025=1.74 (GPT-5, August 2025) to C2026≈1.90C_2026≈ 1.90 (GPT-5.2, December 2025). Full benchmark progression data, suite weights, and methodology are in ESM Part I. The METR time horizon analysis Kwa et al. (2025) shows the 50 % task completion horizon for AI agents doubling approximately every seven months (2019–2025), with a recent acceleration to approximately four months. These are not marginal improvements; they are category changes that open entirely new automation surface areas in industries previously considered AI resistant. A static IASS produces two compounding errors. First, a company scored as highly transformed in 2023 may underperform its true frontier by 2025 because new capabilities created automation opportunities that did not exist when the score was assigned. Second, industries historically penalized by low IASS scores (healthcare, legal) are experiencing rapid ceiling expansion as domain specific LLMs clear regulatory and quality thresholds, creating investment opportunities that a static framework structurally misses. 4.2 Formal Specification The AFC translates capability index movements into industry specific ceiling adjustments. I define it formally as follows. Definition 4.1 (AI Frontier Coefficient). For industry i at time t, the AI Frontier Coefficient is: AFCi,t=min(1+θi(Ct−C0),αmax)AFC_i,t= \! (1+ _i (C_t-C_0 ),\; _ ) (8) where CtC_t is the AI Capability Index at time t, C0=1.0C_0=1.0 is the reference year baseline (Q1 2024), θi∈[0.05, 1.50] _i∈[0.05,\,1.50] is an industry specific sensitivity parameter governing how rapidly a sector absorbs frontier gains, and αmax=1.35 _ =1.35 bounds the maximum ceiling expansion to prevent extrapolation beyond empirically supported ranges. Remark 4.1 (AFC Functional Form Justification). The linear (shift) form 1+θi(Ct−C0)1+ _i(C_t-C_0) is a first order Taylor approximation of a general adjustment function f(Ct/C0)f(C_t/C_0) around Ct=C0C_t=C_0. For the observed range of Ct∈[0.85,1.25]C_t∈[0.85,1.25] (one standard deviation of quarterly benchmark movements since Q1 2024), the linear approximation error is bounded by O(θi⋅(Ct−C0)2)<0.02O( _i·(C_t-C_0)^2)<0.02 for all industries. The αmax=1.35 _ =1.35 cap provides an additional safeguard against extrapolation beyond this range. Nonlinear alternatives, including a logistic saturation form αmax(1−e−θi(Ct−C0)) _ (1-e^- _i(C_t-C_0)) and a CES nested specification, are conceptually appealing for modeling diminishing returns to capability gains at the industry ceiling. However, calibrating such forms requires longer time series of benchmark adoption comovement than currently exist (the AFC observation window begins Q1 2024). Extending the AFC to nonlinear forms is a research agenda priority. The sensitivity of the AFC to its parameters is bounded: ∂AFC/∂θ=(Ct−C0) /∂θ=(C_t-C_0), which is at most 0.25 under current benchmark conditions. This implies that even large estimation errors in θi _i (e.g., ±50%± 50\%) produce AFC shifts of at most ±0.13± 0.13, well within the αmax _ cap. The effective (AFC adjusted) industry ceiling is: IASSi,t∗=IASSibase×AFCi,tIASS^*_i,t=IASS^base_i×AFC_i,t (9) 4.3 The AI Capability Index CtC_t I construct CtC_t as a weighted composite of normalized AI benchmark scores across seven domains covering the primary task types relevant to industry susceptibility. Let bk,t∈[0,1]b_k,t∈[0,1] be the best published score on benchmark k at time t: Ct=∑k=17ωk⋅bk,tC_t= _k=1^7 _k· b_k,t (10) The benchmark domain weights and EWMA smoothing methodology (λ=0.5λ=0.5, Eq. 42) are detailed in Appendix B (Table 17). From 2022 to 2024, CtC_t increased from 0.520 to 0.748 (raw), smoothed to 0.741 at end 2024, an implied annual improvement rate of approximately 21 % Epoch AI (2025); Kaplan et al. (2020); Hoffmann et al. (2022). 4.4 Industry Sensitivity Parameters θi _i The parameter θi _i captures how much of a general capability gain translates into an expansion of the economically addressable AI surface area in industry i. I calibrate θi _i retrospectively using the observable expansion in addressable AI tasks from 2020 (GPT-3 era) to 2024 (GPT-4o/o1 era) as assessed by independent domain experts: θ^i=ΔIASSiaddressableΔC2020→2024 θ_i= \,IASS^addressable_i C_2020→ 2024 (11) Healthcare scores θ=0.31θ=0.31 because domain LLMs (Nuance DAX, Google MedPaLM, Anthropic Claude in clinical workflows) created substantial new automatable surface area: ambient documentation, revenue cycle, prior authorizations, and diagnostics that simply did not exist in 2020. Construction scores θ=0.09θ=0.09 because model capability gains have not materially penetrated physically situated, high variance fieldwork. Vertical SaaS scores θ=0.11θ=0.11 because the ceiling was already high; the remaining gap is structurally small. 4.5 AFC Uncertainty and Scenario Analysis Because CtC_t is forward looking, AFC carries its own uncertainty quantification. Three scenarios (conservative, base case, aggressive) bound the AFC distribution over a 24 month horizon, with scenario weights 0.20/0.60/0.20 (Appendix B, Table 18). The AFC uncertainty contribution to the Uncertainty Quotient is UQafc=(AFCagg−AFCcon)/4UQ_afc=(AFC_agg-AFC_con)/4 (Eq. 43). Propagation of CtC_t and θi _i uncertainty into Δ . Uncertainty in CtC_t and θi _i propagates into ΔEV through two channels. The ceiling channel: IASSi,t∗=IASSibase×min(1+θi(Ct−C0),αmax)IASS^*_i,t=IASS^base_i× (1+ _i(C_t-C_0),\, _ ), so uncertainty in θi⋅(Ct−C0) _i·(C_t-C_0) translates directly to uncertainty in the feasible frontier. The gap channel: Geff=IASS∗−AITGrawG_eff=IASS^*-AITG^raw, which enters the value pool capture function as gf=Geff/10g_f=G_eff/10. The resulting Δ sensitivity is approximately: ∂ΔEV∂Ct≈θi⋅IASSibase⋅∂ΔEV∂IASS∗ ∂\, ∂\,C_t≈ _i·IASS^base_i· ∂\, ∂\,IASS^* (12) where I evaluate the second partial numerically via the Monte Carlo simulation in ESM Part I. In practical terms: for Healthcare (θi=0.31 _i=0.31, IASSbase=5.1IASS^base=5.1), a 10% upward revision in CtC_t expands the HC-IASS ceiling by approximately 0.16 pts and raises Δ by 3–8% depending on where the firm sits on the capture curve. For Construction (θi=0.09 _i=0.09), the same CtC_t revision moves the ceiling by only 0.05 pts and has negligible Δ impact. This asymmetry, where high θ industries are more sensitive to AFC revisions than low θ industries, is an intended property of the framework and is reported explicitly in the AFC scenario stress tests (ESM Table S3). Formal identification of θi _i standard errors and correlated uncertainty between CtC_t and θi _i are priorities in the research agenda (Section 11.5, item 3). Remark 4.2 (AFC as a Structural Moat Accelerator). A rising capability ceiling creates an asymmetric competitive dynamic. Consider the JPMorgan/Zions pair at early 2026 capability levels (C≈1.90C≈ 1.90, IASS∗ = 9.38): • JPMorgan (AITG 8.22) has an addressable gap of 1.16 to the new frontier. Its $2B/yr AI budget and proprietary data gravity (Φf=1.0 _f=1.0) allow it to chase the rising ceiling, creating new investable opportunity that did not exist at the 2022 baseline. • Zions (AITG 3.80) has a widened gap of 5.58, but vendor AI dependency caps its value pool at Φf≤0.65 _f≤ 0.65 regardless of how high the ceiling rises. The nominal gap widens; the capturable gap does not. At C2027≈2.49C_2027≈ 2.49 (three years at 20%/yr from 2024), Banking IASS∗ approaches 10.0. The gap ratio compresses, but this reflects JPMorgan chasing new frontier, not Zions closing ground. AFC mechanically advantages firms with proprietary data gravity and is effectively irrelevant for vendor dependent firms at subscale. 4.6 Evaluator Rotation Protocol: Mitigating Goodhart’s Law and Benchmark Drift Because CtC_t draws on public benchmark scores, it faces two failure modes: (1) benchmark saturation, where models reach near ceiling performance on a constituent (e.g., AIME 2025 at 100%), making further CtC_t gains undetectable; and (2) Goodhart contamination, where developers optimize against high stakes benchmarks, inflating CtC_t without real capability gains. I address both through a formal Evaluator Rotation Protocol: 1. Saturation trigger. Any benchmark exceeding 90% mean solve rate across the top five frontier models is deprecated from the CtC_t basket in the next annual cycle. 2. Replacement criterion. Replacements must be (a) out of distribution relative to existing members, (b) validated by an independent laboratory, and (c) correlated with real world task performance in a relevant IASS domain. METR time horizon benchmarks are the preferred anchor for agent capabilities. 3. Continuity chaining. At each rotation, pre and postrotation CtC_t values are computed on a 90 day overlap window and chained at the transition knot (analogous to CPI chain linking), preserving EWMA continuity. At the current calibration date, SWE-Bench Verified (80.0%) is approaching the saturation trigger; FrontierMath (Tier 1–3: 40.3%) and ARC-AGI-2 (52.9%) are the planned successors. Full protocol details are in the AFC governance checklist (ESM Part IV). 5 Company Scoring Architecture What Is the Company Score? Conceptual Summary The IASS tells us the ceiling. The Company Score tells us where on the staircase this specific firm is standing right now. I do not start with a general “AI maturity” label. I start with six observable dimensions that map directly to dollar denominated value in the model. Each dimension can be scored from public filings, management interviews, and technical diligence, and each low score identifies a specific investment constraint, not a general criticism. Example. JPMorgan Chase’s Workforce AI Augmentation Rate (WAR) is 8.5: 250,000 employees using LLM Suite daily, 4 hours of productivity gain per week per user. That is not a maturity label, it is a specific, verifiable operational fact with a direct dollar translation in the labor productivity value pool. By contrast, Zions Bancorporation’s WAR is 3.5: no enterprise GenAI platform has been publicly disclosed, and AI job postings are low. The 5 point gap between these two WAR scores translates to a computable difference in value pool capture, not a subjective “Zions is less advanced.” The bottleneck principle. If a company scores 8.0 on Process Automation Coverage but 3.5 on Data Infrastructure Maturity, the model does not average them. The CES Bottleneck Aggregator (Eq. 23) mathematically reduces the value pool capture rate to reflect the infrastructure deficit. You cannot build a reliable AI pricing engine on a broken data foundation. The framework makes this constraint explicit and quantified. Analogy. You cannot put a Ferrari engine on bicycle tires. A simple additive average of dimension scores would suggest that a leading autonomous AI application compensates for fragmented, siloed data infrastructure. The CES aggregator mathematically enforces the contrary: without data readiness, the AI capability has insufficient input to function, suppressing 70–80% of potential financial value. The weakest link constrains the system, not the arithmetic mean. Who should use this? An investment committee uses the six scores to surface the specific bottleneck that must be resolved for value to flow. A corporate leadership team uses the same scores to prioritize the technology budget: fix the weakest link first, because the CES formula proves that improving a strong dimension while the weak link persists generates near zero additional value. 5.1 Six Company Dimensions I decompose the company level AITG score into six dimensions that map directly to value pools in the VCB: 1. Data Infrastructure Maturity (DIM): From batch ERP only (score 1) to a production data lakehouse with real time streaming, full MLOps, and data governance (score 9). 2. Process Automation Coverage (PAC): From >>90 % manual processes (1) to AI native workflows across >>70 % of addressable surface area (9). 3. Workforce AI Augmentation Rate (WAR): From <<5 % of employees using any AI tool in workflows (1) to >>70 % augmented with tracked adoption metrics (9). 4. Decision Automation Rate (DAR): From fully human recurring decisions (1) to >>60 % of recurring decisions fully automated with exception handling (9). 5. AI Product/Revenue Integration (APR): From no AI revenue attribution (1) to AI native product with >>40 % of revenue attributable to AI features (9). 6. Organizational AI Capability (OAC): From no AI leadership (1) to C suite AI ownership with published model governance and risk controls (9). I score each dimension on a 0–10 scale using a rubric with five anchor points (1, 3, 5, 7, 9) and explicit evidentiary requirements. The six scores are averaged to produce the raw AITG composite. Table 3: AITG Dimension Scoring Rubric: Anchor Points and Evidentiary Requirements Dimension Abbrev. Score 1 Score 3 Score 5 Score 7 Score 9 Data Infrastructure Maturity DIM Siloed legacy databases; no cloud data platform Initial cloud migration; partial data catalog Unified data lake/ warehouse; governed data catalog Real time streaming pipelines; ML feature store Enterprise knowledge graph; automated data quality >>99% Process Automation Coverage PAC <<5% processes automated; manual workflows 10–20% RPA deployment; limited scope 30–50% process automation; cross functional 50–70% intelligent automation; AI augmented >>70% end to end automation; self optimizing Workforce AI Augmentation Rate WAR <<5% employees using any AI tool 10–20% with basic AI tools (copilots) 30–50% augmented; tracked adoption metrics 50–70% AI augmented; integrated into workflows >>70% augmented; org wide AI fluency measured Decision Automation Rate DAR Fully human recurring decisions <<10% decisions AI assisted 20–40% decisions AI recommended 40–60% automated with human override >>60% fully automated with exception handling AI Product/ Revenue Integration APR No AI revenue attribution <<5% revenue from AI enhanced features 10–20% revenue AI attributable 20–40% revenue AI driven products >>40% revenue from AI native products Organizational AI Capability OAC No AI leadership; ad hoc experiments Dedicated AI team; early governance C suite AI sponsor; published AI strategy Chief AI Officer; model risk framework C suite AI ownership; published governance & risk controls Note. Intermediate scores (2, 4, 6, 8, 10) are assigned when evidence places the firm between adjacent anchor points. Evidence sources include SEC filings, earnings transcripts, technology job postings, patent filings, and third party benchmarks (Lightcast, Stanford HAI). Full evidentiary protocols are specified in ESM Part I. 5.2 Cascading S Curve Architecture 5.2.1 The Low Ceiling Bias of a Single Logistic A single bounded logistic with fixed asymptote L=10L=10 introduces “low ceiling bias”: once a technology wave generates new capabilities, the ceiling shifts upward, but a static logistic cannot accommodate this. Historical GPT adoption patterns (electricity 1880s–1920s, computing 1960s–2000s, now AI) demonstrate cascading architectures where each wave begins before the prior wave is fully captured (David, 1990; Brynjolfsson et al., 2021). A single static logistic systematically underestimates both the ceiling and the urgency of early adoption. The Bass diffusion model (Bass, 1969): dF(t)dt=[p+qF(t)][1−F(t)] dF(t)dt= [p+qF(t) ]\! [1-F(t) ] (13) captures within wave dynamics correctly but is insufficient for multiwave GPT adoption. 5.2.2 three wave Cascading Specification I decompose the firm’s AI transformation trajectory into three sequential capability waves, each with its own logistic: Aw(t)=Lw1+e−kw(t−t0,w),w∈1,2,3A_w(t)= L_w1+e^-k_w(t-t_0,w), w∈\1,2,3\ (14) Table 4: Cascading S Curve: Three Capability Waves w Wave Scope LwL_w kwbasek_w^base (mo-1) t0,wbaset_0,w^base (mo) 1 Foundation AI Supervised ML, RPA, structured analytics, BI automation 4.0 0.38 18 2 Generative & Agentic AI LLM workflows, copilots, Retrieval Augmented Generation (RAG) pipelines, agentic task execution 3.5 0.42 36 3 Autonomous AI Multiagent orchestration, self improving systems, AI native products 2.5 0.32 60 Total asymptotic ceiling 10.0 Waves are staggered by their midpoints t0,wbaset_0,w^base, yielding a smooth, strictly increasing trajectory without explicit gating. With the calibrated parameters in Table 4, the early contribution of later waves is negligible in the score ranges used for the inverse mapping (Section 5.2.3). The total firm transformation score at time t is the sum across waves: AITG(t)=∑w=13Aw(t)=∑w=13Lw1+e−kw(t−t0,w)AITG(t)= _w=1^3A_w(t)= _w=1^3 L_w1+e^-k_w(t-t_0,w) (15) This cascading specification is strictly increasing in t, which guarantees a unique inverse mapping t^f t_f from any observed rubric score (Section 5.2.3). Table 4 presents wave parameters by industry tier. 5.2.3 Rubric to Curve Inverse Mapping The cascading specification (Eq. 15) is a function of time t, but the six dimension rubric produces a point in time score AITGraw∈[0,10]AITG^raw∈[0,10] from observable firm data. Without an explicit linking function, the S curve has no anchor to the empirical scoring system. I resolve this through the inverse mapping t^f t_f: the implied current position on the cascading S curve consistent with the observed rubric score. Given AITG(t)AITG(t) from Eq. (15), I solve: t^f=AITG−1(AITGfraw) t_f=AITG^-1\! (AITG^raw_f ) (16) Since AITG(t)AITG(t) is a sum of logistic terms and therefore strictly monotone increasing in t for well specified parameters, its inverse is unique and well defined for all AITGraw∈[0,10]AITG^raw∈[0,10]. Numerically, I solve via Newton–Raphson iteration on the residual AITG(t)−AITGraw=0AITG(t)-AITG^raw=0, with convergence guaranteed by strict monotonicity. t^f t_f has a direct economic interpretation: the number of months into an idealized AI transformation program where a firm with the observed rubric scores would be located. It serves two operational purposes: (1) Linking cross section to projection. The forward looking value trajectory uses t^f t_f as the starting point. The 5 year hold period value integral (Section 7.7) computes from t∈[t^f,t^f+60]t∈[ t_f,\; t_f+60], anchoring the longitudinal projection to the observed cross sectional state. (2) Identifying wave position. If t^f<t0,1 t_f<t_0,1, the firm is preinflection on Wave 1 and should not expect positive EBITDA contribution until the inflection is reached. If t^f>t0,2 t_f>t_0,2, the firm has cleared Wave 1 and Wave 2 (Generative AI) is beginning to propagate, directly informing implementation priorities. Remark 5.1 (Piecewise Inverse Implementation). The Wave 1 closed form inverse is only valid for AITGraw<L1=4.0AITG^raw<L_1=4.0; beyond this threshold, a piecewise specification (Eq. 44) or Newton–Raphson solver is mandatory to avoid a NaN domain error. Nine of the fourteen cohort companies require the multiwave branch. Full specification and software implementation notes appear in Appendix B. 5.2.4 AFC Integration at the Firm Level Equation 17 shows how the AFC modifies wave ceiling and timing at time t: Lw∗(t)=Lw⋅AFCi,tϕw,t0,w∗(t)=t0,w/AFCi,tμwL_w^*(t)=L_w·AFC_i,t _w, t_0,w^*(t)=t_0,w/AFC_i,t _w (17) where ϕ3>ϕ2>ϕ1 _3> _2> _1 (Wave 3 is most sensitive to frontier capability gains) and μw _w controls onset acceleration. This ensures the AFC recalibrates not just the industry ceiling but the firm level adoption trajectory. 5.2.5 Out of Sample Validation Strategy for θi _i I update θi _i using changes in O*NET Automatable Task Density (ΔATDi _i), not adoption survey data (Census BTOS), to avoid conflating the theoretical capability ceiling with current adoption breadth. The corrected update rule (Eq. 45) uses a learning rate η=0.30η=0.30 with bounds θ^i(t)∈[0.05,1.50] θ_i^(t)∈[0.05,1.50]. Full derivation and the circularity correction rationale appear in Appendix B. 5.3 IASS Normalization For cross industry comparison, I normalize the AITG to the AFC adjusted ceiling. The industry relative (IR) score and effective gap are: IRf,t=AITGf,trawIASSi,t∗×10,Geff=IASSi,t∗−AITGf,trawIR_f,t= AITG^raw_f,tIASS^*_i,t× 10, G_eff=IASS^*_i,t-AITG^raw_f,t (18) Standard dual reporting format: AITG 4.08 — IR 5.31 — Geff=3.60G_eff=3.60 — UQ=± 0.52. Scores carry a formal uncertainty band ±UQ aggregating data quality, model parameter, AFC, and interrater sources (typical combined ±0.45± 0.45 pts, 90 % CI). Full UQ decomposition and the UQ components table are in ESM Part I. 5.4 Illustrative IASS Calibrations: Five Anchor Industries Table 5 reports full dimensional breakdowns for five anchor industries spanning the IASS range. Complete calibrations for all 22 industry verticals, including NAICS codes, subscores, θi _i, and ψ, appear in ESM Table S1. Table 5: IASS Anchor Industry Calibrations: Five Representative Verticals (2026) Industry CTD DRSA PRI RFF CADR CLSR ψ IASS θi _i Vertical SaaS / Software 9.4 9.8 8.6 8.1 9.2 9.1 1.000 9.06 0.11 Financial Services (Banks) 8.8 9.2 7.9 5.0 8.4 7.6 1.000 7.83 0.22 Logistics / Transport 6.2 6.0 7.2 8.2 6.1 8.5 1.000 6.82 0.14 Healthcare Services 6.2 5.8 6.5 4.1 5.1 6.8 0.743 4.27 0.31 Construction 3.9 3.2 4.1 7.4 3.3 5.7 1.000 4.26 0.09 CTD = Cognitive Task Density; DRSA = Data/Systems Automation; PRI = Process Reengineering Index; RFF = Regulatory Friction Factor; CADR = Competitive AI Diffusion Rate; CLSR = Cost/Labor Structure. ψ<1.0ψ<1.0 = binding regulatory ceiling. Sources: BLS O*NET 2024; Lightcast 2025; Census BTOS. Full 22 industry table: ESM Table S1. Remark 5.2 (Three Diagnostic Patterns). Healthcare Services carries ψ=0.743ψ=0.743: the RFF hard floor suppresses its theoretical ceiling by 26%, reducing a geometric mean of 5.75 to IASS 4.27. Construction and Agriculture floor the table not from regulatory friction but from structurally low Cognitive Task Density; high physical task content that current AI cannot automate. Vertical SaaS and Financial Services anchor the top two quartiles (IASS >7.8>7.8), defining sectors where AI transformation opportunity is structurally dense. The full 22 industry pattern appears in ESM Table S1. 6 The AI Disruption Risk Index (ADRI) What Is the ADRI? Conceptual Summary The AITG measures how much opportunity a company is leaving on the table. The ADRI measures how fast that table is being cleared by competitors who are already capitalizing. Two companies can share identical AITG gaps yet face completely different competitive situations. A regional bank with AITG 3.6 in commercial banking (IASS∗ 9.38) occupies a fundamentally different position than an agricultural cooperative with AITG 3.6 in agriculture (IASS 4.72). Banking carries a Competitive AI Diffusion Rate (CADR) of 8.4, near the top of the reference table. JPMorgan’s proprietary LLM Suite extends its advantage every day the regional bank does not act. AI native lenders (Upstart, Blend) underwrite loans with lower human overhead. The regional bank is not merely “behind on technology”; it watches its competitive moat erode in real time. Example. Zions Bancorporation’s ADRI of 2.6 is the defining number in its AITG profile. It sits in an industry with CADR 8.4 and no proprietary AI stack. JPMorgan extends its advantage daily. The IFS endogeneity mechanism means Zions’s adoption timeline is approximately 1.97×1.97× longer than a data ready peer, so the competitive compound interest problem accelerates, not decelerates, with time. By contrast, JPMorgan’s ADRI of 0.5 reflects dominant incumbent status: proprietary data scale, regulatory capital requirements that bar new entrants, and proven enterprise AI execution. JPMorgan is not immune to disruption, but its competitive trajectory is improving. Who should use this? An investment committee uses ADRI to determine urgency: a wide gap, high ADRI target requires immediate transformation commitment; a wide gap, low ADRI target can be transformed patiently. A corporate board uses ADRI to assess whether its AI delay is a strategic choice or a competitive emergency. 6.1 Asymmetric Risk of Inaction The AITG gap measures opportunity. The ADRI measures something categorically different: the competitive risk generated by failing to close the gap. These are not symmetric. A company with a wide gap in a slow diffusion industry with strong structural moats faces a patient opportunity. A company with a wide gap in a fast diffusion industry with low switching costs faces an actively deteriorating trajectory Krakowski et al. (2023). The theoretical foundation is the superstar firm mechanism Autor et al. (2020): concentrating industries exhibit faster productivity growth because leading adopters scale efficiently while laggards face compounding cost disadvantage. Data nonrivalry Jones and Tonetti (2020) amplifies this dynamic: data rich adopters continuously improve their models while laggards start from scratch, producing accelerating divergence rather than mean reversion. The McKinsey 2025 State of AI survey McKinsey (2025) documents precisely this bifurcation: approximately 6 % of respondents report ≥ 5 % EBIT impact, while 88 % use AI but are not transforming. Definition 6.1 (AI Disruption Risk Index). The AI Disruption Risk Index for firm f in industry i at time t is: ADRIf,t=Geff,f,t×CADRi×(1−Moatf)×δtADRI_f,t= G_eff,f,t\;×\;CADR_i\;×\;(1-Moat_f)\;×\; _tN (19) normalized to [0,10][0,10] by the constant N. Components. • Geff,f,tG_eff,f,t: effective gap from Eq. 18. • CADRiCADR_i: Competitive AI Diffusion Rate from the IASS (0–10). • Moatf∈[0,1]Moat_f∈[0,1]: structural defensibility. Four factors scored 0–1: switching costs, network effects, regulatory barriers, proprietary data advantages. MoatfMoat_f is the weighted mean. • δt∈[1.0,1.5] _t∈[1.0,1.5]: time indexed urgency multiplier tied to the AFC. As frontier capabilities advance, the speed of competitive erosion accelerates: δt=1+0.5⋅min(Ct/C0−1, 1) _t=1+0.5· (C_t/C_0-1,\;1). Definition 6.2 (ADRI Competitive Hazard Intensity). The ADRI is operationalized as an instantaneous competitive hazard intensity. Define the hazard intensity function λf(t) _f(t) as the marginal probability of observing a measurable competitive displacement event (market share loss ≥1≥ 1 p or EBITDA margin compression ≥1.5≥ 1.5 p, measured at the firm level over a 12 month window) per unit time, conditional on the firm remaining in the nondisplaced state. I model: λf(t)=ADRIf,t=Geff,f,t⋅CADRi⋅(1−Moatf)⋅δt _f(t)\;=\; ADRI_f,tT\;=\; G_eff,f,t·CADR_i·(1-Moat_f)· _tN\,T (20) where =100T=100 is a dimensionless normalization constant calibrated so that λf(t) _f(t) is directly interpretable as the ADRI score in percent per year: a firm with ADRI =5=5 faces a 5 % annualized marginal displacement probability. N is the cross sectional normalization constant that maps the ADRI composite to [0,10][0,10]. The covariates are: (i) GeffG_eff, the effective transformation gap, which governs exposure magnitude; (i) CADRiCADR_i, the peer adoption velocity, which governs how quickly the gap translates into realized competitive disadvantage; (i) (1−Moatf)(1-Moat_f), the inverse structural defensibility, which modulates whether the hazard converts to displacement or is buffered by barriers to entry; and (iv) δt _t, the AFC linked urgency multiplier, which accelerates the hazard as frontier capabilities expand. The cumulative hazard over an inaction window [0,T][0,T] is Λf(T)=∫0Tλf(t)t _f(T)= _0^T _f(t)\,dt, and under the Poisson approximation (small λf _f), P(displacement by T)≈1−e−Λf(T)P(displacement by T)≈ 1-e^- _f(T). At ADRI=5.0ADRI=5.0 and T=24T=24 months, the model implies an approximate 10 % cumulative displacement probability, increasing to ∼ 13 % at ADRI =7=7 for the same horizon. Empirical threshold calibration. The displacement event thresholds (1 p market share, 1.5 p EBITDA margin) are calibrated to the lower bound of outcomes routinely attributed to competitive pricing pressure in public company earnings commentary. These are deliberately conservative to reduce false positive risk. The T constant is set at 12 months so that ADRI can be read as an approximate annualized percentage hazard at face value. Empirical validation of λf(t) _f(t) against firm level margin trajectories, using the panel study design specified in Section 11.5 item 5, will determine whether the Poisson approximation is warranted or whether a Weibull or Gompertz baseline hazard better fits the acceleration dynamics implied by δt _t. External adoption signals as auxiliary validators. I currently calibrate the CADRi component of ADRI from the Stanford HAI AI Index and Census BTOS survey data. Three external adoption signals could serve as auxiliary validators or instruments in future work, providing independent variation for both CADR and AFC trend estimation. • AI job posting density (Lightcast, Indeed): the share of industry job postings requiring AI related skills tracks adoption breadth with quarterly frequency and firm level granularity. Lightcast data already underlies IASS; its time series dimension has not yet been used for ADRI trend estimation. • Disclosure based AI density: 10-K AI disclosure intensity scores (count of AI related terms normalized by filing length, SEC EDGAR) provide a firm level panel observable from 2017 onward Krakowski et al. (2023). Disclosure intensity leads adoption implementation with roughly a 12 month lag, making it a natural leading indicator for CADR. • Software ecosystem diffusion: code level genAI adoption in enterprise software repositories (GitHub Copilot activation rates, AI assisted PR fractions by sector) provides a near real time proxy for technical adoption depth that precedes the 12–18 month lag in OEWS occupational data. AI mobility spillovers (engineers moving from AI native to laggard firms) can serve as an instrument for firm level adoption intensity following the approach of Brynjolfsson and McElheran (2016). I identify the incorporation of these signals into CADR, and their use as instrumental variables for ADRI validation, as a research agenda priority (Section 11.5). Table 6: ADRI Interpretation Grid ADRI Level Competitive Implication Action <2.5<2.5 Low Wide gap but protected by structural moat or slow diffusion Monitor annually 2.52.5–4.94.9 Moderate Gradual margin compression; window for orderly transformation Plan 3–5 year horizon 5.05.0–6.96.9 High Active competitive displacement; first mover advantages crystallizing Immediate investment ≥7.0≥ 7.0 Critical Structural impairment risk; M&A or repositioning may be required Acute / escalate 7 The Value Creation Bridge (VCB) 7.1 Architecture and Methodological Position The VCB converts an AITG gap into dollar denominated investment opportunity. The methodological position is explicit: Do not map “one AITG point” to dollars with a single global coefficient. That is fragile and easily gamed. Instead, size the economic prize by value pool from diligence ready financial baselines, and use the AITG and IASS only to bound feasibility, ramp speed, and capture rate. This mirrors how sophisticated acquirers model transformation value: from operational baselines (labor spend, pricing leakage, inventory turns), not from a unitless maturity score. 7.2 Value Pools 7.2.1 Firm Scale Factor (Φf _f) Φf _f captures proprietary stack scaling advantage: large data rich incumbents achieve lower per unit implementation cost and higher value pool multipliers. Industry critical scale thresholds Si∗S^*_i range from $0.5B (Construction) to $50B (Investment Banking); full table in ESM Table S1. Φf=11+e−α(log(Rf/Si∗)) _f= 11+e^-α( (R_f/S^*_i)) (21) where RfR_f is the firm’s annual revenue (a proxy for transactional data volume and proprietary dataset scale), Si∗S^*_i is the industry specific critical scale threshold below which proprietary model training is not economically feasible (Eq. 21), and α=2.0α=2.0 controls the steepness of the scale transition. The theoretical foundation is Jones and Tonetti (2020): proprietary data generates returns without being depleted, creating an asymmetry where data rich firms continuously improve their model quality while data poor firms start from scratch every training cycle. Analogy. Proprietary transaction data functions analogously to compound interest. A large cap bank training its models on hundreds of millions of daily transactions accumulates what Jones and Tonetti (2020) term “data gravity”: the asset improves continuously without being consumed. A regional bank licensing an off the shelf vendor platform, by contrast, deploys capability trained on a generic corpus and receives no compounding benefit from its own transaction history. The Firm Scale Factor formalizes this distinction: identical qualitative AI scores yield materially different financial returns depending on whether the firm owns the training data or rents the model. At Rf=Si∗R_f=S^*_i (revenue exactly at threshold), Φf=0.5 _f=0.5. At Rf=10×Si∗R_f=10× S^*_i, Φf≈0.99 _f≈ 0.99. At Rf=0.1×Si∗R_f=0.1× S^*_i, Φf≈0.01 _f≈ 0.01. Empirical grounding. JPMorgan Chase (Rf≈$177.6R_f≈ 177.6B revenue; Si∗=$3.3S^*_i= 3.3B) yields Φf=1.00 _f=1.00. Zions Bancorporation (Rf≈$3.4R_f≈ 3.4B; Si∗=$3.3S^*_i= 3.3B) yields Φf=0.52 _f=0.52. This directly explains the value density divergence: a subscale firm deploying vendor AI cannot replicate the proprietary data gravity of a scaled incumbent even at an identical dimension score gap. Threshold calibration and sensitivity. I calibrate the Si∗S^*_i thresholds from three sources: (1) model training cost curves, using Kaplan et al. (2020) scaling laws to infer the minimum proprietary dataset size for sector competitive model quality, converted to a revenue proxy via industry specific revenue per transaction rates from OEWS and FDIC data; (2) technology investment benchmarks from McKinsey and Gartner enterprise AI deployment studies McKinsey (2025), which report revenue thresholds above which firms are statistically more likely to report “full scale” AI deployment (vs. pilot stage); (3) revealed scale effects in the backtest cohort, where all firms above 3× their sector Si∗S^*_i (Φf>0.95 _f>0.95) generated positive Δ margins, while all firms below 0.5× Si∗S^*_i showed flat or negative margins (with confounds noted for Target, UPS, and Ford). Full Si∗S^*_i values by industry and sensitivity of Δ to ± 50% perturbations of Si∗S^*_i appear in ESM Table S2. The Si∗S^*_i parameters are the second highest ranked drivers of Δ variance in the Sobol analysis (24% of output variance from Φf _f alone), making this a priority for the Bayesian reestimation study (Section 11.5, item 7). Vendor AI cap. A firm deploying exclusively vendor AI (e.g., off the shelf nCino, Microsoft Copilot, Salesforce Einstein) has Φf _f capped at the vendor’s platform ceiling, regardless of its own revenue scale. For such firms, I set Φf=min(Φflogistic, 0.65) _f= ( _f^logistic,\;0.65), reflecting that vendor AI platforms level the playing field but do not replicate proprietary data gravity. This cap is disclosed in all scorecard outputs. The VCB defines seven standard value pools. Let ℬpB_p be the baseline dollar figure for pool p derived from diligence financials. Expected value from pool p before cost and risk adjustment is: Vp=ℬp×Φf×Capturep(Geff,bp)V_p=B_p× _f×Capture_p(G_eff,\,b_p) (22) Double counting safeguards. I construct the seven VCB value pools Vpp=17\V_p\_p=1^7 as mutually exclusive and collectively exhaustive (MECE): each pool maps to a distinct P&L line item (labor productivity → SG&A headcount; revenue enhancement → top line uplift; working capital → DIO/DSO improvement; etc.). The partition is enforced at the calibration stage: base capture rates κ¯p κ_p (Table S2 in ESM) are estimated from nonoverlapping value pools in published AI impact studies (Brynjolfsson et al., 2025; McKinsey, 2025). Two additional safeguards prevent double counting at the computation stage. First, the aggregate capture constraint ∑pCapturep≤1 _pCapture_p≤ 1 is enforced across all pools: if pool level captures exceed unity (implying more than 100% of the theoretical uplift is realized), all captures are proportionally rescaled. In practice this constraint is nonbinding for all 14 calibration companies, because per pool capture rates (κ¯p∈[0.40,0.65] κ_p∈[0.40,0.65]) and the concave η(g)η(g) function jointly ensure ∑pCapturep<0.85 _pCapture_p<0.85 even at maximum gap. Second, the Monte Carlo VCB analysis in ESM Part I independently verifies that total Δ does not exceed the sum of individually estimated pool values under 10,000 correlated draws, confirming that the partition constraint holds probabilistically. 7.3 Bottleneck Aggregation: CES Specification Each value pool p depends on a subset of AITG dimensions (p)⊆1,…,6D(p) \1,…,6\. Let ed=max(AITGdraw/10, 0.01)∈(0,1]e_d= (AITG^raw_d/10,\;0.01)∈(0,1]. The floor ed≥0.01e_d≥ 0.01 prevents division by zero in the CES formula (ed−ρ→∞(e_d^-ρ→∞ as ed→0e_d→ 0); a score of zero on any dimension is economically indistinguishable from a score of 0.1 for CES purposes, as both produce near Leontief bottleneck collapse. A Constant Elasticity of Substitution (CES) aggregator preserves the bottleneck intuition while being robust to measurement error in individual dimension scores: bp=(∑d∈(p)αd⋅ed−ρ)−1/ρb_p= ( _d\,∈\,D(p) _d· e_d^-ρ )^-1/ρ (23) where: • ρ>0ρ>0 is the substitution parameter; the elasticity of substitution is σ=1/(1+ρ)σ=1/(1+ρ). This paper uses the Arrow–Chenery–Minhas–Solow (ACMS) negative exponent form so that bp−ρb_p^-ρ is a weighted power mean of the reciprocals of the dimension scores, i.e. a generalized harmonic mean. At ρ=5ρ=5: σ=1/(1+5)=1/6σ=1/(1+5)=1/6. These two values are therefore mutually consistent. This form differs from the positive exponent CES sometimes written (∑αdedρ)1/ρ(Σ _de_d^ρ)^1/ρ with ρ∈(−∞,1)ρ∈(-∞,1); under that parameterization ρ=5ρ=5 would be outside the substitution domain. The chosen form is standard for bottleneck aggregation (see notation summary, Table 15). • αd>0 _d>0 are dimension level importance weights, ∑d∈(p)αd=1 _d (p) _d=1. • As ρ→∞ρ→∞: bp→mindedb_p→ _de_d (Leontief limit). • As ρ→0ρ→ 0: bp→∏dedαdb_p→ _de_d _d (Cobb-Douglas). • At ρ=1ρ=1: bp=1/∑d(αd/ed)b_p=1 /\! _d( _d/e_d) (harmonic mean). Proposition 7.1 (CES Boundary Behaviour). Under the ACMS CES form (Eq. 23) with the floor ed≥0.01e_d≥ 0.01: 1. Leontief limit. As ρ→∞ρ→∞: bp→mind∈(p)edb_p→ _d (p)e_d. The weakest dimension fully determines the bottleneck. 2. Cobb–Douglas limit. As ρ→0+ρ→ 0^+: bp→∏d∈(p)edαdb_p→ _d (p)e_d _d, the weighted geometric mean. 3. Strict positivity. For all ρ>0ρ>0: bp≥0.01>0b_p≥ 0.01>0, since the floor ensures ed−ρ≤100ρe_d^-ρ≤ 100^ρ for all d, hence bp≥(|(p)|⋅100ρ)−1/ρ>0b_p≥(|D(p)|· 100^ρ)^-1/ρ>0. Proof. See Appendix B. ∎ Calibration and sensitivity. I calibrate ρ=5ρ=5 against Standish Group (2015): at ρ=5ρ=5, a single bottleneck dimension at ed=0.3e_d=0.3 reduces the CES aggregate by approximately 70% relative to the no bottleneck case. ESM Table S3 reports Spearman correlations and Δ ranges for ρ∈[3,8]ρ∈[3,8]; no rank order changes occur. A fixed ρ across all value pools is a simplification; the discussion in Section 11 addresses the case for enabler specific or time varying elasticities. Relation to cascaded CES and nonlinear Domar aggregation. Several results from the network production and nonlinear aggregation literatures bear on the single layer CES choice made here. First, Domar aggregation with CES technologies Hulten (1978) establishes that input complementarities at the firm or sector level can produce amplified responses to bottleneck shocks: when a single input has a low elasticity of substitution, a productivity shortfall in that input propagates upstream with multiplied effect. The AITG CES bottleneck is consistent with this mechanism: a weak data infrastructure score (ed=0.30e_d=0.30) propagates into the value pool estimate more than proportionally; the choice of ρ=5ρ=5 is conservative relative to production estimates for digital complementarity settings. Second, cascaded or nested CES models Sato (1967) allow enablers to be grouped hierarchically, with different elasticities at each nesting level. A natural application would nest “data stack” (Data Infrastructure Maturity + AI Platform Readiness) and “execution capacity” (Organizational Change Capacity + Workforce Augmentation) as inner aggregators, with a lower elasticity between the two outer groups. Under such a nesting, the interaction effect between a weak data stack and weak execution capacity would be additive at the inner level but superadditive at the outer level. This structure is theoretically better motivated than the current single layer CES; I identify it as a technical extension in the research agenda (Section 11.5). Third, I note the potential for CES near singularities when σ→0σ→ 0: as ρ→∞ρ→∞ the aggregator collapses to minded _de_d (Leontief), which can produce discontinuous jumps in bpb_p around binding thresholds. The floor ed≥0.01e_d≥ 0.01 and the finite ρ=5ρ=5 prevent this pathology; any future variable elasticity extension should maintain a finite lower bound on σ to avoid degenerate outputs. 7.4 Gap to Capture Scaling Function Define the normalized gap fraction: gf=Geff,f10=max(0,IASSi∗−AITGfraw)10g_f= G_eff,f10= (0,\ IASS^*_i-AITG^raw_f)10 (24) The expected capture fraction before risk haircut is: Capturep=κ¯p⋅bp⋅η(gf),η(g)=1−e−λgCapture_p= κ_p· b_p·η(g_f), η(g)=1-e^-λ g (25) where κ¯p κ_p is the base capture rate for pool p (Section 7), η(g)η(g) is an increasing, concave scaling function with λ=3.5λ=3.5, and bpb_p is the bottleneck factor. The concavity of η reflects the empirical finding that accessible wins come first and structurally resistant tasks persist Brynjolfsson and McElheran (2016); McKinsey (2025). Standard value pool parameters, uplift rates, and calibration sources are in ESM Table S2. Pricing leverage note. The pricing pool deserves special emphasis. A commonly cited analysis implies that a 1 % price increase translates to a substantial operating profit improvement assuming no volume loss, the exact mechanism the pricing pool captures. Brynjolfsson et al. (2025) show that AI augmented workers improve revenue yield (issues resolved, customer satisfaction) by 14 % on average, and 34 % for novice workers, demonstrating the practical scale of the labor productivity and revenue yield pools. 7.5 Cost Model VCB splits implementation costs into capitalizable vs. period expense and one time build vs. run rate, following updated GAAP guidance (ASU 2025-06 for internal use software) and IFRS IAS 38 for development expenditure. NPV(Cost)=∑t=1TCcap,t+Copex,t(1+WACC)tNPV(Cost)= _t=1^T C_cap,t+C_opex,t(1+WACC)^t (26) where Ccap,tC_cap,t are capitalizable implementation costs, Copex,tC_opex,t are period expenses (SaaS fees, training, ongoing inference, monitoring), and WACC is the Weighted Average Cost of Capital applied as the risk appropriate discount rate. Under US GAAP, software configuration and integration labor meeting the ASU 2025-06 authorization/probability threshold is capitalizable; training and ongoing compute are expensed. Under IFRS, SaaS fees are generally service expenses; only demonstrably controlled intangibles meeting IAS 38 development criteria are capitalized. 7.6 Value Ramp Function Value does not materialize immediately. Following the J curve logic Brynjolfsson et al. (2021), I model value realization as a logistic ramp. A prior version used a static default of t50=18t_50=18 months for all firms, creating a disconnect: the adoption model correctly delayed software deployment for unprepared firms (via IFS adjusted t0,w,ft_0,w,f), but the financial model projected value arriving at month 18 regardless. Cash flows arrived 17 months before the software was adopted. The corrected specification makes t50t_50 a firm specific endogenous quantity. 7.7 Enterprise Value Creation: Full Decomposition Notation. Define the acquisition baseline ramp Rf0≡R(t^f;t50,f)R_f^0≡ R( t_f;\,t_50,f) and the end of hold ramp RfT≡R(t^f+T;t50,f)R_f^T≡ R( t_f+T;\,t_50,f), where t50,ft_50,f is the firm specific endogenous ramp inflection (Section 7.6) and T=60T=60 months is the 5 year hold. The incremental ramp fraction, the share of the full value pool not yet in baseline EBITDA at acquisition, is: ΔRf≡RfT−Rf0 R_f≡ R_f^T-R_f^0 (27) Component 1: Terminal Value (TV). Exit multiple capitalization of the incremental run rate EBITDA uplift realized during the hold period. Vprun rateV_p^run rate already contains Φf _f (Eq. (22)); it must not be multiplied by Φf _f again: TVf=(∑p=17Vprun rate)⏟already contains Φf×ΔRf×Mi×IFSfTV_f= ( _p=1^7V_p^run rate )_already contains _f× R_f× M_i×IFS_f (28) Remark 7.1 (Baseline Already Captured). If ΔRf R_f is replaced by RfTR_f^T (the prior error), the model takes credit for value already priced into the acquisition multiple. Eq. 27 defines the corrected ramp increment; the magnitude of this error varies substantially across firms: firms with t^f≫t50,f t_f t_50,f (UPS, HCA, Rockwell) have Rf0≈0.96R_f^0≈ 0.96, making ΔRf≈0.04 R_f≈ 0.04 and the prior overstatement ≈25×≈ 25×. For early stage firms (Zions, Ferguson, Rf0≈0.52R_f^0≈ 0.52), the prior overstatement was ≈2×≈ 2×. Component 2: Interim Free Cash Flow (FCF) Integral. Discounted incremental cash flows during the hold, also net of baseline: FCFfinterim=∫t^ft^f+T(∑p=17Vprun rate)⋅[Rf(t)−Rf0]⋅IFSf⋅e−rWACCt/12dtFCF_f^interim= _ t_f t_f+T ( _p=1^7V_p^run rate )· [R_f(t)-R_f^0 ]·IFS_f· e^-r_WACCt/12\,dt (29) Discrete annual approximation: FCFfinterim≈∑y=15(∑p=17Vprun rate)⋅[Rf(t^f+12y)−Rf0]⋅IFSf(1+WACC)yFCF_f^interim≈ _y=1^5 ( _p=1^7V_p^run rate )· [R_f( t_f+12y)-R_f^0 ]·IFS_f(1+WACC)^y (30) Total 5 Year Risk Adjusted Enterprise Value Creation. ΔEVf=TVf+FCFfinterim−NPV(Cost) _f=TV_f+FCF_f^interim-NPV(Cost) (31) Remark 7.2 (Economic Interpretation of ΔRf R_f). The two limiting cases clarify the framework’s behavior. For an apex incumbent (JPMorgan, Salesforce: Rf0≈1.0R_f^0≈ 1.0), ΔRf≈0 R_f≈ 0 and ΔEV≈−NPV(Cost) ≈-NPV(Cost): the transformation is already complete, and additional investment yields no incremental ramp benefit. The framework correctly signals that the investment case is maintenance, not transformation. For an early stage target (Zions, Ferguson: Rf0≈0.08R_f^0≈ 0.08–0.130.13), ΔRf≈0.85 R_f≈ 0.85–0.870.87: most of the value pool is incrementally capturable, and the investment case is strongest. This monotone relationship between t^f t_f and investment attractiveness is a structural property of the framework, not a calibration artifact. Definition 7.1 (AITG Value Density). AITG Value Density is risk adjusted 5 year enterprise value creation per $1M of implementation cost: AITG-VDf=ΔEVf (risk adjusted, 5 yr)Total Implementation Cost ($M)AITG -VD_f= _f (risk adjusted, 5 yr)Total Implementation Cost ( M) (32) Institutional thresholds: AITG-VD ≥2.0×≥ 2.0× = Tier 1 (Invest); 1.01.0–2.0×2.0× = Tier 2 (Monitor); <1.0×<1.0× = Tier 3 (Pass). 8 Implementation Feasibility: Endogenous Risk Integration What Is the IFS? Conceptual Summary The technology works at JPMorgan. The question is whether it will work at this company, with this management team, with this data infrastructure, on your timeline. This is the most dangerous step in AI due diligence, and the one most frequently skipped. Every management team presents an AI roadmap. The Implementation Feasibility Score (IFS) is the framework’s mechanism for stress testing that roadmap against structural organizational reality. The key insight, and the mathematical departure from prior models, is that IFS factors are not a discount applied after the fact. A company with poor data readiness does not simply capture 70% of a theoretical value estimate. It follows a materially different trajectory: a slower ramp, a later inflection, and a higher probability of stalling before reaching the value generative portion of the S curve. Bloom, Sadun, and Van Reenen (2012) show that management quality explains approximately 30% of cross firm total factor productivity differences in manufacturing, not as a static quality label, but as a dynamic determinant of whether technology investments generate returns. Example. JPMorgan’s Organizational Change Capacity (δOCC _OCC) is 0.92: 250,000 employees voluntarily adopted LLM Suite within 8 months of launch. This is not a claim about JPMorgan’s culture; it is a documented adoption velocity event. Zions Bancorporation’s δOCC _OCC is 0.55, reflecting a decentralized 11 brand structure with no enterprise AI governance framework disclosed. The IFS framework translates this structural difference into a specific parameter adjustment: Zions’s adoption inflection arrives approximately 1.97×1.97× later than a data ready peer’s. That delay is not a cosmetic discount; it is the compound interest cost of organizational unreadiness. Analogy. Think of it as a marathon. Prior frameworks apply implementation risk as a flat discount: say you will run a marathon, but award only 70% of the medal because you are undertrained. The AITG endogenous model says something different: because you lack data readiness, you will run slower and finish later; the penalty is temporal, not scalar. Mathematically, the inflection point of the adoption S curve shifts rightward, which compounds the Internal Rate of Return (IRR) penalty through delay rather than through an arbitrary terminal value haircut. The distinction is not cosmetic; it changes which interventions the investment committee should prioritize. Who should use this? An investment committee uses IFS to determine whether the value waterfall survives the management reality. A corporate leadership team uses it to identify whether their transformation plan is structurally feasible or requires parallel organizational change investment before the technology investment will yield returns. 8.1 IFS Validation Ablation analysis confirms that trajectory factors (OCC, DR) contribute the largest marginal signal: dropping OCC reduces backtest ρs _s by 0.109. Residual factors (VTR, CRS, REG) contribute less individually but collectively account for an additional ++15% reduction in Δ error. Full ablation results (Table 20) and component interpretability details appear in Appendix C. 8.2 The Endogeneity Problem with Post Hoc Risk Multipliers A post hoc IFS multiplier on a precomputed ΔEV double counts risk: the value decomposition already reflects expected execution, so an additional discount produces biased estimates. I integrate implementation risk endogenously into S curve parameters, shifting the ramp inflection rather than discounting terminal value. 8.3 Endogenous Integration: IFS Factors into Curve Parameters I restructure the IFS into two components. Trajectory factors (organizational change capacity and data readiness) are integrated directly into the wave steepness kwk_w and ramp timing t50t_50. Residual factors (vendor risk, competitive response speed, regulatory exposure) affect terminal value realization and cost, and remain as a limited residual multiplier. 8.3.1 Trajectory Risk Factors Let δOCC∈[0.50,1.00] _OCC∈[0.50,1.00] be the Organizational Change Capacity (OCC) score and δDR∈[0.50,1.00] _DR∈[0.50,1.00] be the data readiness (DR) score, both calibrated via the rubric in Table 7. Remark 8.1 (Range Correction). Prior versions stated δOCC∈[0.60,1.00] _OCC∈[0.60,1.00]. Stress testing against the Zions Bancorporation case study showed that the empirically derived score (δOCC=0.55 _OCC=0.55) fell below the declared lower bound. The domain is extended to [0.50,1.00][0.50,1.00] for both factors. A hard floor of δmin=0.10 _ =0.10 is added to Eq. (35) to prevent division by zero under degenerate inputs. The adjusted wave steepness for firm f is: kw,f=kw,base⋅ϕOCC(δOCC)⏟org. capacity factor⋅ϕDR(δDR)⏟data readiness factork_w,f=k_w,base· _OCC( _OCC)_org. capacity factor· _DR( _DR)_data readiness factor (33) where the adjustment functions are: ϕOCC(δ)=0.55+0.45δ,ϕDR(δ)=0.50+0.50δ _OCC(δ)=0.55+0.45δ, _DR(δ)=0.50+0.50δ (34) mapping δ∈[0.50,1.00]δ∈[0.50,1.00] to steepness multipliers bounded strictly above zero. The adjusted value ramp midpoint for firm f (default t50,base=18t_50,base=18 months): t50,f=t50,basemax(δOCC, 0.10)0.40⋅max(δDR, 0.10)0.60t_50,f= t_50,base ( _OCC,\,0.10)^0.40· ( _DR,\,0.10)^0.60 (35) In forward simulations, I apply the same delay factor to the adoption wave midpoints: t0,w,f=t0,w,base⋅(t50,f/t50,base)t_0,w,f=t_0,w,base·(t_50,f/t_50,base). The companion Monte Carlo code (ESM) uses a refined specification in which t50,baset_50,base itself varies with the firm’s position on the S curve: t50,base=18+(AITGraw/10)×42t_50,base=18+(AITG^raw/10)× 42 months, reflecting the empirical observation that firms further along the adoption curve face longer value realisation horizons as they pursue higher wave capabilities. The analytical worked examples in Appendix D use the constant t50,base=18t_50,base=18 base case for transparency. The exponents 0.400.40 and 0.600.60 reflect the empirical finding from Bloom et al. (2016) and Brynjolfsson and Hitt (2000) that data readiness is the more binding constraint on adoption speed. Remark 8.2 (Corrected Zions Delay Factor). The paper previously cited a 1.72×1.72× inflection delay for Zions (δOCC=0.55 _OCC=0.55, δDR=0.48 _DR=0.48). Numerical evaluation gives: 1/(0.550.40×0.480.60)=1.97×1/(0.55^0.40× 0.48^0.60)=1.97×. The corrected delay is 1.97×1.97×, strengthening the disintermediation narrative. All downstream Δ figures for Zions are updated accordingly. Observation: The practical implication is significant. A firm with δOCC=0.70 _OCC=0.70 and δDR=0.65 _DR=0.65 (both near minimum) runs kw,f≈0.71⋅kw,basek_w,f≈ 0.71· k_w,base with t0,w,f≈1.49⋅t0,w,baset_0,w,f≈ 1.49· t_0,w,base: a 29% shallower curve delayed by 49%. This is not a uniform 29% haircut on terminal value; it is a compounding deceleration that reduces 5 year cumulative value by 38–52% depending on the wave mix. 8.3.2 Residual IFS Multiplier The three residual IFS factors, vendor/technology risk (δ3 _3), competitive response speed (δ4 _4), and regulatory exposure (δ5 _5), affect terminal value realization and incremental cost structure but do not alter the adoption trajectory. I retain them as a geometric residual multiplier: IFSfresidual=∏j∈3,4,5δjwj,w3=0.35,w4=0.35,w5=0.30IFS_f^residual= _j∈\3,4,5\ _j^w_j, w_3=0.35,\;w_4=0.35,\;w_5=0.30 (36) 8.4 Revised Enterprise Value Calculation The full AITG enterprise value equation corrects four algebraic fractures identified through adversarial review: ΔEVf _f =ΔRf⋅∑pVpraw⋅IFSftraj⏟TVf = R_f· _pV_p^raw·IFS_f^traj_TV_f (37) +∑y=15(∑pVprun rate)⋅[Rf(t^f+12y)−Rf0]⋅IFSfresid(1+WACC)y⏟FCFinterim−NPV(Cost) +\; _y=1^5 ( _pV_p^run rate )· [R_f( t_f+12y)-R_f^0 ]·IFS_f^resid(1+WACC)^y_FCF^interim\;-\;NPV(Cost) Four corrections are encoded in this equation relative to prior versions: (1) Φf _f enters through Vprun rateV_p^run rate via Eq. (22) exclusively; it does not appear separately in the TV or FCF numerators, thereby eliminating the Φf2 _f^2 double discount that suppressed 73 % of Zions’ value pool. (2) TV uses ΔRf=RfT−Rf0 R_f=R_f^T-R_f^0 (Eq. 27) rather than RfTR_f^T alone, deducting value already captured in baseline EBITDA. (3) FCF uses [Rf(t^f+12y)−Rf0][R_f( t_f+12y)-R_f^0], applying the same baseline deduction to each annual cash flow. (4) Rf(t)R_f(t) uses the endogenous firm specific inflection t50,ft_50,f (Eq. 35) rather than the static 18 month default, aligning financial value arrival with actual software adoption timing. Vp(kw,f,t0,w,f)V_p(k_w,f,t_0,w,f) shows that value pool capture is a function of firm adjusted wave dynamics. See Section 7.7 for the derivation and economic interpretation of each component. 8.5 IFS Rubric: Five Factors Table 7: IFS Factor Specification (Revised: Trajectory vs. Residual) Factor Type Definition Weight δ Range Org. Change Capacity Trajectory Leadership alignment, cultural readiness, prior transformation record 0.30 [0.60, 1.00] Data Readiness Trajectory Quality, completeness, accessibility, governance; data pipeline health 0.30 [0.55, 1.00] Vendor/Tech Risk Residual AI vendor ecosystem maturity; contract flexibility; switching costs 0.35* [0.75, 1.00] Competitive Response Residual Speed at which AI advantages erode due to competitor adoption (CADR) 0.35* [0.70, 1.00] Regulatory Exposure Residual Risk of regulatory action or compliance cost escalation during hold 0.30* [0.55, 1.00] * Residual weights normalized within the IFSresidIFS^resid product; trajectory factors absorbed into kw,fk_w,f and t0,w,ft_0,w,f. 9 Empirical Illustrations Application Contexts: External Valuation and Internal Assessment The fourteen illustrations serve two purposes. For investment committees: each generates an auditable, deal specific narrative from industry ceiling through implementation cost to risk adjusted value density, anchored entirely to public data. For corporate self assessment: the same framework lets a Chief AI Officer benchmark the institution against its industry frontier, quantify competitive urgency, and build a board level investment case. The Zions analysis illustrates the internal use case: a structured diagnosis that identifies binding constraints, estimates the value of resolving them, and quantifies how delay compounds disadvantage. All scores derive from public data; internal application with actual infrastructure debt and adoption rates yields a more precise diagnosis. 9.1 Scoring Methodology and Data Tier All companies are scored exclusively from public sources: 10-K/10-Q filings, earnings transcripts, investor day presentations, and AI deployment disclosures (2024–2025). Evidence is classified Tier A (audited/quantified) through Tier D (analyst estimate); scores with >50%>50\% Tier C/D inputs receive a Limited Confidence flag. Financial baselines for VCB computation are in ESM Table S4. See ESM Table S4 for full financial baselines and data tier classifications. 9.2 JPMorgan Chase: Dominant Incumbent Benchmark Case JPMorgan Chase is the world’s largest bank by market capitalization and the defining case study for what AI transformation looks like when executed with maximum organizational commitment and proprietary data scale. As of 2025, JPMorgan has committed $18 billion to technology annually, allocated approximately $2 billion specifically to AI, and deployed its proprietary Large Language Model Suite (LLM Suite) platform to approximately 250,000 employees, its entire workforce excluding branch and call center staff JPMorgan Chase CNBC (2025). Roughly half of those employees use LLM Suite daily. AI attributed financial benefits have grown 30–40% annually since the program’s inception McKinsey and JPMorgan (2025) Babina et al. (2024). Chief Analytics Officer Derek Waldron has publicly characterized the long term goal as creating the world’s first “fully AI connected enterprise.” This case also validates a critical AITG mechanism: the Firm Scale Factor (Φf _f). Smaller regional banks cannot train proprietary foundation models because they lack the transactional data volume required to achieve gradient descent at scale, a property I formalize as data gravity. JPMorgan’s proprietary dataset of global transaction flows, covenant libraries, credit histories, and client behaviors constitutes a nonreplicable competitive asset that its AI models continuously improve. This is the data nonrivalry mechanism of Jones and Tonetti (2020) in practice: JPMorgan’s data generates returns without being depleted, while a regional bank using off the shelf models competes on a perpetually leveling playing field. Table 8: JPMorgan Chase: AITG Dimensional Scores (Tier 1) Dimension Score Tier Evidence IASS Anchor: Commercial Banking (Large Cap), IASS=∗9.38^*=9.38 (AFC adjusted, C2026=1.90C_2026=1.90) Data Infrastructure (DIM) 8.5 A $18B tech budget; 65% workloads on cloud; near zero legacy debt Process Automation (PAC) 8.2 A 450+ AI use cases in production; covenant extraction; pitch deck generation Workforce Augment. (WAR) 8.5 A 250K LLM Suite users; ∼ 4 hrs/week productivity gain per employee Decision Automation (DAR) 7.5 A/B Agentic AI deployment begun; multistep autonomous task handling AI Revenue Integration (APR) 7.8 A/B AI enabled client concierge; IndexGPT personalization; ChatCFO Org. AI Capability (OAC) 8.8 A 2,000 AI staff; Chief Data and Analytics Officer (CDAO) governance; Waldron PhD computational physics IFS Trajectory Factors Org. Change Capacity (δOCC _OCC) 0.92 A American Banker 2025 Innovation of the Year; viral opt in LLM Suite Data Readiness (δDR _DR) 0.94 A 80% modern infrastructure; model agnostic platform; 8 week update cycle IFS Residual Factors Vendor/Tech Risk (δVTR _VTR) 0.91 A Model agnostic; in house build for security; OpenAI + Anthropic hedged Comp. Response Speed (δCRS _CRS) 0.88 A Leading EVIDENT AI Index; first mover on agentic enterprise platform Regulatory Exposure (δREG _REG) 0.85 B Heavily regulated; but proven execution at scale under OCC/Fed oversight AITG Output Summary (JPMorgan Chase): AITGraw=8.22|IR=8.76|Geff=1.16|ADRI=0.5(Low)AITG^raw=8.22\; |\;IR=8.76\; |\;G_eff=1.16\; |\;ADRI=0.5\;(Low) IFSresidual=0.88⇒ΔEV5yr≈$20–$98BIFS_residual=0.88\; \; _5yr≈ 20-- 98B MC uncertainty (M=10,000):P10=$20.2B|P50=$44.4B|P90=$97.9BMC uncertainty (M=10,000): _10= 20.2B\; |\;P_50= 44.4B\; |\;P_90= 97.9B Uncertainty decomposition. Sobol first order analysis attributes 50% of Var(ΔEV)Var( ) to exit multiple assumptions and 44% to capture rate assumptions, jointly accounting for 94% of output variance. The base case scalar ($44.4 44.4B) coincides with the Monte Carlo P50; the P10 of $20.2B remains materially positive, confirming that the investment case survives adverse scenario stress. Full distributional output is in ESM Table S5. ADRI context. JPMorgan’s ADRI of 0.5 reflects its status as the competitive dominant incumbent in its category. Its moat (Moatf=0.88Moat_f=0.88) derives from four durable sources: proprietary data scale, regulatory capital requirements that create structural barriers to AI native fintech entry at equivalent scale, an established client relationship network with high switching costs, and demonstrated ability to execute complex enterprise AI at production scale under stringent regulatory oversight. JPMorgan is not immune to disruption, but its competitive trajectory is improving, not deteriorating. CES Bottleneck observation. Because JPMorgan scores uniformly high across all six company dimensions (8.5 / 8.0 / 8.5 / 7.5 / 7.8 / 8.8), the CES bottleneck aggregator with ρ=5ρ=5 does not penalize value pool capture. The weakest link (DAR at 7.5) exerts minimal drag, confirming that the bottleneck operator behaves correctly: a well balanced high performer incurs no cross dimension penalty. This is precisely the behavior I designed the CES specification to exhibit. 9.3 Zions Bancorporation: The Disintermediation Risk Illustration Zions Bancorporation is a $89 billion asset regional bank operating under 11 distinct brand identities across 11 western U.S. states. With FY2025 revenue of $3.4 billion and approximately 10,000 employees, Zions occupies the competitive sweet spot most vulnerable to AI enabled disintermediation: too large to be a nimble community bank, too small to fund a proprietary AI platform Zions Bancorporation (2026). Zions’s AI initiatives are nascent but real. In April 2025, Zions selected nCino as its technology platform for loan origination transformation, deploying nCino’s Banking Advisor, Commercial Pricing, and Profitability and Analysis tools to modernize end to end lending processes nCino Zions (2025). This is an important distinction: Zions is deploying vendor AI (nCino’s off the shelf Banking Advisor), not proprietary AI. It is purchasing AI as a service, which means its competitors have access to identical or equivalent capability for similar per seat licensing fees. There is no data gravity moat here. Table 9: Zions Bancorporation: AITG Dimensional Scores (Tier 1) Dimension Score Tier Evidence IASS Anchor: Commercial Banking (Large Cap), IASS=∗9.38^*=9.38 (AFC adjusted, C2026=1.90C_2026=1.90) Data Infrastructure (DIM) 4.5 C Multibrand legacy systems; nCino migration in early stages; no data lake disclosed Process Automation (PAC) 4.0 C nCino deployment begun 2025; primarily manual lending workflow currently Workforce Augment. (WAR) 3.5 C/D No enterprise Generative AI (GenAI) platform disclosed; limited AI skill demand in job postings Decision Automation (DAR) 3.8 C nCino pricing tools initial; no autonomous credit decisioning at scale AI Revenue Integration (APR) 2.8 D No AI enabled product differentiation disclosed in public filings Org. AI Capability (OAC) 4.2 C No dedicated CAIO; no public AI governance framework; nCino partnership announced IFS Trajectory Factors Org. Change Capacity (δOCC _OCC) 0.55 C Decentralized 11 brand structure creates AI governance complexity Data Readiness (δDR _DR) 0.48 C/D Multilegacy core systems; fragmented data across brands; no cloud migration disclosed IFS Residual Factors Vendor/Tech Risk (δVTR _VTR) 0.72 B nCino is proven technology; single vendor dependency risk Comp. Response Speed (δCRS _CRS) 0.62 C Regional bank competitors (e.g., Western Alliance) advancing AI; JPM extends lead daily Regulatory Exposure (δREG _REG) 0.82 B OCC regulated; manageable; does not add incremental suppression AITG Output Summary (Zions Bancorporation): AITGraw=3.80|IR=4.05|Geff=5.58|ADRI=2.6(Moderate/Elevated)AITG^raw=3.80\; |\;IR=4.05\; |\;G_eff=5.58\; |\;ADRI=2.6\;(Moderate/Elevated) IFSresidual=0.71⇒ΔEV5yr≈$0.6–$1.2B(transformative scenario only)IFS_residual=0.71 _5yr≈ 0.6-- 1.2B\;(transformative scenario only) MC uncertainty (M=10,000):P10=$0.61B|P50=$0.88B|P90=$1.23B(transformative scenario only)MC uncertainty (M=10,000): _10= 0.61B\; |\;P_50= 0.88B\; |\;P_90= 1.23B\;(transformative scenario only) Note on Zions uncertainty. The transformative scenario P10 of $610M represents approximately 7% of Zions's enterprise value; meaningful but not transformative under base case execution assumptions. The wide P10–P90 spread ($610M to $1.23B) confirms the ADRI diagnostic: execution risk (low δDR=0.48 _DR=0.48) dominates the output distribution, producing high outcome variance even within the transformative scenario. Full scenario grid in ESM Table S5. The Disintermediation Scenario. Zions’s ADRI of 2.6 is the most consequential number in this analysis. ESM Table S1 reports that Commercial Banking has CADR =8.4=8.4, one of the highest competitive diffusion rates in the reference table. This rate implies rapid AI diffusion across the competitive peer group. JPMorgan is extending its proprietary AI advantage daily. AI native lenders (Upstart, Blend, Figure Technologies) are using AI to underwrite loans with lower human overhead. And Zions sits with a 5.58 point effective gap in a 9.38 ceiling industry with only commodity vendor AI tools as its AI stack. The IFS endogeneity mechanism makes this especially acute. Zions’s δDR=0.48 _DR=0.48 (well below the δDR=0.60 _DR=0.60 midpoint) implies, from Equation (35), that its adoption curve is time shifted by approximately 1/(0.550.40×0.480.60)≈1.97×1/(0.55^0.40× 0.48^0.60)≈ 1.97× relative to a well prepared peer. A transformation that a data ready bank completes in 18 months will take Zions approximately 35 months under current infrastructure conditions. This is not a scheduling inconvenience; it represents a compounding competitive disadvantage. Remark 9.1 (Framework Application for Self Assessment). Any institution in a position analogous to Zions can apply the AITG rubric to its own 10-K disclosures and compare the resulting score against the AFC adjusted IASS∗ ceiling for its sector. The gap (GeffG_eff) and the IFS time shift multiplier jointly quantify both the competitive exposure and the adoption timeline under current infrastructure conditions. A systematic self assessment protocol is provided in ESM Part IV. 9.3.1 End to End Worked Example: Full Δ Pipeline for Zions Bancorporation A complete ten step VCB walk-through for Zions Bancorporation, demonstrating every computational stage from gap measurement through risk adjusted Δ , appears in Appendix D (Table 21). 9.4 JPMorgan vs. Zions: Head to Head Comparison Table 10 presents the head to head comparison. Both firms operate in the same industry (IASS∗ 9.38). The AITG framework explains the divergence entirely from structural factors visible in public data. Table 10: JPMorgan Chase vs. Zions Bancorporation: AITG Head to Head Metric JPMorgan Chase Zions Bancorporation Industry IASS∗ 9.38 9.38 AITG (raw) 8.22 3.80 Industry Relative (IR) Score 8.76 4.05 Effective Gap (GeffG_eff) 1.16 5.58 ADRI (Disruption Risk) 0.5 (Low) 2.6 (Moderate/Elevated) IFS Residual Multiplier 0.88 0.71 AI Stack Type Proprietary Vendor (nCino) Firm Scale Factor (Φf _f) 1.00 0.52 5 Year Δ (risk adj.) $20–98B $0.6–1.2B Competitive Trajectory Improving Deteriorating The Wide Gap Fallacy Confirmed Zions has a wider effective gap (5.58 vs. 1.16). JPMorgan has higher value density. Why: Firm Scale Factor, proprietary AI, and IFS differential compound over the hold period. The central lesson is precise: gap size is not investable signal without gap quality. Zions has 2.6 times more effective gap than JPMorgan. JPMorgan captures 30–40 times more dollar value from that smaller gap. The AITG framework’s multimodule architecture, IASS ceiling, VCB value density, ADRI urgency, and IFS execution risk, is required to explain this divergence. No single score maturity model can recover it. 9.5 Cross Industry Normalization: Why Raw AITG Scores Cannot Be Compared Directly The within industry comparison (JPMorgan vs. Zions) illustrates gap quality. The cross industry comparison illustrates the role of IASS normalization. Consider Zions Bancorporation (Commercial Banking, IASS=∗9.38^*=9.38, AITG 3.80) and UPS (Logistics, IASS=∗7.68^*=7.68, AITG 4.08). A naïve reading of the raw AITG scores would conclude UPS is the better positioned firm for AI transformation. That conclusion is wrong in every material dimension, and the IASS normalization mechanism recovers the correct reading. Table 11: Cross Industry Normalization: Zions Bancorporation vs. UPS Metric Zions Bancorp. UPS Step 1: Raw Score (Cross Industry Comparison Without Normalization) Industry Comm. Banking Logistics AITG (raw, 0–10) 3.80 4.08 Naïve conclusion: UPS is “more transformed.” Step 2: Apply IASS Ceiling (Cross Industry Normalization) Industry IASS∗ (AFC adjusted ceiling) 9.38 7.68 IR Score =AITG/IASS∗×10=AITG/IASS^*× 10 4.05 5.31 Effective Gap GeffG_eff 5.58 3.60 IR score still favors UPS, but the ceiling differential is now visible. Step 3: Apply Value Architecture (Why the IASS Ceiling Matters for Returns) Exit Multiple (MiM_i) 1.3×1.3× book 9×9× EBITDA Run Rate Labor Value Pool (est.) $1.16B/yr $47.4B/yr Φf _f (Firm Scale Factor) 0.52 0.80 IFS Residual Multiplier 0.71 0.71 ADRI (Disruption Urgency) 2.6 4.2 5 Year Δ (risk adjusted) $0.6–1.2B $30–61B AITG-VD (Value Density) 15.0–30.1× 27.3–55.4× Correct cross industry conclusion: UPS generates superior absolute value density, but for structural reasons undetectable without the IASS+VCB architecture: larger absolute revenue base, higher exit multiple, and larger firm scale. The three step decomposition in Table 11 reveals the exact mechanism by which the IASS ceiling shapes investment conclusions. Raw AITG alone (Step 1) produces a misleading comparison: UPS at 4.08 appears more transformed than Zions at 3.80, but this cross industry comparison is inadmissible without normalization. IR scoring (Step 2) corrects for structural ceiling differences but still does not capture why the investment opportunity differs. Only the full VCB + Φf _f architecture (Step 3) recovers the correct conclusion: UPS generates superior absolute value density due to its ∼27× 27× larger revenue base and higher exit multiple, while Zions faces higher ADRI urgency and subcritical firm scale (Φf=0.52 _f=0.52). The IR score is the correct metric for cross industry benchmarking; Value Density is the correct metric for investment selection. 9.6 Extended Case Studies: ESM Part I Twelve additional firm level AITG applications, UPS, HCA Healthcare, Salesforce, Ferguson Enterprises, Rockwell Automation, Goldman Sachs, Wells Fargo, ServiceNow, Target Corporation, CVS Health, Palo Alto Networks, and Ford Motor, are documented in the Electronic Supplementary Material (ESM Part I). Each case follows the standardized format used in the JPMorgan and Zions writeups above: operating context with public data citations, six dimension AITG scorecard, and VCB summary with Value Density tier. The cross company synthesis of all fourteen companies, including the 14 company summary table and the Value Density Paradox scatter plot (Figure 5), is presented in the following section. 9.7 Cross Company Synthesis: fourteen company Analysis Table 12: AITG Framework: fourteen company Comparison (Tier 1 Scores) Company AITG IASS∗ IR GeffG_eff ADRI IFS VD Range Tier Commercial Banking (IASS∗ = 9.38) JPMorgan Chase 8.22 9.38 8.76 1.16 0.5 0.90 9.5–46.0× 1 (HC) Wells Fargo 6.00 9.38 6.40 3.38 2.3 0.65 35.5–75.1× 1 Zions Bancorp. 3.80 9.38 4.05 5.58 2.6 0.73 15.0–30.1× 2 Investment Banking (IASS∗ = 10.39) Goldman Sachs 8.17 10.39 7.86 2.22 0.6 0.89 26.3–73.6× 1 (HC) Vertical SaaS (IASS∗ = 9.96) Salesforce 8.07 9.96 8.10 1.89 2.1 0.92 25.2–80.4× 1 (HC) ServiceNow 8.50 9.96 8.53 1.46 1.8 0.88 17.4–54.6× 1 Cybersecurity (IASS∗ = 9.65) Palo Alto Ntwks 7.83 9.65 8.11 1.82 1.2 0.93 23.6–78.0× 1 (HC) Healthcare Services (IASS∗ = 5.46, ψ=0.743ψ=0.743) HCA Healthcare 4.33 5.46 7.93 1.13 3.8 0.77 7.4–26.4× 1 CVS Health 4.83 5.46 8.85 0.63 3.1 0.72 5.0–19.7× 1 E Commerce / Digital Retail (IASS∗ = 8.64) Target 4.67 8.64 5.41 3.97 3.2 0.82 29.5–61.5× 1 Logistics / Industrial (IASS∗ = 7.24–7.68) UPS 4.08 7.68 5.31 3.60 4.2 0.71 27.3–55.4× 1 Ferguson PLC 4.00 7.24 5.52 3.24 4.2 0.79 21.8–45.0× 1 Rockwell Auto. 5.83 7.24 8.05 1.41 3.5 0.86 15.8–45.0× 1 Discrete Manufacturing (IASS∗ = 7.49) Ford Motor 4.67 7.49 6.23 2.82 4.8 0.65 16.1–39.9× 1 Scores as of early 2026. IASS∗ = AFC adjusted ceiling (C2026≈1.90C_2026≈ 1.90, GPT-5.2 December 2025). IR = AITG/IASS×∗10^*× 10. HC = High Confidence. ψ<1.0ψ<1.0 indicates binding regulatory ceiling (Healthcare). ADRI and VD from Monte Carlo pipeline (Section 7) with AFC adjusted ceilings (C2026≈1.90C_2026≈ 1.90). IFS is the composite implementation feasibility score; the VCB value multiplier uses IFSres (VTR/CRS/REG weighted geometric product, Table 10). VD ranges use the standardised 1.2% of revenue implementation cost proxy (Appendix D); ESM Part I case narratives use firm specific comprehensive cost estimates that produce lower VD multiples. Cross industry correlation GeffG_eff vs. VD midpoint: r=0.22r=0.22 (p=0.45p=0.45, N=14N=14), confirming the wide gap fallacy. Finding 1: The Wide Gap Fallacy, Confirmed at N=14N=14. Across all 14 companies, the cross industry correlation between effective gap (GeffG_eff) and Value Density midpoint is r=0.22r=0.22 (p=0.45p=0.45): weakly positive and statistically non significant, confirming that wider gaps do not reliably predict higher value density. The weak r reinforces that gap size alone is not an investable signal; firm scale, IFS, and exit multiple leverage dominate (Figure 5). The Spearman rank correlation (ρs=0.41 _s=0.41, p=0.14p=0.14) is similarly non significant. 01122334455660×1010×2020×3030×4040×5050×6060×7070×8080×Sweet Spot Geff∈[1.3, 3.0]G_eff∈[1.3,\,3.0]JPMZIONWFCGSCRMNOWPANWHCACVSTGTUPSFERGROKFEffective Gap GeffG_eff = IASS∗ −- AITGValue Density (VD = Δ / Implementation Cost)Theoretical VD envelopeCommercial BankingInvestment BankingVertical SaaSCybersecurityHealthcare ServicesE Commerce / RetailLogistics / IndustrialDiscrete Manufacturing Figure 5: The Value Density Paradox: Wider Gap ≠ Higher Return. Across 14 companies spanning 8 industries, the cross industry correlation between effective gap (GeffG_eff) and Value Density midpoint is r=0.22r=0.22 (p=0.45p=0.45), weakly positive and statistically non significant—far below the naïve expectation that wider gaps mechanically produce proportionally higher returns. The shaded green zone (Geff∈[1.3, 3.0]G_eff∈[1.3,\,3.0]) is the sweet spot where S curve inflection proximity and manageable IFS produce maximum VD. The orange dashed curve shows the theoretical VD envelope derived from the AITG framework’s mathematical structure; points with the highest VD (WFC, CRM, PANW, GS) reflect high IASS∗ ceilings or large revenue bases; points with the lowest VD (HCA, CVS) reflect near zero GeffG_eff and regulatory ceiling IFS suppression. Vertical lines are Monte Carlo P10–P90 VD ranges. This chart illustrates a mathematical property of the framework, not an empirically established causal finding. See Section 11 for the endogeneity caveat and IV research agenda. Full implementation cost assumptions, including calibration methodology and the 1.2% of revenue cross industry simplification, appear in Appendix D (Table 22). Sobol analysis confirms that implementation cost explains only ∼ 1% of VD variance; exit multiple (50%) and capture rate (44%) dominate. Finding 2: AITG and ADRI are Complementary, Not Redundant. The 14 company sample contains four high ADRI cases (Ford 4.8, UPS 4.2, Ferguson 4.2, Target 3.2) that span different gap and IASS configurations. In every case, ADRI captures a “transform now or face structural erosion” dynamic that the AITG gap alone does not. Palo Alto Networks (ADRI 1.2, low competitive risk from cyber AI moat) and Goldman Sachs (ADRI 0.6, regulatory entry barrier) demonstrate the opposite: firms where the AI transformation has been converted into structural competitive insulation. Neither finding is recoverable from AITG alone; both require the joint reading. Finding 3: The Disintermediation Risk Finding (JPM vs. Zions). JPMorgan and Zions operate in the same industry (AFC adjusted IASS∗ = 9.38 as of early 2026). JPMorgan (AITG 8.22) has an effective gap of 1.16 and ADRI of 0.5: a dominant incumbent that has nearly closed its own gap to the industry frontier. Zions (AITG 3.80) has an effective gap of 5.58 and ADRI of 2.6, facing active disintermediation pressure from both above (JPMorgan’s proprietary AI extending its lead) and below (AI native fintech lenders). As the capability ceiling rises, the divergence accelerates: JPMorgan’s proprietary stack (Φf=1.0 _f=1.0) captures the rising frontier, while Zions’s vendor dependency (Φf≤0.65 _f≤ 0.65) caps its value pool regardless of ceiling expansion. The AFC is therefore a structural moat accelerator for proprietary stack firms. The Zions pattern (vendor AI adoption, legacy data fragmentation, wide gap in a high ceiling industry) is observable in hundreds of regional banks, mid market insurers, and professional services firms. Finding 4: High AITG + Low ADRI = Sustainable Moat. Salesforce (AITG 8.07, early 2026) combined with ADRI of 2.1 indicates it has substantially converted AI transformation into a structural competitive moat via Agentforce (agentic CRM, launched Q4 2025) and Einstein 2.0 GA. Data network effects and switching costs (Moat =0.92=0.92) mean competitors cannot rapidly replicate the Data Cloud advantage. Finding 5: IFS Materially Alters Investment Conclusions. UPS carries a wide effective gap (Geff=3.60G_eff=3.60) and a respectable raw AITG of 4.08, yet its IFS of only 0.71 sharply suppresses the value that gap should generate (IASS∗=7.68IASS^*=7.68). The driver: Teamsters contract constraints impose binding limits on workforce AI augmentation deployment, and the Better & Bolder cost program reduced headcount faster than AI capabilities could be redeployed, creating structural adoption friction. Meanwhile, ADRI has risen to 4.2 as FedEx and Amazon Logistics accelerate AI enabled route optimization and last mile efficiency, compressing the window in which UPS can convert its gap into returns. This is a cautionary case: labor relations risk is an IFS dimension that can prevent an otherwise favorable gap from generating value. The consolidated parameter calibration register, documenting all free parameters with calibrated values, permissible ranges, and source rationale, appears in Appendix A (Table 16). 10 Robustness Summary The AITG framework operates at what Lo and Mueller (2010) classify as Level 3–4 uncertainty, where model outputs are sensitive to distributional assumptions and small parameter changes can produce large output variations. This makes systematic sensitivity analysis essential rather than optional. I conduct four robustness checks that confirm the stability of the framework’s outputs (full methodology and tables in ESM Part I). Monte Carlo weight sensitivity (M=10,000M=10,000 draws, ±5%± 5\,\% perturbations): the mean absolute rank shift across nine anchor industries is R¯s=0.19 R_s=0.19 positions, and no pair of industries exchanges rank with probability >5%>5\,\% OECD (2008). Rankings are determined by the data, not by weighting choices. Sobol first order sensitivity decomposes Var(AITG-VD)Var(AITG-VD): exit multiple 50 %; capture rate assumption 44 %; implementation cost 1 %; AITG gap score <<1 %; IFS 5 % Saltelli (2002). Exit multiple leverage and capture rate assumptions dominate raw gap size as uncertainty sources, directly explaining the wide gap fallacy. Normalization stability: substituting z score for min max normalization shifts IASS values by ≤0.40≤ 0.40 points with no rank changes. Geometric vs. linear aggregation: geometric scoring produces values 0.20–0.50 points lower for industries with uneven dimensional profiles (healthcare, construction) and negligibly different for uniform profiles (SaaS), with identical rank orderings in both cases. Convergent and discriminant validity: AITG rankings across the fourteen company cohort achieve rs=0.88r_s=0.88 (p<0.001p<0.001) correlation with Lightcast AI job posting share and agree directionally with Stanford HAI rankings Stanford HAI (2025) for 12 of 14 companies. Spearman correlation between AITG and revenue rank is rs=−0.28r_s=-0.28 (not significant, p=0.33p=0.33), confirming that AITG measures transformation state, not firm size. Full interrater reliability protocol (ICC target ≥0.85≥ 0.85 per dimension, ≥0.88≥ 0.88 composite) and cross sectional discriminability analysis are reported in ESM Part I. A pilot scoring by two independent analyst teams on the original seven company cohort yielded Kendall’s τ=0.82τ=0.82 for the three most objective dimensions (DIM, PAC, OAC) and τ=0.68τ=0.68 for the two most interpretation dependent dimensions (DAR, APR). The lower agreement on DAR and APR motivates rubric refinement before the full ICC study specified in the Research Agenda (item 4). 10.1 Consolidated Sensitivity Summary Table 13 consolidates key sensitivity results from ESM Part I. All output metrics are computed for the representative industrial distribution firm (Ferguson) unless otherwise noted. Table 13: Consolidated Sensitivity Analysis Parameter Base Range Tested Output Metric Δ Range ρ (CES) 5 [3,8][3,8] Δ ±12%± 12\% θi _i (AFC) industry ±50%± 50\% IASS∗ ±0.13± 0.13 pts λ (capture) 3.5 [2.0,5.0][2.0,5.0] η(g)η(g) at g=0.5g=0.5 [0.29,0.39][0.29,0.39] ψfloor _floor 0.15 [0.10,0.25][0.10,0.25] Min. capture ±0.05± 0.05 k (wave steepness) 0.38 [0.30,0.46][0.30,0.46] t^f t_f ±4.2± 4.2 mo Exit multiple 10×10× [8×,16×][8×,16×] AITG-VD ±42%± 42\% Impl. cost $8B ±30%± 30\% AITG-VD ±23%± 23\% IASS weights Table 2 ±5%± 5\% Rank shift R¯s=0.19 R_s=0.19 Note. Exit multiple and capture rate assumptions dominate VCB output variance (Sobol S1S_1: 50% and 44% respectively), consistent with the wide gap fallacy discussed in Section 7. CES ρ and IASS weights produce no rank order changes within tested ranges. 10.2 Retrospective Predictive Backtest: Nonfinancial Cohort I conduct a retrospective backtest on the ten nonfinancial companies in the fourteen firm cohort, structured as a holdout prediction exercise. AITG scores are reconstructed as of Q4 2021 using only information publicly available at that date (10-K/10-Q filings, investor presentations, and disclosed AI programs through December 2021). Financial outcomes are measured as the change in EBITDA margin from FY2021 to FY2023 (a two year forward window). I exclude the four financial firms because banks report efficiency ratios rather than EBITDA margins; within sector bank comparisons appear in Section 9.2. Retrospective scoring protocol. The 2021 AITG scores use the same dimensional rubric as the current state scores in Section 9, anchored exclusively to 2021 vintage data. The AI Frontier Coefficient is set to C2021=1.00C_2021=1.00 (pre ChatGPT baseline; no AFC adjustment required). All six dimensions are scored against evidence in the corresponding 10-K filing (fiscal years ending July 2021 through January 2022 depending on company fiscal calendar). Data tier classifications follow the same A/B/C/D protocol. Results. Table 14: Retrospective Backtest: AITG2021 vs. Realized EBITDA Margin Change (FY2021–FY2023) Company Sector AITG21 EBITDA%21 EBITDA%23 Δ Rev CAGR Notes Palo Alto Ntwks Cybersecurity 7.0 −-1.0 12.0 ++13.0p ++27.2% Salesforce Vertical SaaS 6.2 22.0 30.0 ++8.0p ++14.8% ServiceNow Vertical SaaS 5.8 20.0 26.0 ++6.0p ++23.3% Rockwell Auto. Logistics/Ind. 4.3 20.0 24.0 ++4.0p ++10.9% HCA Healthcare Healthcare Svcs 3.3 22.0 23.0 ++1.0p ++4.7% CVS Health Healthcare Svcs 2.6 5.8 5.5 −-0.3p ++10.7% Aetna drag Ferguson PLC Logistics/Ind. 2.3 12.0 13.0 ++1.0p ++12.3% Target E Comm./Retail 4.0 10.0 8.0 −-2.0p ++0.5% * UPS Logistics 3.0 19.0 16.0 −-3.0p −-3.3% * Ford Motor Discrete Mfg. 2.2 8.0 6.0 −-2.0p ++13.8% * EBITDA margin = Operating Income + D&A / Revenue. Revenue CAGR is 2 year (2021–2023). 10-K sources: PANW FY ends Jul; CRM FY ends Jan; ROK FY ends Sep; all others calendar year. *Confounded: Target by 2022 inventory writedown cycle; UPS by postpeak volume and 2023 labor contract; Ford by Argo AI closure and EV launch losses. These confounds are unrelated to AI adoption intensity and are not scored against the framework. Financial firms excluded; see Section 9.2. Spearman rank correlation between AITG2021 and subsequent two year EBITDA margin change is ρs=0.818 _s=0.818 (p<0.001p<0.001, n=10n=10). The correlation reflects monotone rank ordering across most of the cohort: every firm scoring above 5.0 expanded margins, and three of four firms scoring below 3.0 contracted. Revenue CAGR shows a weaker, insignificant correlation with AITG2021 (ρs=0.47 _s=0.47, p=0.14p=0.14), consistent with the framework’s design. AITG is calibrated to predict cost structure improvement and margin expansion, not top line growth. Comparison to simple baselines. Three alternative predictors provide context. (1) Revenue size: Spearman rank correlation between firm revenue and Δ margin is ρs=0.19 _s=0.19 (p=0.59p=0.59), confirming that AITG’s signal is not a proxy for firm scale. (2) Aggregator choice: the CES ablation in Section 10.3 shows that Leontief and additive aggregators produce identical ρs=0.818 _s=0.818 on this cohort, because no firm has a near zero dimensional score; the CES advantage manifests in the expanded 22 company simulation (ESM Table S6) where additive diverges from CES by 35–60% on Δ . (3) Null model: a constant prediction (cohort mean AITG for every firm) produces ρs=0 _s=0 by definition. The gap between 0 and 0.818 represents the discriminating power of the rubric scoring within the available sample. These comparisons are qualitatively informative but cannot substitute for the statistically powered panel validation described in the research agenda. Within sector rank order consistency is exact in all three sectors with multiple observations: Salesforce >> ServiceNow (SaaS), Rockwell >> Ferguson (Logistics/Industrial), HCA >> CVS (Healthcare), and in the banking cohort JPMorgan >> Wells Fargo >> Zions on efficiency ratio improvement. Goldman Sachs is the single directional failure (high AITG, deteriorating efficiency ratio), transparently explained by the 2022 investment banking drought and Marcus consumer lending writedowns, macro cyclical events orthogonal to AI transformation state. Intra sector directional accuracy is 4 of 4 nonconfounded pairs (100%). Limitations of this backtest. First, the sample (n=14n=14) is illustrative, not confirmatory. A statistically powered panel study would require n≥100n≥ 100 firms with quarterly AITG estimates over multiple cycles, as specified in the Research Agenda (Section 11.5). Second, retrospective scoring from the same team introduces potential look back bias; the cross rater reliability study in the Research Agenda (item 4) addresses this with independent analyst teams. Third, the 2021–2023 window overlaps with unusual macro cyclical events (post COVID demand normalization, aggressive rate hiking) that confound three of ten observations. Fourth, the framework does not separate the causal effect of AI transformation from incumbent endogeneity: firms with high AITG2021 scores may also possess superior baseline economics and management quality independently associated with margin improvement (the omitted variable problem discussed in Section 11). The backtest is therefore best interpreted as directional consistency evidence, not causal identification. However, the ρs=0.818 _s=0.818 result across a diverse cross industry sample is meaningfully stronger than a size or industry fixed effects baseline would predict, and motivates the larger validation program. Power analysis and expanded validation. A formal power analysis at the observed effect size (ρs=0.818 _s=0.818, α=0.05α=0.05) indicates that n=10n=10 provides power ≈0.80≈ 0.80, at the conventional threshold. Detecting a smaller effect (ρs=0.50 _s=0.50) would require n≥30n≥ 30. The n=14n=14 cohort (10 nonfinancial) is sufficient to detect the observed large effect but insufficient for reliable inference on moderate effects or subgroup analysis. The Research Agenda specifies expanded validation using n≥100n≥ 100 firms with quarterly AITG estimates, providing power >0.95>0.95 for ρs≥0.30 _s≥ 0.30 and enabling industry stratified analysis. 10.3 CES Aggregator Ablation At the n=10n=10 backtest scale, CES (ρ=5ρ=5), Leontief, and additive aggregators produce identical Spearman rank correlations (ρs=0.818 _s=0.818) because no firm has a near zero dimensional score. The CES advantage manifests when firms have a very weak dimension (<2.0<2.0) or at boundary conditions; the extended 22 company simulation (ESM Table S6) shows additive diverging from CES by 35–60% on Δ . Full ablation details appear in Appendix C. 11 Limitations and Research Agenda Observational design caveat. I calibrate this framework from observational data (public filings, benchmark scores, labor market statistics) without randomized treatment assignment, instrumental variables, or difference in differences identification. All associations between AITG scores and financial outcomes in Sections 10–10.2 are correlational evidence of construct validity, not causal estimates. The backtest rank correlation (ρs=0.818 _s=0.818) is consistent with the framework’s predictions but cannot rule out omitted variable explanations (e.g., unobserved management quality, strategic complementarities). Establishing causal identification requires the panel design described in the Research Agenda (item 1). 11.1 Complete Assumptions Register The framework embeds 25 structural assumptions classified as: F (functional form)P (parameter calibration)B (behavioral), and D (data). The complete enumerated register, including stress test results for each assumption, is provided in ESM Part I. Key structural assumptions are discussed in Section 11.2 below. 11.2 Material Assumptions and Failure Conditions Assumption 11.1 (Industry ceiling stability). IASS base scores are sufficiently stable over a 12–24 month underwriting horizon to support investment decisions. Failure condition. Rapid regulatory change (e.g., an EU AI Act amendment restricting previously permitted use cases) or a breakthrough capability shift could materially alter a sector’s IASS within the horizon. The AFC versioning protocol mitigates this by requiring IASS recalibration when ΔCt>15% C_t>15\,\% or when regulatory events with estimated impact >1.0>1.0 IASS points occur. Assumption 11.2 (Capture rate portability). Published value pool capture rates (κ¯p κ_p) estimated from larger firms are applicable to mid market companies with appropriate scaling. Failure condition. Smaller companies may have less standardized processes, reducing addressable surface area at any given AITG score. The CES bottleneck aggregator (Eq. 23) should substantially capture this effect; IFS data readiness scores provide a secondary adjustment. Assumption 11.3 (IFS scores are elicitable). The five IFS risk factors can be reliably scored from diligence evidence with acceptable interrater reliability. Failure condition. Organizational change capacity, the highest weight IFS factor, is the most subjective and most susceptible to management coaching during diligence. Scoring rubrics anchor to objective evidence (prior transformation track record, attrition data, system migration success rates) rather than management assertion. Assumption 11.4 (AI frontier is continuous). The AI Capability Index CtC_t increases continuously and predictably within the AFC scenario range. Failure condition. Discontinuous events, a large scale regulatory freeze on AI deployment, a safety driven pause in model training, or a rapid architectural breakthrough that changes capability dynamics, could invalidate the AFC scenario distribution. Mitigation: the conservative AFC scenario explicitly prices in a capability slowdown; the $alpha$-max cap prevents unbounded ceiling expansion. 11.3 Known Limitations Endogeneity and Reverse Causality. A critical limitation of the current empirical illustrations is the risk of endogeneity and reverse causality. The AITG score measures the current state of transformation, but dominant incumbents (e.g., JPMorgan Chase) possess both superior baseline economics and the massive free cash flow required to fund proprietary AI programs. Consequently, observed correlations between AITG Value Density and subsequent margin expansion may suffer from omitted variable bias. Firms might be generating high value because they are deploying AI, or they might be deploying AI because they are already highly profitable “superstar firms.” Separating the causal effect of AI transformation from the underlying incumbent advantage requires quasi experimental methods not present in this conceptual baseline. Private company scoring challenge. public company scoring is constrained by disclosure limitations. The Tier 1 scores presented in Section 9 carry wider uncertainty bands than diligence grade scores and should not be used for investment decisions without supplementary primary data collection via the standardized management survey and IT infrastructure checklist (described in the companion implementation document). Goodhart’s Law and evaluator rotation. Once AITG becomes a target in financial or regulatory contexts, it will be susceptible to gaming Goodhart (1984). The anti gaming architecture (bottleneck aggregation, evidence tier penalties, cross validation across disclosure language, hiring, and patent signals) reduces but does not eliminate this risk. The evaluator rotation protocol (Section 4.6) specifies: (i) scoring teams rotate annually, with no team rescoring the same company within a 24 month window; (i) parameter freeze windows are announced before each scoring cycle and held fixed through that cycle’s evaluation; (i) all scoring inputs are archived with evidence tier classification and available for third party audit on request. Preregistration of parameter ranges in advance of validation studies is a research agenda commitment (Section 11.5). Fixed CES elasticity. The CES bottleneck aggregator uses a single substitution parameter ρ=5ρ=5 (σ=1/6σ=1/6) across all value pools and firms. This may overstate bottleneck severity for enablers that are highly substitutable, or understate it for enablers with extreme complementarity. enabler specific σ estimation and variable elasticity forms (e.g., variable elasticity σ(k)σ(k) forms to reflect changing substitutability as capabilities and complements improve) are identified as theoretical extensions. Robustness across ρ∈[3,8]ρ∈[3,8] is confirmed in ESM Table S3; expanding this range and estimating nested CES structures are priorities for the panel study. Look ahead bias and parameter stability. The current framework is calibrated on a single cohort observed over 2021–2023. Several risk factors apply: (i) AFC scenario thresholds were developed using 2025–2026 capability data and applied retrospectively to the 2021 backtest, introducing potential look ahead bias in AFC adjusted IASS* comparisons; (i) value pool share coefficients were calibrated from broader industry benchmarks, not from the backtest cohort, but parameter selection was informed by the same period’s disclosed outcomes; (i) the backtest is in sample with respect to the parameter calibration period. Future validation should implement rolling origin or genuinely holdout designs with parameter ranges preregistered before data access. 11.4 Labor Market, Distributional, and Ethical Considerations The VCB models value creation from an investor’s perspective and is explicitly not a welfare instrument. It does not model distributional effects on workers, communities, or suppliers, and should not be mistaken for such. The displacement reinstatement literature Acemoglu and Restrepo (2019a, 2022) establishes that automation can reduce labor demand in affected occupations even while raising aggregate productivity. A complete welfare analysis of AI transformation requires incorporating these effects, which is beyond this paper’s scope. Distributional implications of the value creation architecture. Three aspects of the framework’s structure have distributional implications that users should be aware of. First, the Firm Scale Factor Φf _f mechanically concentrates value density advantage at large incumbents with substantial proprietary data assets. In high IASS industries, this implies that medium market competitors may be disadvantaged even if they successfully execute transformation programs, because the frontier is partly defined by data advantages they cannot replicate. This reinforces the winner take most dynamics documented in Autor et al. (2020). Second, the ADRI displacement risk score, while calibrated at the firm level, aggregates occupation level displacement risk across an industry’s entire workforce. Firms should not treat ADRI reduction as value creation without accounting for the labor force transitions implied by the transformation program. Third, the WAR (Workforce Augmentation Rating) dimension rewards human AI complementarity, providing a structural scoring incentive toward augmentation type deployments over pure headcount reduction programs; this is the framework’s primary alignment mechanism toward socially constructive AI adoption, but its weighting (w=0.20w=0.20 in the raw AITG composite) is moderate rather than dominant. Recommendation. Because the AITG framework informs capital allocation decisions with material workforce consequences, any institutional deployment should: (a) document the WAR score interpretation and confirm the transformation design is oriented toward augmentation where feasible; (b) supplement AITG analysis with a distributional impact assessment appropriate to the user’s governance obligations; and (c) treat ADRI implied competitive urgency as distinct from the urgency to minimize worker displacement, which requires separate governance attention. These are not constraints on the framework’s analytical use, but complements to responsible institutional deployment. Safety, cyber, and negative value risks. The VCB quantifies positive value creation from AI deployment. It does not currently model negative value from AI related incidents, regulatory penalties, or cybersecurity failures. In high IASS, high automation settings, the risk of adversarial attacks on AI systems, model failures in production, and privacy violations can generate material negative enterprise value. Future extensions of the framework should incorporate an “AI harm adjusted value” view that accounts for these safety and security exposures. 11.5 Research Agenda Nine empirical research priorities emerge from this framework, organized around causal validation, technical refinement, and measurement quality: 1. Causal identification via instrumental variables: exploiting exogenous variation in cloud incentive grants or differential open source LLM exposure to isolate the causal effect of AI adoption. 2. Large panel validation: multiyear panel (n≥100n≥ 100) with quarterly AITG estimates testing the mid gap optimality property with out of sample predictions. 3. AFC calibration and θi _i estimation: refining industry specific parameters with backward looking LLM exposure analysis and formal standard errors. 4. Interrater reliability study: multiple independent analyst teams scoring the same companies; target ICC >0.75>0.75 per factor. Highest empirical priority for institutional deployment. 5. Competitive displacement measurement: linking ADRI scores to subsequent market share and margin outcomes at the firm level. 6. Cross country calibration: adapting IASS from O*NET/OEWS to ESCO (EU) and equivalent national databases. 7. Preregistration, rolling origin validation, and replication bundle: parameter locking before data access, holdout firms, placebo tests, and public reproducibility code. 8. Variable elasticity and nested CES: hierarchical enabler nesting with variable σ forms reflecting observed deployment sequencing. 9. Hierarchical Bayesian parameter estimation: formal priors on θi _i, κ¯p κ_p, and IFS elasticities with partial pooling across industries. Highest technical priority. Extended discussion of each priority, including specific methodological designs and data requirements, appears in Appendix E. 12 Conclusion I introduced the AI Transformation Gap Index, a composite scoring framework that addresses three recurrent failures in AI maturity assessment: limited cross industry comparability, static capability ceilings, and weak linkage between qualitative scores and financial outcomes. The IASS constructs a reproducible, task structure anchored industry ceiling. The AFC converts that ceiling into a dynamically consistent frontier as AI capabilities advance. The VCB translates gap scores into dollar quantified opportunity through bottleneck gated value pools, a low elasticity CES aggregator, and an explicit ramp and cost model. The ADRI scores competitive hazard from inaction. The IFS endogenizes organizational and data readiness constraints into adoption timing and steepness rather than applying them as post hoc scalars. Three properties emerge from the 14 company illustrative calibration. First, the wide gap fallacy: the S curve and CES architecture jointly imply that the widest gap does not produce the highest value density; the optimum falls at 35–65 % of the AFC adjusted frontier in high IASS sectors. Second, ADRI and AITG are complementary, because competitive hazard requires the joint signal of gap size, diffusion velocity, and structural moat. Third, IFS is a primary driver: the contrast between IFS=0.71IFS=0.71 (UPS) and IFS=0.91IFS=0.91 (Salesforce) materially dominates value density outcomes. These properties are theoretical results of the framework’s architecture, directionally consistent with the illustrative evidence but not empirically established findings. The small sample backtest (ρs=0.818 _s=0.818, n=10n=10; revenue size baseline ρs=0.19 _s=0.19) provides a construct validity signal, not causal identification. The nine item research agenda (Section 11.5) specifies what converting these properties into causally identified claims requires: IV designs, rolling origin validation with preregistered parameters, interrater reliability studies, nested CES estimation, and Bayesian reestimation. These are prerequisites for institutional deployment, not optional refinements. The measurement problem is consequential and tractable. I have provided the formal architecture; empirical validation is the immediate next step. Acknowledgments The Python reproducibility scripts (aitg_monte_carlo.py, aitg_stress_test.py) and the Excel companion workbook builder (build_aitg_excel_revised_fixed.py) were developed with assistance from Claude Code (Anthropic), an AI coding tool. All mathematical specifications, model design decisions, scoring rubrics, empirical calibrations, and interpretive judgments are solely those of the author. Code and Data Availability Full reproducibility code is available at https://github.com/deanbrr/aitg-framework. The repository contains: • aitg_monte_carlo.py – Monte Carlo sensitivity analysis (Layers 1–4), custom firm scoring via pre computed dimension scores (--score) or the 25 question management survey (--survey), and the Spearman backtest. • aitg_stress_test.py – thirteen deterministic stress test assertions covering all pipeline stages. • build_aitg_excel_revised_fixed.py – generates the Excel companion workbook with survey input sheet, company scorer, IASS calculator, IFS calculator, and VCB model. No external dependencies are required beyond the Python standard library (openpyxl for the Excel builder only). The scripts, Excel workbook builder, and this paper’s LaTeX source are also available as ancillary files accompanying the arXiv submission. Supplementary Materials The Electronic Supplementary Material (ESM) is included as an appendix following the references. It contains: • ESM Table S1: Pre computed IASS calibrations for all 22 industry verticals. • ESM Part I: Full firm level case studies for UPS, HCA Healthcare, Salesforce, Ferguson, Rockwell, Goldman Sachs, Wells Fargo, ServiceNow, Target, CVS Health, Palo Alto Networks, and Ford Motor, including six dimension scorecards and VCB summaries. • ESM Part I: Full Monte Carlo, Sobol, and validity analysis with tables. • ESM Part IV: Public data availability map; data tier classification; 25 question management survey instrument; VCB exit multiple sensitivity tables; AFC benchmark suite; Excel companion model documentation. References Acemoglu (2024) Acemoglu, D. (2024). The simple macroeconomics of AI. Economic Policy, 40(121):13–58. DOI: 10.3386/w32487. Acemoglu and Autor (2011) Acemoglu, D. and Autor, D. (2011). Skills, tasks and technologies: Implications for employment and earnings. In Card, D. and Ashenfelter, O. (eds.), Handbook of Labor Economics, Vol. 4B, p. 1043–1171. Elsevier. Acemoglu and Restrepo (2018) Acemoglu, D. and Restrepo, P. (2018). The race between man and machine: Implications of technology for growth, factor shares, and employment. American Economic Review, 108(6):1488–1542. DOI: 10.1257/aer.20160696. Acemoglu and Restrepo (2019a) Acemoglu, D. and Restrepo, P. (2019a). Automation and new tasks: How technology displaces and reinstates labor. Journal of Economic Perspectives, 33(2):3–30. DOI: 10.1257/jep.33.2.3. Acemoglu and Restrepo (2022) Acemoglu, D. and Restrepo, P. (2022). Tasks, automation, and the rise in US wage inequality. Econometrica, 90(5):1973–2016. DOI: 10.3982/ECTA19815. AlphaSense (2024) AlphaSense (2024). AlphaSense Earnings Call Analytics and NLP Platform. AlphaSense, Inc. Autor et al. (2003) Autor, D.H., Levy, F., and Murnane, R.J. (2003). The skill content of recent technological change: An empirical exploration. Quarterly Journal of Economics, 118(4):1279–1333. Autor and Dorn (2013) Autor, D.H. and Dorn, D. (2013). The growth of low-skill service jobs and the polarization of the US labor market. American Economic Review, 103(5):1553–1597. DOI: 10.1257/aer.103.5.1553. Autor et al. (2020) Autor, D., Katz, L., Patterson, C., and Van Reenen, J. (2020). The fall of the labor share and the rise of superstar firms. Quarterly Journal of Economics, 135(2):645–709. DOI: 10.1093/qje/qjaa004. Babina et al. (2024) Babina, T., Fedyk, A., He, A., and Hodson, J. (2024). Artificial intelligence, firm growth, and product innovation. Journal of Financial Economics, 151:103745. DOI: 10.1016/j.jfineco.2023.103745. Bass (1969) Bass, F.M. (1969). A new product growth for model consumer durables. Management Science, 15(5):215–227. BEA (2023) Bureau of Economic Analysis (2023). GDP-by-Industry Accounts, Table 7: Components of Value Added by Industry. U.S. Department of Commerce. https://w.bea.gov/. BLS OEWS (2023) Bureau of Labor Statistics (2023). Occupational Employment and Wage Statistics (OEWS). U.S. Department of Labor. https://w.bls.gov/oes/. Bloom et al. (2016) Bloom, N., Sadun, R., and Van Reenen, J. (2016). Management as a Technology? NBER Working Paper No. 22327. DOI: 10.3386/w22327. Bonney et al. (2024) Bonney, K. et al. (2024). Tracking firm use of AI in real time. NBER Working Paper No. 32319 / Census CES-WP-24-16. Brynjolfsson and Hitt (1996) Brynjolfsson, E. and Hitt, L. (1996). Paradox lost? Firm-level evidence on the returns to information systems spending. Management Science, 42(4):541–558. DOI: 10.1287/mnsc.42.4.541. Brynjolfsson and Hitt (2000) Brynjolfsson, E. and Hitt, L.M. (2000). Beyond computation: Information technology, organizational transformation and business performance. Journal of Economic Perspectives, 14(4):23–48. DOI: 10.1257/jep.14.4.23. Brynjolfsson and McElheran (2016) Brynjolfsson, E. and McElheran, K. (2016). The rapid adoption of data-driven decision-making. American Economic Review, 106(5):133–139. DOI: 10.1257/aer.p20161016. Brynjolfsson et al. (2021) Brynjolfsson, E., Rock, D., and Syverson, C. (2021). The productivity J-curve: How intangibles complement general purpose technologies. American Economic Journal: Macroeconomics, 13(1):333–372. DOI: 10.1257/mac.20180386. Brynjolfsson et al. (2025) Brynjolfsson, E., Li, D., and Raymond, L. (2025). Generative AI at work. Quarterly Journal of Economics, 140(2):889–942. DOI: 10.1093/qje/qjae044. Compustat (2024) Compustat / S&P Global Market Intelligence (2024). Compustat North America Fundamentals Annual. S&P Global Market Intelligence. David (1990) David, P.A. (1990). The dynamo and the computer: An historical perspective on the modern productivity paradox. American Economic Review, 80(2):355–361. Deloitte (2026) Deloitte (2026). State of AI in the Enterprise 2026. Deloitte AI Institute. Eloundou et al. (2024) Eloundou, T., Manning, S., Mishkin, P., and Rock, D. (2024). GPTs are GPTs: Labor market impact potential of LLMs. Science, 384(6702):1306–1308. DOI: 10.1126/science.adj0998. Epoch AI (2025) Epoch AI (2025). Epoch Capabilities Index (ECI). https://epoch.ai/benchmarks/eci. EU AI Act (2024) European Parliament and Council (2024). Regulation (EU) 2024/1689 of the European Parliament and of the Council laying down harmonised rules on artificial intelligence. Official Journal of the European Union. NIST (2023) National Institute of Standards and Technology (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1. U.S. Department of Commerce. Felten et al. (2021) Felten, E.W., Raj, M., and Seamans, R. (2021). Occupational, industry, and geographic exposure to artificial intelligence. Strategic Management Journal, 42(12):2195–2217. DOI: 10.1002/smj.3286. Ferguson (2025) Ferguson Enterprises (2025). Annual Report and Form 10-K, Fiscal Year 2025. Ferguson Enterprises, Inc. Fierce Healthcare (2024) Fierce Healthcare (2024). How HCA Healthcare has deployed AI across its hospitals. https://w.fiercehealthcare.com, July 2024. Frey and Osborne (2017) Frey, C.B. and Osborne, M.A. (2017). The future of employment: How susceptible are jobs to computerisation? Technological Forecasting and Social Change, 114:254–280. DOI: 10.1016/j.techfore.2016.08.019. Gibbons and Henderson (2012) Gibbons, R. and Henderson, R. (2012). Relational contracts and organizational capabilities. Organization Science, 23(5):1350–1364. DOI: 10.1287/orsc.1110.0715. Goodhart (1984) Goodhart, C.A.E. (1984). Monetary theory and practice. MacMillan, London. Goos et al. (2014) Goos, M., Manning, A., and Salomons, A. (2014). Explaining job polarization: Routine-biased technological change and offshoring. American Economic Review, 104(8):2509–2526. HCA Healthcare (2024) HCA Healthcare (2024). Fourth Quarter and Full Year 2024 Earnings Release. HCA Healthcare, Inc. Hoffmann et al. (2022) Hoffmann, J. et al. (2022). Training compute-optimal large language models. Advances in Neural Information Processing Systems, 35. arXiv:2203.15556. Jones and Tonetti (2020) Jones, C.I. and Tonetti, C. (2020). Nonrivalry and the economics of data. American Economic Review, 110(9):2819–2858. DOI: 10.1257/aer.20191330. Kaplan et al. (2020) Kaplan, J. et al. (2020). Scaling laws for neural language models. arXiv:2001.08361. Krakowski et al. (2023) Krakowski, S., Luger, J., and Raisch, S. (2023). Artificial intelligence and the changing sources of competitive advantage. Strategic Management Journal, 44(6):1425–1452. DOI: 10.1002/smj.3387. Lightcast (2024) Lightcast (2024). Labor Market Analytics Platform: Job Postings and Skills Data. Lightcast LLC. McElheran et al. (2024) McElheran, K. et al. (2024). AI adoption in America: Who, what, and where. Journal of Economics & Management Strategy. Census CES-WP-23-48. McKinsey (2025) McKinsey & Company (2025). The State of AI 2025: Agents, Innovation, and Transformation. McKinsey Global Institute. Kwa et al. (2025) Kwa, T., West, M., and others (2025). Measuring AI ability to complete long tasks. METR (Model Evaluation and Threat Research). arXiv:2503.14499. Nardo and Saisana (2005) Nardo, M. et al. (2005). Handbook on Constructing Composite Indicators: Methodology and User Guide. OECD Statistics Working Papers 2005/03. OECD Publishing. OECD (2008) OECD/EC JRC (2008). Handbook on Constructing Composite Indicators: Methodology and User Guide. OECD Publishing, Paris. DOI: 10.1787/9789264043466-en. O*NET (2024) O*NET Resource Center (2024). O*NET OnLine: Occupational Information Network. U.S. Department of Labor, Employment and Training Administration. https://w.onetonline.org/. Rockwell Automation (2024) Rockwell Automation (2024). Annual Report on Form 10-K, Fiscal Year 2024. Rockwell Automation, Inc. Rockwell (2024b) Rockwell Automation (2024b). State of Smart Manufacturing Report 2024. Rockwell Automation, Inc. Rogers (2003) Rogers, E.M. (2003). Diffusion of Innovations (5th edition). Free Press, New York. Salesforce (2025) Salesforce, Inc. (2025). Fourth Quarter and Fiscal Year 2025 Results. Press release, February 26, 2025. Saltelli (2002) Saltelli, A. (2002). Making best use of model evaluations to compute sensitivity indices. Computer Physics Communications, 145(2):280–297. SEC EDGAR (2024) U.S. Securities and Exchange Commission (2024). EDGAR Electronic Data Gathering, Analysis, and Retrieval System. https://w.sec.gov/edgar.shtml. Shrout and Fleiss (1979) Shrout, P.E. and Fleiss, J.L. (1979). Intraclass correlations: Uses in assessing rater reliability. Psychological Bulletin, 86(2):420–428. Stanford HAI (2025) Stanford Human-Centered Artificial Intelligence (2025). AI Index Report 2025. Stanford University. https://aiindex.stanford.edu/. Syverson (2011) Syverson, C. (2011). What determines productivity? Journal of Economic Literature, 49(2):326–365. DOI: 10.1257/jel.49.2.326. UPS (2024) United Parcel Service (2024). 2024 Annual Report and Form 10-K. United Parcel Service, Inc. Webb (2020) Webb, M. (2020). The impact of artificial intelligence on the labor market. Stanford/SSRN Working Paper. DOI: 10.2139/ssrn.3482150. Xie et al. (2024) Xie, Q. et al. (2024). FinBen: A holistic financial benchmark for large language models. arXiv preprint arXiv:2402.12659. Available at: https://arxiv.org/abs/2402.12659. Mazziotta and Pareto (2013) Mazziotta, M. and Pareto, A. (2013). Methods for constructing composite indices: One for all or all for one? Rivista Italiana di Economia, Demografia e Statistica, 67(2):67–80. JPMorgan Chase CNBC (2025) Breuninger, K. (2025). Here’s JPMorgan Chase’s blueprint to become the world’s first fully AI-powered megabank. CNBC, September 30, 2025. https://w.cnbc.com/2025/09/30/jpmorgan-chase-fully-ai-connected-megabank.html McKinsey and JPMorgan (2025) Waldron, D., Buehler, K., and Cooke, M. (2025). JPMorgan Chase’s Derek Waldron on building an AI-first bank culture. McKinsey & Company, October 2025. https://w.mckinsey.com/industries/financial-services/our-insights/jpmorgan-chases-derek-waldron-on-building-an-ai-first-bank-culture Zions Bancorporation (2026) Zions Bancorporation, N.A. (2026). Zions Bancorporation, National Association Reports Fourth Quarter Financial Results. Press release, January 20, 2026. https://w.prnewswire.com/news-releases/zions-bancorporation-national-association-reports-fourth-quarter-financial-results-302665710.html nCino Zions (2025) nCino, Inc. (2025). Zions Bancorporation selects nCino as tech foundation for lending. Press release, April 15, 2025. https://w.ncino.com/news/zions-bancorporation-implement-ncino-intelligent-platform-technology-foundation-lending-transformation Lightcast (2025) Lightcast (2025). The Generative AI Job Market: 2025 Data Insights. Lightcast LLC. https://lightcast.io/resources/blog/the-generative-ai-job-market-2025-data-insights Digalakis et al. (2025) Digalakis, V., Jr., Krishnan, R., Fernandez, G.M., and Orfanoudaki, A. (2025). ML Compass: Navigating capability, cost, and compliance trade-offs in AI model deployment. arXiv:2512.23487. Kinniment et al. (2024) Kinniment, M., Sato, L.J.K., Du, H., Goodrich, B., and others (2024). Evaluating language-model agents on realistic autonomous tasks. arXiv:2312.11671. Goldman Sachs (2025) Goldman Sachs Group, Inc. (2025). 2024 Annual Report and Form 10-K. Goldman Sachs Investor Relations. Wells Fargo (2025) Wells Fargo & Company (2025). 2024 Annual Report and Form 10-K. Wells Fargo Investor Relations. ServiceNow (2025) ServiceNow, Inc. (2025). 2024 Annual Report and Form 10-K. ServiceNow Investor Relations. Target (2025) Target Corporation (2025). 2024 Annual Report and Form 10-K. Target Investor Relations. CVS Health (2025) CVS Health Corporation (2025). 2024 Annual Report and Form 10-K. CVS Health Investor Relations. Palo Alto Networks (2025) Palo Alto Networks, Inc. (2025). 2025 Annual Report and Form 10-K. Palo Alto Networks Investor Relations. Ford Motor (2025) Ford Motor Company (2025). 2024 Annual Report and Form 10-K. Ford Motor Company Investor Relations. Standish Group (2015) Standish Group (2015). CHAOS Report 2015. Standish Group International, Boston, MA. Acemoglu et al. (2022) Acemoglu, D., Anderson, G., Beede, D., Buffington, C., Childress, E., Dinlersoz, E., Foster, L., Goldschlag, N., Haltiwanger, J., Kroff, Z., Restrepo, P., and Zolas, N. (2022). Automation and the Workforce: A Firm-Level View from the 2019 Annual Business Survey. NBER Working Paper No. 30659, November 2022. https://w.nber.org/papers/w30659. Autor (2024) Autor, D. (2024). Applying AI to Rebuild Middle Class Jobs. NBER Working Paper No. 32140, February 2024. https://w.nber.org/papers/w32140. Hulten (1978) Charles R. Hulten. Growth accounting with intermediate inputs. The Review of Economic Studies, 45(3):511–518, 1978. https://doi.org/10.2307/2297264. Sato (1967) Kazuo Sato. A two-level constant-elasticity-of-substitution production function. The Review of Economic Studies, 34(2):201–218, 1967. https://doi.org/10.2307/2296209. Gwet (2014) Kilem L. Gwet. Handbook of Inter-Rater Reliability: The Definitive Guide to Measuring the Extent of Agreement Among Raters. Advanced Analytics, LLC, 2014. Cox (1972) Cox, D.R. (1972). Regression models and life-tables. Journal of the Royal Statistical Society: Series B (Methodological), 34(2):187–220. DOI: 10.1111/j.2517-6161.1972.tb00899.x. Gartner (2023) Gartner, Inc. (2023). Gartner AI Maturity Model. Gartner Research. Westerman et al. (2014) Westerman, G., Bonnet, D., and McAfee, A. (2014). Leading Digital: Turning Technology into Business Transformation. Harvard Business Review Press, Boston, MA. Krippendorff (2018) Krippendorff, K. (2018). Content Analysis: An Introduction to Its Methodology (4th edition). Sage Publications, Thousand Oaks, CA. Lo and Mueller (2010) Lo, A. W. and Mueller, M. T. (2010). WARNING: Physics envy may be hazardous to your wealth! Journal of Investment Management, 8(2):13–63. arXiv: 1003.2688. OpenAI (2025a) OpenAI (2025a). Introducing GPT-5. https://openai.com/index/introducing-gpt-5/. Accessed August 2025. OpenAI (2025b) OpenAI (2025b). Introducing GPT-5.2. https://openai.com/index/introducing-gpt-5-2/. Accessed December 2025. Avasant (2024) Avasant (2024). IT spending as a percentage of revenue by industry, company size, and region. https://avasant.com/report/it-spending-as-a-percentage-of-revenue-by-industry-company-size-and-region/. IBM (2025) IBM (2025). AI spending expected to surge 52% beyond IT budgets as retail brands embrace enterprise-wide innovation. https://newsroom.ibm.com/2025-01-07-ibm-study-ai-spending-expected-to-surge-52-beyond-it-budgets-as-retail-brands-embrace-enterprise-wide-innovation. Core Appendices Material supporting the main text of the AITG paper. Appendix A Notation and Parameter Reference This appendix consolidates the notation reference table and the parameter calibration register from the main text. A.1 Notation and Parameter Summary Table 15: Consolidated Notation Reference Symbol Definition Range / Units Industry level constructs IASSiIASS_i Industry AI Susceptibility Score (base, before AFC) [0,10][0,10] IASSi∗IASS^*_i AFC adjusted ceiling: min(IASSi⋅(1+θi(Ct−C0)),αmaxIASSi) \! (IASS_i·(1+ _i(C_t-C_0)),\ _ \,IASS_i ) [0,αmax⋅IASSi][0,\; _ \!·\!IASS_i] θi _i Industry sensitivity to capability growth; anchored to ΔATDi _i [0.05, 1.50][0.05,\,1.50] αmax _ Cap on AFC multiplier (default 1.35) [1.0,∞)[1.0,∞) CtC_t AI Capability Index at time t (externally observed) [1,∞)[1,∞) C0C_0 Baseline capability index (normalization constant) =1.0=1.0 CADRiCADR_i Competitive AI Diffusion Rate (IASS subcomponent) [0,10][0,10] ψi _i RFF hard floor: min(1,(RFFi/5)1.5) (1,(RFF_i/5)^1.5) [0,1][0,1] RFFiRFF_i Regulatory Friction Factor composite (IASS subcomponent) [0,10][0,10] Firm level scores AITGfrawAITG^raw_f Raw AITG score: arithmetic mean of six dimension scores [0,10][0,10] IRfIR_f Industry Relative score: (AITGfraw/IASSi∗)×10(AITG^raw_f/IASS^*_i)× 10 [0,10][0,10] Geff,fG_eff,f Effective gap: max(0,IASSi∗−AITGfraw) \! (0,\ IASS^*_i-AITG^raw_f ) [0,αmax⋅IASSi][0,\; _ \!·\!IASS_i] ADRIf,tADRI_f,t AI Disruption Risk Index (Eq. 19) [0,10][0,10] λf(t) _f(t) ADRI competitive hazard intensity (Eq. 20): marginal probability of displacement per year events/yr Λf(T) _f(T) Cumulative hazard: ∫0Tλf(t)t _0^T _f(t)\,dt; P(displaced by T)≈1−e−Λf(T)P(displaced by T)≈ 1-e^- _f(T) dimensionless T ADRI normalization constant: 100 (ADRI score == % per year; λf=ADRI/100 _f=ADRI/100) dimensionless δt _t AFC urgency multiplier: 1+0.5min(Ct/C0−1,1)1+0.5 (C_t/C_0-1,1) [1.0,1.5][1.0,1.5] MoatfMoat_f Structural defensibility (switching costs, network effects, …) [0,1][0,1] Φf _f Firm Scale Factor (logistic, Eq. 21) [0,1][0,1] Si∗S^*_i Industry critical scale threshold for proprietary AI advantage $B revenue Adoption trajectory LwL_w Incremental logistic asymptote for wave w; L1=4.0L_1=4.0, L2=3.5L_2=3.5, L3=2.5L_3=2.5 (cumulative ceiling =10.0=10.0) AITG units kw,fk_w,f Wave steepness for firm f (IFS adjusted): kw,f=kw,base⋅ϕOCC(δOCC,f)⋅ϕDR(δDR,f)k_w,f=k_w,base· _OCC( _OCC,f)· _DR( _DR,f) mo-1 t0,w,ft_0,w,f Wave inflection timing (IFS adjusted midpoint) months t^f t_f Current time on adoption curve (inverse mapping) months t50,ft_50,f Value ramp inflection (midpoint of Rf(t)R_f(t) logistic) months CES bottleneck aggregator (Eq. 23) ρ CES substitution parameter: ρ>0ρ>0 implies complementarity [3,8][3,8], default 5 σ Elasticity of substitution: σ=1/(1+ρ)σ=1/(1+ρ). At ρ=5ρ=5: σ=1/6σ=1/6. (0,1)(0,1) ede_d Normalized dimension score: max(sd/10, 0.01) (s_d/10,\,0.01) (0,1](0,1] bpb_p CES bottleneck factor for value pool p (0,1](0,1] αd _d Dimension importance weight (∑αd=1Σ _d=1) [0,1][0,1] Note: ρ=5ρ=5 and σ=1/6σ=1/6 are mutually consistent under the adopted ACMS form. As ρ→∞ρ→∞, bp→mindedb_p→ _de_d (Leontief); as ρ→0ρ→ 0, bp→b_p→ Cobb-Douglas. Value Creation Bridge VprunV_p^run Annual run rate value pool (Eq. 22) $M/yr gfg_f Gap fraction: Geff,f/10G_eff,f/10 [0,1][0,1] η(gf)η(g_f) Capture function: 1−e−λgf1-e^-λ g_f, λ=3.5λ=3.5 (0,1)(0,1) ΔRf R_f Realized ramp increment: Rf(t^f+T)−Rf(t^f)R_f( t_f+T)-R_f( t_f) [0,1][0,1] IFSresIFS_res IFS residual multiplier: product of risk factor scores (0,1](0,1] TVpTV_p Terminal value contribution of pool p $M ΔEV Total risk adjusted enterprise value uplift $M or $B IFS risk factors δOCC _OCC Organizational change capacity [0,1][0,1] δDR _DR Data readiness [0,1][0,1] δVTR _VTR Vendor/technology risk [0,1][0,1] δCRS _CRS Competitive response speed [0,1][0,1] δREG _REG Regulatory exposure [0,1][0,1] A.2 Parameter Calibration and Empirical Grounding Table 16 consolidates all free parameters in the AITG framework with their calibrated values, permissible ranges, and source rationale. This disclosure responds to the legitimate concern that composite indicator frameworks often embed opaque parameter choices; here, every parameter is traceable to a documented source or sensitivity tested range. Table 16: Consolidated Parameter Calibration Register Parameter Symbol Base Value Range Tested Source / Rationale CES bottleneck aggregator Substitution parameter ρ 5 [3,8][3,8] Grid search against Standish Group (2015) project data; ESM Table S3 Dimension floor edmine_d 0.01 — Division by zero guard; economically negligible AI Frontier Coefficient Industry sensitivity θi _i [0.05,1.50][0.05,1.50] Full range Calibrated per industry from 2020–2024 benchmark adoption comovement AFC cap αmax _ 1.35 [1.20,1.50][1.20,1.50] Expert elicitation; prevents extrapolation beyond observed CtC_t range EWMA decay λewma _ewma 0.50 [0.3,0.7][0.3,0.7] Standard quarterly smoothing; half weight on most recent quarter Cascading S curve (adoption trajectory) Wave ceilings L1,L2,L3L_1,L_2,L_3 4.0, 3.5, 2.5 — Sum=10.0=10.0 by construction; wave proportions from GPT diffusion literature Wave steepness k1,k2,k3k_1,k_2,k_3 0.38, 0.42, 0.32 ±20%± 20\% mo-1; calibrated to observed enterprise adoption timelines Wave midpoints t0,1,t0,2,t0,3t_0,1,t_0,2,t_0,3 18, 36, 60 ±6± 6 mo Months; Wave 3 set 30% slower per Kinniment et al. (2024) Capture and value ramp Capture rate λ 3.5 [2.0,5.0][2.0,5.0] Controls η(g)=1−e−λg/10η(g)=1-e^-λ g/10 concavity; ESM Sobol analysis Capture floor ψfloor _floor 0.15 [0.10,0.25][0.10,0.25] Minimum realisable capture at near zero gap Ramp steepness krampk_ramp 0.18 [0.12,0.24][0.12,0.24] mo-1; value ramp logistic parameter IFS parameters OCC exponent βOCC _OCC 0.40 [0.30,0.50][0.30,0.50] Timing delay weight; Bloom et al. (2016) DR exponent βDR _DR 0.60 [0.50,0.70][0.50,0.70] Timing delay weight; complements OCC IFS residual weights VTR/CRS/REG 0.35/0.35/0.30 ±0.05± 0.05 Expert elicitation; equal VTR/CRS, slightly lower REG Firm scale and IASS Scale logistic steepness αΦ _ 2.0 [1.5,3.0][1.5,3.0] Controls Φf _f sigmoid sharpness RFF floor RminR_ 5.0 — Regulatory penalty threshold; NIST/EU AI Act calibration RFF penalty exponent γ 1.5 [1.0,2.0][1.0,2.0] Controls ψ suppression severity IASS weights wdw_d Table 2 ±5%± 5\% OECD composite methodology; Monte Carlo tested Note. “Range Tested” indicates the interval explored in sensitivity analysis (ESM Part I). Parameters marked “—” are structural constants not subject to calibration. All parameters are disclosed at their operational values; no parameters are estimated from the backtest outcome data. Appendix B Technical Derivations This appendix collects the key formulas, derivations, and technical remarks from the main text. B.1 IASS Sub Dimension Formulas B.1.1 Cognitive Task Density Let occupations be indexed o (O*NET-SOC codes) and industries i (4 digit NAICS). Let Ei,oE_i,o be BLS OEWS employment of occupation o in industry i. Let ao∈[0,1]a_o∈[0,1] be occupation o’s automatable cognitive task share, constructed by applying a fixed task classifier to O*NET task statements following the Autor–Levy–Murnane taxonomy Autor et al. (2003); Autor and Dorn (2013). CTDi=∑oEi,o⋅ao∑oEi,oCTD_i= _oE_i,o· a_o _oE_i,o (38) The definition of aoa_o follows Autor and Dorn (2013): the routine task intensity index is RTIk=lnTkR−lnTkA−lnTkMRTI_k= T_k^R- T_k^A- T_k^M, where: TkA=12(DCPk+GED-MATHk),TkR=12(STSk+FINGDEXk),TkM=EYEHANDkT_k^A= 12(DCP_k+GED -MATH_k), T_k^R= 12(STS_k+FINGDEX_k), T_k^M=EYEHAND_k (39) (DCPDCP = direction/control/planning; STSSTS = set limits/tolerances/standards; FINGDEXFINGDEX = finger dexterity; EYEHANDEYEHAND = eye hand foot coordination). High RTI occupations are classified as cognitively automatable at higher weight in aoa_o. Raw CTDiCTD_i is winsorized at the 5th/95th percentile across industries to limit outlier leverage, then min max normalized to [0,10][0,10]: x~d,i=xd,i−minjxd,jmaxjxd,j−minjxd,j×10 x_d,i= x_d,i- _jx_d,j _jx_d,j- _jx_d,j× 10 (40) Remark B.1 (Min Max Normalization and Recalibration). The [0,10][0,10] normalization anchors dimension scores to the observed cross industry distribution at the calibration date (Q4 2024). Because normalization bounds are determined by the empirical minimum and maximum across the 22 calibrated industries, a new industry with an extreme score could shift the bounds for all industries. In practice, winsorization at the 5th/95th percentile (applied before normalization) limits this effect to ≤0.4≤ 0.4 points per industry. As the AI landscape evolves, periodic recalibration of normalization bounds is required; the AFC versioning protocol (Section 4.6) specifies recalibration triggers. B.1.2 Process Repeatability Index (PRI) For industry i, let fi,kf_i,k be normalized task phrase frequencies from trailing twelve month job postings. Define task entropy: Hi=−∑kfi,klogfi,kH_i=- _kf_i,k f_i,k (41) Lower entropy ⇒ higher standardization ⇒ higher PRI. The PRI combines the complement of rank normalized entropy with a standardization language index (frequency of “workflow,” “queue,” “ticketing,” “SOP”). B.2 AFC CtC_t Weight Table and EWMA Table 17: AI Capability Index CtC_t: Benchmark Domains and Weights (ωk _k) k Domain Representative Benchmarks ωk _k 1 Language Understanding MMLU, HellaSwag, ARC 0.20 2 Mathematical Reasoning MATH, GSM8K 0.15 3 Code Generation HumanEval, MBPP, SWE-bench 0.15 4 Multimodal (Vision) MMMU, ChartQA 0.10 5 Agentic / Tool Use WebArena, WorkArena, ToolBench 0.15 6 Domain Specific MedPerf, FinBench, LegalBench 0.15 7 Long Context / RAG RULER, LongBench 0.10 Total 1.00 Note. Each bk,tb_k,t is the highest normalized score achieved by any frontier model family (GPT, Claude, Gemini, Llama, Mistral) on the domain’s benchmark suite at the end of quarter t. Weights reflect the relative importance of each capability domain for enterprise AI transformation; they are calibrated via expert elicitation across the six IASS dimensions and are subject to periodic recalibration as the benchmark landscape evolves. Model families with scores on fewer than 3 of 7 domains are excluded to prevent single benchmark distortion. Time smoothing. Raw benchmark scores can exhibit transient spikes when a new frontier model releases on a single benchmark before the result generalizes across the suite. To prevent such anomalies from artificially shocking the industry ceiling, I compute CtC_t as an Exponentially Weighted Moving Average (EWMA) of the past three quarters of benchmark high water marks: Ctsmooth=λ⋅Ctraw+(1−λ)⋅Ct−1smooth,λ=0.5C_t^smooth=λ· C_t^raw+(1-λ)· C_t-1^smooth, λ=0.5 (42) where CtrawC_t^raw is the weighted composite from Eq. (10) at the end of quarter t and λ=0.5λ=0.5 places equal weight on the most recent quarter and the exponentially decaying history. The half weight decay means that a single benchmark spike in quarter t propagates at 50% weight into the AFC and falls to below 6% weight within three quarters. The smoothed series is the operational input to all AFC calculations in this paper. B.3 AFC Scenario Table and Uncertainty Table 18: AFC Scenarios (24 Month Horizon) Scenario Assumption Annual ΔCt C_t AFC (24 mo) Δ HC-IASS Weight Conservative Capability slowdown; inference costs persist 8–12 % 1.04 +0.5 pts 0.20 Base Case Observed 2022–24 trajectory continues 18–25 % 1.10 +1.1 pts 0.60 Aggressive Reasoning models + agents deployed at scale 35–50 % 1.22 +2.2 pts 0.20 The AFC uncertainty contribution to the overall Uncertainty Quotient is: UQafc=AFCagg−AFCcon4UQ_afc= AFC_agg-AFC_con4 (43) approximating the 90 % confidence half width under a roughly normal scenario distribution. B.4 Piecewise Analytic Inverse with NaN Guard Remark B.2 (Piecewise Analytic Inverse with NaN Guard). The Wave 1 closed form approximation: t^f≈t0,1−1k1ln(L1AITGraw−1) t_f≈ t_0,1- 1k_1 \! ( L_1AITG^raw-1 ) is only valid for AITGraw<L1=4.0AITG^raw<L_1=4.0. For AITGraw≥L1AITG^raw≥ L_1, the argument of the logarithm becomes nonpositive (L1/AITGraw−1≤0L_1/AITG^raw-1≤ 0), producing a domain error (NaN crash) in any software implementation. Across the fourteen company cohort, this affects 9 of the 14 companies (all firms with AITG ≥L1=4.0≥ L_1=4.0): ServiceNow (8.50), JPMorgan (8.22), Goldman Sachs (8.17), Salesforce (8.07), Palo Alto (7.83), Wells Fargo (6.00), Rockwell (5.83), CVS (4.83), Target (4.67), Ford (4.67), and HCA (4.33). The corrected specification is a three branch piecewise inverse: t^f=t0,1−1k1ln(L1AITGraw−1)AITGraw<L1t0,2−1k2ln(L2AITGraw−L1−1)L1<AITGraw<L1+L2Newton-Raphson(AITG(t)=AITGraw)AITGraw≥L1+L2 t_f= casest_0,1- 1k_1 \! ( L_1AITG^raw-1 )&AITG^raw<L_1\\[8.0pt] t_0,2- 1k_2 \! ( L_2AITG^raw-L_1-1 )&L_1<AITG^raw<L_1+L_2\\[8.0pt] Newton-Raphson\! (AITG(t)=AITG^raw )&AITG^raw≥ L_1+L_2 cases (44) Branch 1 (AITGraw<L1−εAITG^raw<L_1- , ε=0.01 =0.01): Wave 1 dominant; exact within ±0.3± 0.3 months for AITGraw∈[0.5,3.8]AITG^raw∈[0.5,3.8]. The ε guard prevents division by zero at AITGraw=L1=4.0AITG^raw=L_1=4.0. Branch 2 (4.0<AITGraw<7.54.0<AITG^raw<7.5): Wave 2 dominant; analogous formula centred on t0,2t_0,2. Branch 3 (AITGraw≥7.5AITG^raw≥ 7.5): Wave 3 active; Newton–Raphson on the full cascading function Eq. (15) is mandatory. Convergence is guaranteed by strict monotonicity in 4–8 iterations from starting estimate t0=60t_0=60 months. Software implementation note. Any implementation must branch on the input value before calling the logarithm. The guard if AITG_raw >= L1: use_numerical_solver() prevents the NaN crash. I solved all fourteen companies in Section 9 numerically to produce the t^f t_f values in Table 12. B.5 Out of Sample Validation Strategy for θi _i Retrospective calibration of θi _i from 2020–2024 constitutes an in sample fit. I address this through rolling reestimation, with a critical distinction from prior formulations: the updating signal must be the frontier expansion in addressable task surface area, not current adoption breadth. The circularity flaw in adoption based updating. A prior version of the update rule used Census BTOS AI adoption rates as the observable proxy for ΔIASSiobserved _i^observed. This conflates two categorically distinct objects: (1) the theoretical capability ceiling that the IASS measures, and (2) the current adoption rate, which the BTOS measures. Healthcare could experience massive expansion in its theoretical AI ceiling (via domain specific clinical LLMs) while showing near zero change in BTOS adoption rates due to FDA clearance lags. If the BTOS signal drives θi _i updates, the Healthcare ceiling remains artificially suppressed, defeating the AFC’s purpose. Corrected update rule: O*NET task automatability expansion. I update θi _i only when the addressable task surface area in industry i expands, observable through O*NET task automatability reevaluations rather than survey based adoption rates. The corrected rule is: θ^i(t)=θ^i(t−1)+η⋅ΔATDi(t)ΔC(t)if ΔC(t)≥εθ^i(t−1)if ΔC(t)<ε θ_i^(t)= cases θ_i^(t-1)+η· _i^(t) C^(t)&if C^(t)≥ \\[6.0pt] θ_i^(t-1)&if C^(t)< cases (45) with ε=0.005 =0.005 (capability stall threshold: no update when the AI Capability Index grows by less than 0.5 % annually). Bounds: θ^i(t)∈[0.05, 1.50] θ_i^(t)∈[0.05,\;1.50]; values outside this range are clamped. The lower bound prevents near zero sensitivity (implying total AI immunity) for any industry with observable task automatability; the upper bound prevents the AFC from more than doubling the IASS ceiling within a single recalibration cycle. where: • ΔATDi(t) _i^(t) is the annual change in the Automatable Task Density (ATD) for industry i, defined as the weighted share of O*NET task importance scores classified as AI automatable under the current capability frontier CtC_t: ATDi(t)=∑j∈iwj⋅[Auto(j,Ct)=1]∑j∈iwjATD_i^(t)= _j _iw_j·1\! [Auto(j,C_t)=1 ] _j _iw_j (46) where iJ_i is the set of O*NET task items for industry i, wjw_j is the task importance weight, and Auto(j,Ct)=1Auto(j,C_t)=1 iff task j is classifiable as automatable under AI capability level CtC_t (assessed annually via O*NET reevaluation methodology O*NET (2024)). • ΔC(t) C^(t) is the annual increment in the AI Capability Index (Eq. 10). • η=0.30η=0.30 is the learning rate, unchanged. Remark B.3 (Separation of Frontier from Adoption). The corrected specification enforces a clean separation: θi _i responds to what AI can now do in industry i (measured by O*NET task automatability reclassification), while the AFC separately models the capability index CtC_t driving frontier expansion. Adoption data (Census BTOS McElheran et al. (2024), Lightcast Lightcast (2024)) remain relevant for validating GeffG_eff and calibrating the ADRI competitive diffusion rate, but are explicitly excluded from the θi _i update loop. This separation prevents the regulatory compliance lag problem from suppressing frontier estimates in heavily regulated industries. B.6 CES Boundary Proof The following proof sketch establishes the boundary behaviour of the CES bottleneck aggregator (Proposition 7.1). Proof sketch. (1) Write bp=(∑dαded−ρ)−1/ρb_p=( _d _de_d^-ρ)^-1/ρ. As ρ→∞ρ→∞, the sum is dominated by the term with the smallest ede_d (largest ed−ρe_d^-ρ), so bp→(αd∗⋅ed∗−ρ)−1/ρ=ed∗⋅αd∗1/ρ→ed∗=mindedb_p→( _d^*· e_d^*^-ρ)^-1/ρ=e_d^*· _d^*^1/ρ→ e_d^*= _de_d. (2) Take logarithms: lnbp=−1ρln(∑dαded−ρ) b_p=- 1ρ ( _d _de_d^-ρ). As ρ→0+ρ→ 0^+, apply L’Hôpital’s rule to obtain lnbp→∑dαdlned b_p→ _d _d e_d. (3) Since ed≥0.01e_d≥ 0.01 for all d, ed−ρ≤0.01−ρ=100ρe_d^-ρ≤ 0.01^-ρ=100^ρ. Hence ∑dαded−ρ≤100ρ _d _de_d^-ρ≤ 100^ρ and bp≥(100ρ)−1/ρ=0.01>0b_p≥(100^ρ)^-1/ρ=0.01>0. ∎ Appendix C Robustness and Sensitivity Details This appendix provides the detailed robustness tables and ablation analyses summarized in the main text. C.1 Sensitivity to Weighting: First Robustness Check Following OECD Handbook guidance OECD (2008), I evaluate ranking stability by perturbing weights uniformly within ±0.05± 0.05 of baseline values. Table 19 reports the 5th/95th percentile IASS range under 10,000 weight draws for each anchor industry. Table 19: IASS Anchor Calibration with Robustness Bounds Industry CTD DRSA PRI RFF CADR CLSR IASS P5–P95 θi _i Financial Services 8.8 9.2 7.9 5.0 8.4 7.6 7.83 7.50–8.15 0.22 Healthcare Services 6.2 5.8 6.5 4.1 5.1 6.8 4.27 3.97–4.57 0.31 Industrial Distribution 5.4 6.1 7.8 8.9 4.7 6.3 6.43 6.01–6.84 0.14 Construction 3.9 3.2 4.1 7.4 3.3 5.7 4.26 3.90–4.61 0.09 Vertical SaaS 9.4 9.8 8.6 8.1 9.2 9.1 9.06 8.79–9.28 0.11 θi _i: AFC sensitivity (Sec. 4). Weights: CTD 0.25; DRSA 0.20; PRI 0.20; RFF 0.15; CADR 0.10; CLSR 0.10. Rankings are stable: no pair of anchor industries exchanges rank under any weight draw. The largest interindustry gap (Financial Services 7.83 vs. Healthcare 4.27 = 3.56 points) exceeds the widest confidence interval (0.60) by a factor of 5.9, confirming that ranking uncertainty does not affect cross industry comparison at this scale of separation. C.2 IFS Ablation Analysis Table 20 reports the effect of dropping individual IFS factors on backtest rank correlation and mean absolute Δ error across the fourteen company cohort. The full five factor specification serves as the baseline; each row removes one factor while holding all others at calibrated values. Table 20: IFS Ablation: Impact of Dropping Individual Factors IFS Specification Backtest ρs _s Mean |ΔEV|| | Error (%) Δ vs. Full Full IFS (OCC+DR+VTR+CRS+REG) 0.818 — Baseline Drop OCC 0.709 ++18% −-0.109 Drop DR 0.745 ++14% −-0.073 Drop VTR 0.800 ++5% −-0.018 Drop CRS 0.791 ++7% −-0.027 Drop REG 0.809 ++3% −-0.009 Note. Trajectory factors (OCC, DR) contribute the largest marginal signal because they enter the adoption curve endogenously via kw,fk_w,f and t50,ft_50,f, not as terminal value multipliers. Removing OCC reduces ρs _s by 0.109, confirming that organizational change capacity is the single most informative IFS factor. Residual factors (VTR, CRS, REG) contribute less individually but collectively account for an additional ++15% reduction in Δ error. Full ablation details in ESM Part I. C.3 IFS Component Interpretability Each IFS factor maps to an observable organizational characteristic with a clear economic interpretation: • OCC (Organizational Change Capacity): measures management’s demonstrated ability to execute technology driven transformation, proxied by prior restructuring success, voluntary tool adoption rates, and AI governance maturity. Higher OCC ⇒ steeper adoption S curve (faster kw,fk_w,f). • DR (Data Readiness): assesses data infrastructure maturity: cloud migration status, data governance scores, CDO tenure, and data quality audit history. Higher DR ⇒ earlier ramp inflection (t50,ft_50,f decreases). • VTR (Vendor/Technology Risk): captures dependency concentration on external AI vendors, model access fragility, and SOC 2 compliance posture. Affects terminal value realization via residual multiplier. • CRS (Competitive Response Speed): measures the industry’s AI adoption velocity relative to the firm, proxied by Lightcast CADR, competitor press, and CB Insights data. Higher CRS ⇒ less residual friction. • REG (Regulatory Exposure): quantifies the regulatory overhang from NIST AI RMF, EU AI Act, HIPAA/FDA applicability, and pending legislation. Higher REG ⇒ lower residual friction (better regulatory preparedness). C.4 CES Aggregator Ablation I address the concern that the low elasticity CES choice is asserted rather than validated by comparing three aggregators across the ten nonfinancial backtest firms: 1. CES (ρ=5ρ=5, σ=1/6σ=1/6): the proposed specification. 2. Leontief (min): bp=mindedb_p= _de_d, the most restrictive bottleneck. 3. Additive (mean): bp=∑dαdedb_p= _d _de_d, fully compensable. At the n=10n=10 scale, all three aggregators produce identical Spearman rank correlations with realized margin outcomes (ρs=0.818 _s=0.818) because no backtest firm has a near zero dimensional score; the compensability difference is inactive when all dimensions exceed 2.0. The CES advantage manifests in two regimes: (a) firms with one very weak dimension (score <2.0<2.0), where additive overestimates value capture relative to CES and Leontief, and (b) near threshold boundary conditions that generate NaN in pure Leontief implementations. The extended 22 company simulation in ESM Table S6 includes three synthetic firms with single dimension near zero scores; in that expanded set, additive diverges from CES by 35–60% on Δ while CES and Leontief agree within 8%. The Leontief minimum operator is not rejected by the backtest data but is computationally fragile (zero bounding pathology) and unduly sensitive to scoring noise in the lowest dimension. CES at ρ=5ρ=5 preserves the bottleneck intuition while eliminating these failure modes. Robustness across ρ∈[3,8]ρ∈[3,8] is confirmed in ESM Table S3. Appendix D Worked Examples and Cost Assumptions This appendix provides the detailed step by step VCB worked example and implementation cost assumptions. D.1 Zions VCB Step by Step Table 21 traces the four algebraic corrections through a single representative value pool (Labor Productivity) to show exactly how inputs map to ΔEV . Table 21: VCB Step by Step: Zions Labor Productivity Pool Step Component Input Output Equation / Note 1 Revenue baseline $3.4B rev Bp=$3.4B×0.08=$0.272B_p= 3.4B× 0.08= 0.272B 8% labor productivity uplift rate (Eq. 22) 2 Firm Scale Factor Φf _f Rf=$3.4BR_f= 3.4B, Si∗=$3.3BS^*_i= 3.3B Φf=0.515 _f=0.515 Eq. 21: 1/(1+e−2ln(3.4/3.3))1/(1+e^-2 (3.4/3.3)); just above threshold 3 CES bottleneck bpb_p ePAC=0.40e_PAC=0.40, eWAR=0.35e_WAR=0.35, eOAC=0.42e_OAC=0.42 bp=0.382b_p=0.382 Eq. 23: (∑αded−5)−1/5(Σ _de_d^-5)^-1/5; normalized inputs ed=sd/10∈[0,1]e_d=s_d/10∈[0,1]; weak WAR suppresses pool 4 Gap fraction gfg_f Geff=5.58G_eff=5.58 gf=0.558g_f=0.558 Eq. 24: gf=Geff/10g_f=G_eff/10; Geff=IASS∗−AITG=9.38−3.80G_eff=IASS^*-AITG=9.38-3.80 5 Concave capture η gf=0.558g_f=0.558, λ=3.5λ=3.5 η=0.859η=0.859 Eq. 25: 1−e−3.5×0.5581-e^-3.5× 0.558; wide gap yields high capture at correct λ 6 Raw pool value VpV_p Steps 1–5 Vp≈$30MV_p≈ 30M Eq. 22: $0.272B×0.515×0.65×0.382×0.859 0.272B× 0.515× 0.65× 0.382× 0.859; κ¯p=0.65 κ_p=0.65 7 IFS delay: t50,ft_50,f δOCC=0.55 _OCC=0.55, δDR=0.48 _DR=0.48, t0=18t_0=18 t50,f≈35.5 mot_50,f≈ 35.5 mo Eq. 35: base t0=18t_0=18 (Wave 1; Zions AITG <L1<L_1); 1.97×1.97× delay 8 Ramp increment ΔRf R_f t^f≈23.8 t_f≈ 23.8, t50=35.5t_50=35.5, T=60T=60 ΔRf=0.891 R_f=0.891 Eq. 27: Rf(83.8)−Rf(23.8)=0.999−0.108R_f(83.8)-R_f(23.8)=0.999-0.108; logistic ramp k=0.18k=0.18 9 IFS residual VTR=0.72, CRS=0.62, REG=0.82 IFSres=0.710IFS_res=0.710 Eq. 36: 0.720.35×0.620.35×0.820.300.72^0.35× 0.62^0.35× 0.82^0.30 10 Terminal Value (pool) Steps 6,8,9; Mi=10M_i=10 TVp≈$190MTV_p≈ 190M Eq. 28: $30M×0.891×10×0.710 30M× 0.891× 10× 0.710; 10×10× exit multiple Labor Productivity pool contribution $190M Across 7 pools at base case assumptions, total ΔEV≈$0.90B ≈ 0.90B Step 3 uses normalized CES inputs ed=sd/10∈[0,1]e_d=s_d/10∈[0,1]; reporting raw dimension scores (e.g., 4.0) rather than normalized inputs inflates the pool by an order of magnitude. Step 2 applies Φf _f once via Vprun rateV_p^run rate (Eq. 22); it does not appear again in the terminal value numerator. Step 4 uses Geff=IASS∗−AITG=9.38−3.80=5.58G_eff=IASS^*-AITG=9.38-3.80=5.58 (not the ADRI score of 2.6, which is a separate composite index). Step 5 uses λ=3.5λ=3.5 from Eq. 25, the calibrated concavity parameter; using λ=1.0λ=1.0 underestimates capture by 60%. Step 7 uses base t0=18t_0=18 months (Wave 1), not 36: Zions’s AITG of 3.80 falls below the Wave 1 asymptote (L1=4.0L_1=4.0) and is correctly assigned to the Foundation AI wave. The 1.97×1.97× IFS delay then shifts the inflection from 18 to 35.5 months. A data ready peer (δDR=0.90 _DR=0.90) would reach t50≈18t_50≈ 18 mo and capture ΔRf≈0.677 R_f≈ 0.677, a 3.1×3.1× ramp advantage over Zions arising entirely from data infrastructure readiness. D.2 Implementation Cost Assumptions Table 22: Implementation Cost Assumptions Underlying VD Estimates Company Est. Cost Basis 5 yr Δ VD Range VD Sensitivity JPMorgan Chase $2.3B Revenue × 1.2% implementation rate $20–98B 9.5–46.0× – Zions Bancorp. $38M Revenue × 1.2% implementation rate $0.6–1.2B 15.0–30.1× ±0.4×± 0.4× per $100M UPS $1.1B Revenue × 1.2% implementation rate $30–61B 27.3–55.4× ±0.3×± 0.3× per $500M HCA Healthcare $780M Revenue × 1.2% implementation rate $5.8–20.6B 7.4–26.4× ±0.5×± 0.5× per $500M Salesforce $420M Revenue × 1.2% implementation rate $10.6–33.7B 25.2–80.4× ±2.0×± 2.0× per $500M Ferguson $360M Revenue × 1.2% implementation rate $7.8–16.0B 21.8–45.0× ±0.4×± 0.4× per $100M Rockwell Auto. $100M Revenue × 1.2% implementation rate $1.6–4.5B 15.8–45.0× ±0.6×± 0.6× per $200M VD = Δ / Total Implementation Cost. Sensitivity = change in VD midpoint per stated cost increment. Calibration note. The 1.2% of revenue implementation rate is a cross industry simplification. Published benchmarks suggest meaningful variation: financial services firms allocate approximately 1.2–1.6% of revenue to AI transformation (Gartner IT spending data: ∼ 8% total IT spend × 15–20% AI share), retail firms approximately 3.3% (IBM, 2025), and discrete manufacturers 0.4–0.8% (Avasant, 2024). The uniform 1.2% rate falls within this range and was adopted for tractability. Sobol first order decomposition (Section 10) confirms that implementation cost explains only ∼ 1% of VD variance—exit multiple (50%) and capture rate (44%) dominate—so industry specific cost variation does not materially alter firm rankings, tier assignments, or the qualitative findings. The Monte Carlo (Layer 4) stress tests implementation cost with a LogNormal(0, 0.30) multiplier, spanning roughly 0.6%–2.4% of revenue, which covers the observed cross industry range. Cost assumptions must nonetheless be explicitly stated and challenged in any investment decision context. Appendix E Extended Research Agenda Nine empirical research priorities emerge from this framework, organized around causal validation, technical refinement, and measurement quality: 1. Causal Identification via Instrumental Variables (IV). Future empirical validation must employ IV strategies to isolate the causal effect of AI adoption from incumbent endogeneity. Promising IV designs include exploiting exogenous variation in cloud computing incentive grants (e.g., state level subsidies) or differential firm level exposure to staggered open source LLM releases based on preexisting tech stack composition. These instruments can be used to predict firm level AI investment, allowing observation of the unconfounded effect on the AITG Effective Gap (GeffG_eff) and subsequent margin expansion. 2. Large panel validation of the mid gap optimality property. The framework’s mathematical prediction that medium gap firms in high IASS sectors generate superior risk adjusted value density is a theoretical property of the S curve and CES architecture, not yet an empirically established finding. A multiyear panel study (2018–2026) with quarterly AITG estimates and out of sample predictions of margin expansion, SG&A/productivity deltas, and risk adjusted returns is required to establish whether this property holds empirically after controlling for firm size, capital structure, and exit multiple heterogeneity. 3. AFC calibration and θi _i estimation. Refining industry specific θi _i parameters with systematic backward looking analysis using the Eloundou et al. Eloundou et al. (2024) LLM exposure estimates by occupation, combined with observed AITG changes in early adopter industries, would produce more defensible AFC sensitivity estimates. Standard errors for θi _i and scenario analysis separating capability growth from adoption dynamics are required for AFC to meet identification standards. 4. Interrater reliability study. A formal reliability study with multiple independent analyst teams scoring the same companies on the full AITG rubric (six company dimensions and five IFS factors) would establish whether the target reliability is achievable in practice. Agreement metrics should include Gwet’s AC1/AC2 Gwet (2014) for ordinal scales and Krippendorff’s α for full rubric consistency; both correct for chance agreement differently and provide complementary perspectives. Target: ICC >0.75>0.75 per factor, overall α>0.70α>0.70. The full scoring rubric, including worked examples and anchor vignettes for each score level on each dimension, should be published as a standalone replication document to enable third party scoring and interrater studies. Factors with ICC <0.60<0.60 should be subject to rubric redesign before use in investment decisions. This study is the single highest priority empirical prerequisite for institutional deployment of the framework. 5. Competitive displacement measurement. Linking ADRI scores to subsequent market share outcomes and margin trajectories at the firm level would test the core ADRI hypothesis, that wide gap + high CADR + low moat predicts competitive deterioration, and distinguish “risk from inaction” from generic competitive intensity or industry cyclicality. 6. Cross country calibration. IASS calibration to date uses U.S. labor market task composition (O*NET/OEWS). International application requires country specific task data (ESCO for EU, equivalent national databases for other regions) and jurisdiction specific regulatory exposure scores. ISCO-88 based RTI estimates Goos et al. (2014) provide a starting point for EU calibration. 7. Preregistration, rolling origin validation, and replication bundle. Future validation studies should implement: (i) parameter range preregistration before data access, locking the AFC scenario distribution and value pool share coefficients before observing the outcome cohort; (i) rolling origin evaluation building AITG at t0t_0 and predicting outcomes over [t0,t0+2][t_0,t_0+2] on holdout firms not used in calibration; (i) placebo outcome tests using CapEx intensity (where AI adoption has limited direct traction) to check for spurious correlations; (iv) release of a replication bundle containing anonymized scoring data, parameter values, and reproducible code sufficient for external auditors to reconstruct the backtest ρs _s from raw inputs. 8. Variable elasticity and nested CES structures. The current single layer CES with fixed ρ=5ρ=5 is a simplification. A more realistic structure would nest enablers hierarchically (e.g., Data + Infrastructure gate → Talent + Governance gate → commercial enablement), reflecting the observed sequencing of enterprise AI deployment. Variable σ forms, where elasticity decreases as a firm approaches the frontier (substitution becomes harder at the margin), could better reflect the complementarity dynamics documented in the organizational IT complementarities literature Brynjolfsson and Hitt (2000); Gibbons and Henderson (2012). Estimating nested structures from panel data is a companion technical priority. 9. Hierarchical Bayesian parameter estimation. The current deterministic parameterization relies on Sobol first order sensitivity indices to assess robustness and bound overfitting risk. Formally resolving identifiability, given the framework’s degrees of freedom across 22 industry verticals, requires a hierarchical Bayesian specification. Explicit priors on θi _i, κ¯p κ_p, and the IFS elasticities, combined with partial pooling across industries, would regularize parameter updates and allow posterior credible intervals to replace the current Monte Carlo uncertainty bands. This approach would also enable principled updating as panel data from the longitudinal backtest accumulates: each new firm year observation narrows the posteriors under standard Bayesian updating. The Bayesian extension is the highest priority technical research priority and is being designed for a companion empirical paper. Electronic Supplementary Material (ESM) The AI Transformation Gap Index (AITG) Dean Barr dean@dsconsult.ai February 2026 Appendix S1 ESM Part I: Full IASS Industry Calibration Reference This part supports Main Paper Sections 3–4 (IASS construction and scoring architecture). The main manuscript presents five anchor industry examples drawn from this complete calibration table. All 22 verticals are reported here with full dimensional breakdowns. Source methodology follows the OECD composite indicator approach OECD (2008) using BLS O*NET, Lightcast, Census BTOS, and SEC filing data as detailed in the main text. One of the practitioner barriers to adopting composite indicator frameworks is the cost of calibrating the industry baseline from scratch for each engagement. I resolve this barrier by pre computing and publishing IASS calibrations for 22 industry verticals, using real data from the following sources: • Cognitive Task Density (CTD): Bureau of Labor Statistics O*NET task importance scores for cognitive and information processing activities, 2024 OEWS program BLS OEWS (2023); O*NET (2024). • Data and System Automation (DRSA): McKinsey Automation Potential estimates by occupation cluster McKinsey (2025); Deloitte (2026); Lightcast Generative AI skill demand index (2024–2025) Lightcast (2025). • Process Re engineering Index (PRI): Dun & Bradstreet sector digitization scores and Gartner digital density rankings. • Regulatory Friction Factor (RFF): National Institute of Standards and Technology (NIST) AI Risk Management Framework NIST (2023) sector mapping; EU AI Act high risk classification list EU AI Act (2024); HIPAA, FDA SaMD regulations. • Competitive AI Diffusion Rate (CADR): Census Bureau Business Trends and Outlook Survey (BTOS) AI adoption rates by sector 2023–2025 McElheran et al. (2024); Lightcast AI job posting concentration index. • Cost/Labor Structure (CLSR): BLS Occupational Employment and Wage Statistics 2023–2024 BLS OEWS (2023). Table S1 reports the calibrated scores. All sub scores are on the 0–10 scale. θi _i is the industry level AITG sensitivity coefficient calibrated from 2020–2024 observable adoption data (Census BTOS, Lightcast AI skill demand, and McKinsey State of AI surveys). IASS is computed via the geometric aggregation and RFF floor equations in the main paper (Section 3). Industries with RFF <5.0<5.0 incur the regulatory suppression penalty ψ<1.0ψ<1.0, which is reported in the ψ column. Table S1: Table S1: Pre Computed IASS Reference Table: 22 Industry Verticals (2026 Calibration) Industry NAICS CTD DRSA PRI RFF CADR CLSR ψ IASS θi _i IASS∗ Financial Services Investment Banking / Securities 5211 9.2 9.5 8.4 5.5 9.0 8.2 1.000 8.30 0.28 10.39 Commercial Banking (Large Cap) 5221 8.8 9.2 7.9 5.0 8.4 7.6 1.000 7.83 0.22 9.38 Insurance (P&C + Life) 5241 8.5 8.8 8.2 6.5 7.2 7.4 1.000 7.92 0.19 9.27 Healthcare Healthcare Services (Hospital) 6221 6.2 5.8 6.5 4.1 5.1 6.8 0.743 4.27 0.31 5.46 Life Sciences / Pharma 5417 7.5 7.0 6.0 3.8 7.0 6.5 0.663 4.14 0.25 5.07 Digital Health / HealthTech 5112 8.8 7.5 7.8 5.5 7.8 7.5 1.000 7.54 0.22 9.03 Technology Vertical SaaS / Software 5112 9.4 9.8 8.6 8.1 9.2 9.1 1.000 9.06 0.11 9.96 Cybersecurity / InfoSec 5415 9.0 9.2 7.8 7.5 9.5 8.8 1.000 8.57 0.14 9.65 Professional Services Legal Services 5411 9.1 7.2 6.8 7.5 8.2 8.5 1.000 7.80 0.20 9.20 Accounting / Tax Services 5412 8.9 8.2 8.5 7.0 7.8 8.2 1.000 8.17 0.18 9.49 Retail & Commerce E Commerce / Digital Retail 4541 6.8 7.5 7.8 8.5 8.2 7.2 1.000 7.55 0.16 8.64 Grocery / Traditional Retail 4451 6.1 6.2 7.5 8.8 7.0 8.2 1.000 6.93 0.13 7.74 Industrial / Supply Chain Logistics / Transportation 4920 6.2 6.0 7.2 8.2 6.1 8.5 1.000 6.82 0.14 7.68 Industrial Distribution 5085 5.4 6.1 7.8 8.9 4.7 6.3 1.000 6.43 0.14 7.24 Discrete Manufacturing 3330 6.1 6.5 7.2 7.8 6.5 6.2 1.000 6.60 0.15 7.49 Other Sectors Energy / Utilities 2211 5.5 7.0 7.5 6.0 5.5 6.1 1.000 6.25 0.13 6.98 Media & Entertainment 5110 7.8 8.5 6.5 7.5 8.8 6.5 1.000 7.56 0.18 8.78 Real Estate / REITs 5311 5.2 5.5 6.1 7.0 6.0 5.5 1.000 5.73 0.12 6.35 Education (Higher Ed) 6113 7.2 5.5 5.0 6.5 6.5 6.0 1.000 6.07 0.14 6.83 Government / Public Sector 9200 7.0 5.5 6.5 5.0 3.5 7.0 1.000 5.83 0.08 6.25 Agriculture / AgTech 1100 3.2 5.0 5.5 7.0 4.5 4.8 1.000 4.72 0.10 5.14 Construction 2300 3.9 3.2 4.1 7.4 3.3 5.7 1.000 4.26 0.09 4.61 Sources: BLS O*NET 2024; Lightcast AI Skill Demand 2025; Census BTOS 2024; McKinsey State of AI 2025. ψ<1.0ψ<1.0 indicates regulatory suppression is binding (RFF <5.0<5.0). θi _i calibrated from 2020–2024 observable adoption data. IASS=∗IASS×min(1+θi(Ct−C0),αmax)^*=IASS× (1+ _i(C_t-C_0),\; _ ) with Ct=1.90C_t=1.90, C0=1.0C_0=1.0, αmax=1.35 _ =1.35. Remark S1.1 (Reading the Reference Table). Three patterns are immediately diagnostic for capital allocation. First, the Healthcare Services and Life Sciences rows carry ψ<1.0ψ<1.0, reflecting that the RFF hard floor suppresses their theoretical ceiling substantially: Hospital Healthcare’s geometric mean of 5.75 is reduced by 26% to IASS 4.27 once the FDA and HIPAA friction penalty is applied. Second, Construction and Agriculture floor the table not because of regulatory friction but because their Cognitive Task Density (CTD) and Data/Systems Automation (DRSA) dimensions are structurally low; these industries have high physical task content that current AI cannot automate. Third, the four Vertical SaaS, Cybersecurity, Investment Banking, and Accounting rows all clear IASS >8.0>8.0, defining the top quartile where AI transformation opportunity is structurally dense. Table S2: Table S2: Standard Value Pool Parameters and Calibration Sources Value Pool Base Capture κ¯p κ_p Pool Dims (p)D(p) Typical Uplift % Calibration Source Labor Productivity 0.65 PAC, WAR, OAC 6–12% Brynjolfsson et al. (2025); McKinsey 2024 Revenue Enhancement 0.55 DAR, APR, OAC 4–9% Babina et al. (2024); Goldman internal Working Capital 0.60 DIM, PAC, DAR 8–15% McKinsey 2024 Global AI Report Risk / Compliance 0.50 DIM, DAR, OAC 10–20% Lightcast 2025; BTOS 2024 Operational Cost 0.62 PAC, WAR, DAR 5–10% Eloundou et al. (2024) Customer Experience 0.48 WAR, APR, OAC 8–18% Brynjolfsson, Li, Raymond (2025) Data Monetization 0.40 DIM, DAR, APR 12–25% Jones & Tonetti (2020); proprietary κ¯p κ_p = base capture rate before gap scaling η(g)η(g) and IFS residual haircut. Uplift % = baseline AI attributable improvement rate for the pool at Geff≈2.0G_eff≈ 2.0 (sweet spot firms). Pool dimensions (p)D(p): DIM=Data Infrastructure, PAC=Process Automation, WAR=Workforce Augmentation, DAR=Decision Automation, APR=AI Revenue Integration, OAC=Org. AI Capability. Full methodology in main paper Section 8. Table S3: Table S3: CES Bottleneck Robustness Across Substitution Parameter ρ∈[1,10]ρ∈[1,10] Profile Type ρ=1ρ=1 (Harmonic) ρ=3ρ=3 ρ=5ρ=5 (Proposed) ρ=8ρ=8 Leontief Additive Rank stable? Balanced high (all dims ≈ 8) 7.96 7.93 7.92 7.91 7.50 8.00 ✓ One weak dim (5 dims=8, 1 dim=3) 5.88 4.71 4.12 3.58 3.00 7.17 ✓ Two weak dims (4 dims=7, 2 dims=2) 3.40 2.48 2.12 1.82 2.00 5.33 ✓ Near zero dim (5 dims=7, 1 dim=0.5) 1.20 0.74 0.60 0.52 0.50 5.92 ✓ Simulated bpb_p scores for illustrative dimension profiles. “Balanced high” represents apex incumbents (e.g., JPMorgan, PANW). “One weak dim” is the typical mid gap firm with a single bottleneck. “Near zero” tests the Leontief fragility that motivates CES: the Leontief and CES converge but the CES avoids division by zero pathology. Additive dramatically overestimates at near zero. All rank orderings across profiles are identical for ρ∈[3,8]ρ∈[3,8]; CES dominates Leontief on stability and Additive on non compensability. Full ablation vs. realized outcomes in main paper Section 10.2. Table S4: Table S4: Financial Baselines and Data Tier Classification (Fourteen Company Cohort) Company Sector Rev ($B) EBITDA ($B) Emp (K) S*i ($B) Φf _f Data Tier JPMorgan Chase Comm. Banking 177.6 n/a 316 3.3 1.00 A Goldman Sachs Inv. Banking 47.3 n/a 46 2.0 1.00 A Wells Fargo Comm. Banking 82.6 n/a 219 3.3 1.00 A Zions Bancorp. Comm. Banking 3.4 n/a 10 3.3 0.52 A Salesforce Vert. SaaS 36.5 10.3 73 5.0 1.00 A ServiceNow Vert. SaaS 10.9 3.1 23 5.0 0.82 A Palo Alto Ntwks Cybersecurity 8.0 2.1 14 2.5 0.89 A HCA Healthcare Healthcare 67.5 14.5 300 10.0 0.88 B CVS Health Healthcare 380.0 17.8 300 10.0 1.00 A Target Retail 107.0 10.1 440 8.0 0.88 A UPS Logistics 91.0 14.6 500 5.0 1.00 A Ferguson PLC Ind. Dist. 29.0 3.9 33 3.0 1.00 B Rockwell Auto. Disc. Mfg. 9.1 2.2 28 3.0 0.87 B Ford Motor Disc. Mfg. 185.0 12.3 186 5.0 1.00 A Rev and EBITDA from FY2024/FY2025 10-K or most recent annual report (calendar year end unless noted). Banks: EBITDA not applicable; revenue = net interest + non interest income. Si∗S^*_i = industry critical scale threshold from IASS calibration (ESM Table S1). Φf=1/(1+exp(−2ln(Rf/Si∗))) _f=1/(1+ (-2 (R_f/S^*_i))). Tier A: audited/quantified; Tier B: analyst estimate or limited disclosure. Full source citations in main paper Section 9. Table S5: Table S5: Monte Carlo Δ Distributional Output: P10 / P50 / P90 (Select Firms) Company Sector P10 ($B) P50 ($B) P90 ($B) P90/P10 Dom. Var. Source Action Signal JPMorgan Chase Comm. Banking 20.18 44.44 97.95 4.9× Exit multiple (50%) Invest Goldman Sachs Inv. Banking 4.38 6.70 15.11 3.4× Capture rate (44%) Invest Salesforce Vert. SaaS 4.72 10.28 24.80 5.3× Exit multiple (50%) Invest ServiceNow Vert. SaaS 0.97 1.58 3.66 3.8× Capture rate (44%) Invest Palo Alto Ntwks Cybersecurity 1.12 2.52 5.32 4.8× Exit multiple (50%) Invest Rockwell Auto. Disc. Mfg. 0.88 1.47 2.91 3.3× Impl. cost (1%) Monitor Target Retail 24.25 38.39 57.12 2.4× IFS (5%) Invest HCA Healthcare Healthcare 3.64 5.62 8.44 2.3× Reg. ceiling (psi) Invest UPS Logistics 24.81 36.74 52.35 2.1× Exit multiple (50%) Invest Zions Bancorp. Comm. Banking 0.61 0.88 1.23 2.0× IFS (5%) Monitor Wells Fargo Comm. Banking 20.94 34.45 52.32 2.5× Impl. cost (1%) Invest CVS Health Healthcare 18.14 29.44 44.06 2.4× Capture rate (44%) Invest Ferguson Ind. Dist. 6.22 9.43 13.60 2.2× Exit multiple (50%) Monitor Ford Motor Disc. Mfg. 27.09 46.07 71.62 2.6× Exit multiple (50%) Invest M=10,000M=10,000 Monte Carlo draws. Input uncertainty sources: exit multiple ±2×± 2× (uniform), capture rate ±25%± 25\% (uniform), implementation cost ±30%± 30\% (lognormal), AITG gap score ±0.5± 0.5 pts (normal), IFS ±0.08± 0.08 (normal). All five sources sampled jointly; distributions propagated through the full VCB pipeline. Dominant variance source = Sobol first order index. Action signals: Invest =P10>$1B=P_10> 1B and VDP50>5×VD_P50>5×; Monitor =P50>0=P_50>0 and VD∈[2.5,5]VD∈[2.5,5]; Diligence =P10=P_10 near zero; Do Not Invest =P50=P_50 near zero. For acquisition context only; not investment advice. Table S6: Table S6: Extended CES Ablation Including Synthetic Near Zero Dimension Cases Firm / Profile Min Dim CES (ρ=5ρ=5) Leontief Additive Δ CES vs. Add. Note Real firms (from Section 10.1 backtest) PANW (balanced high) 6.5 7.02 6.50 7.08 −1%-1\% CES≈ ; balanced CRM (near balanced) 6.0 6.22 6.00 6.28 −1%-1\% Minimal CES penalty Target (one weak) 3.5 4.02 3.50 4.17 −4%-4\% Mild bottleneck UPS (two weak) 2.5 2.98 2.50 3.45 −14%-14\% CES penalizes Ford (multi weak) 1.5 2.21 1.50 3.02 −27%-27\% Additive overstates Synthetic corner cases Synth-A: 5 dims=8, 1 dim=1.0 1.0 2.78 1.00 6.83 −59%-59\% Leontief and CES both low; Add. badly wrong Synth-B: 5 dims=7, 1 dim=0.1 0.1 0.59 0.10 5.84 −90%-90\% Near zero; Add. 50×50× overstates Synth-C: 4 dims=6, 2 dims=1.5 1.5 2.23 1.50 5.00 −55%-55\% Two bottlenecks; Add. masks both Dimension scores normalized to ed=sd/10∈(0,1]e_d=s_d/10∈(0,1]. CES ρ=5ρ=5, σ=1/6σ=1/6. “Min Dim” = minimum raw dimension score. “Δ CES vs. Add.” = percentage by which CES Δ is lower than Additive, computed through the full VCB pipeline holding all other inputs fixed. All real firm rank orderings are identical for CES vs. Leontief; differences emerge only in magnitude. Leontief is not rejected by the data but suffers from: (1) zero bounding instability at near zero dims, and (2) maximum sensitivity to single dimension scoring noise. CES at ρ=5ρ=5 preserves Leontief’s rank behavior while eliminating both failure modes. Full main text discussion in Section 10.2. Appendix S2 ESM Part I: Extended Firm Level Empirical Case Studies This part supports Main Paper Section 9 (Empirical Illustrations). The following twelve case studies apply the full AITG framework to public companies across eight industries. Each follows a standardized format: operating context with public data citations; scored metrics (AITG Raw, Industry Ratio, GeffG_eff, ADRI, IFS); and a VCB Summary reporting 5 year risk adjusted EV creation range, implementation cost range, and AITG Value Density tier. Evidence tiers (A–D per the main manuscript’s scoring methodology) are noted where relevant. The full 14 company cross company synthesis (including the 14 company summary table, Table 24, and the Value Density Paradox scatter, Figure 5) appears in the main manuscript. Cost basis note. Each case study reports a comprehensive implementation cost range reflecting total estimated transformation spend (capex, opex, talent, integration) over the 5 year hold period. The 14 company summary table (Table 24) and the Monte Carlo VCB pipeline use a standardised 1.2% of revenue proxy cost (Appendix D, Table 22), which yields higher VD multiples. The two cost bases serve different purposes: the case study estimates are grounded in firm specific disclosures for narrative realism; the 1.2% proxy enables cross company comparability and Monte Carlo tractability. Readers should not directly compare VD ranges between the case narratives and the summary table without accounting for this difference. AFC recalibration note. The Δ and VD estimates in the case narratives below were derived prior to the final AFC recalibration to Ct=1.90C_t=1.90 (GPT-5.2, December 2025). The summary table and Monte Carlo pipeline in the main manuscript reflect the fully recalibrated IASS∗ ceilings; the case study Δ ranges should be treated as illustrative of the VCB methodology rather than as point estimates tied to the final calibration. S2.1 UPS: Logistics Under Structural Transformation UPS’s “Better Not Bigger” strategy, articulated by CEO Carol Tomé in 2020 and evolved into “Better and Bolder” post Teamsters contract (August 2023), explicitly targets automation driven cost reduction UPS (2024). By FY2024, 62 % of U.S. volume processed through automated facilities (up 430 bps YoY); RFID readers installed in ∼ 60,000 package cars; the “Efficiency Reimagined” initiative targets $1.0B additional annualized savings. Labor compensation runs 52–54 % of revenue, the primary value pool. Table S7: UPS: AITG Dimensional Scores (Tier 1) Dimension Score Evidence Signal Data Infrastructure (DIM) 5.0 Cloud investment ongoing; RFID rollout complete Process Automation (PAC) 4.5 62 % automated facilities; Teamster constraints persist Workforce Augment. (WAR) 3.5 AI tools in operations; broad workforce augmentation nascent Decision Auto. (DAR) 4.0 AI routing, pricing, network planning in production AI Product/Revenue (APR) 3.0 Smart logistics features; no standalone AI revenue line Org. AI Capability (OAC) 4.5 Dedicated AI team; no public Chief AI Officer AITG Raw 4.08 IR Score 5.31 (4.08/7.68×104.08/7.68× 10) GeffG_eff 3.60 7.68−4.087.68-4.08 ADRI 4.2 Moderate High; Moat =0.45=0.45 IFS 0.71 Teamster constraints; change capacity limited UQ ±0.54± 0.54 VCB Summary. Primary value pools: labor productivity ($2.4–3.6B addressable), network planning ($0.6–0.9B), dynamic pricing ($0.4–0.7B). Conservative 5 year EBITDA uplift: $4.8–7.2B. Risk adjusted EV creation (exit multiple 9×9×, IFS =0.71=0.71): $28–40B. Implementation cost: $8–14B. AITG-VD: 1.6–2.5×. Tier 2. ADRI context. FedEx, Amazon Logistics, and technology native regional carriers are advancing AI aggressively. UPS’s Teamster constraints materially slow adoption, elevating competitive displacement risk. ADRI =4.2=4.2 (Moderate, transitioning to High). The five year strategic question is whether the “Better and Bolder” automation roadmap executes before AI native competitors compound their operational advantage. S2.2 HCA Healthcare: High Ceiling, Aggressive Deployment HCA is the largest for profit hospital system in the United States ($70.6B FY2024 revenue, 188+ hospitals) HCA Healthcare (2024). Labor costs fell to 44.1 % of revenue (down 130 bps YoY); contract labor fell 25.7 %. HCA has deployed AI across ambient clinical documentation (Nuance DAX Copilot; Commure/Augmedix across 188+ hospitals), revenue cycle automation, clinical decision support, nurse staffing optimization, and patient communication Fierce Healthcare (2024). CFO Mike Marks described the AI investment thesis: “Administrative AI use cases spanning IT, supply chain, HR, revenue cycle showed the shortest pathway to value: millions of transactions, more centralized operations, more standardized data.” Table S8: HCA Healthcare: AITG Dimensional Scores (Tier 1) Dimension Score Evidence Signal Data Infrastructure (DIM) 5.5 Epic EHR; Google Cloud; AWS Bedrock; MLOps emerging Process Automation (PAC) 4.8 Revenue cycle AI deployed; clinical workflows scaling Workforce Augment. (WAR) 4.2 78 % physician DAX Copilot satisfaction; broad rollout Decision Auto. (DAR) 4.0 Staffing AI deployed; sepsis prediction; pricing manual AI Product/Revenue (APR) 2.5 No external AI product; AI improves service quality Org. AI Capability (OAC) 5.0 Dedicated AI teams; CFO publicly leading AI agenda AITG Raw 4.33 IR Score 7.93 (4.33/5.46×104.33/5.46× 10) GeffG_eff 1.13 max(0, 5.46−4.33)=1.13 (0,\;5.46-4.33)=1.13 ADRI 3.8 Moderate; Moat =0.65=0.65 (scale data, regulatory barriers) IFS 0.77 Regulatory friction; facility level data variance UQ ±0.49± 0.49 VCB Summary. Primary value pools: clinical documentation ($800M–1.4B), revenue cycle ($1.1–1.8B), workforce productivity ($0.9–1.4B). 5 year risk adjusted EV creation (exit multiple 12×12×, IFS =0.77=0.77): $35–57B. Implementation cost: $6–11B. AITG-VD: 2.8–4.6×. Tier 1. ADRI context. HCA’s scale creates a data moat via nonrivalry Jones and Tonetti (2020): 400,000+ weekly nurse shift handoffs Fierce Healthcare (2024) generate training data at a rate smaller health systems cannot replicate. ADRI =3.8=3.8 (Moderate) primarily because HCA is already adopting aggressively, reducing the competitive gap risk. S2.3 Salesforce: Narrow Gap, Maximum Value Density Salesforce (FY2025 revenue $37.9B, non GAAP operating margin 33.0%) is the paradigmatic AI native SaaS case: AI is the product, not merely the tool Salesforce (2025). Agentforce launched in Q3 FY2025 with 5,000 deals closed; on help.salesforce.com it handled 380,000 conversations with 84 % resolution rate and 2 % human escalation. Data Cloud & AI ARR reached $900M (up 120 % YoY); all top-10 Q4 wins included Data and AI. SG&A fell from ∼ 72 % to ∼ 57 % of revenue over four years (a 15.9 percentage point improvement). Free cash flow: $12.4B. Table S9: Salesforce: AITG Dimensional Scores (Tier 1) Dimension Score Evidence Signal Data Infrastructure (DIM) 8.5 50T+ records in Data Cloud; Hyperforce global deployment Process Automation (PAC) 8.2 Agentforce; automated service flows at scale Workforce Augment. (WAR) 7.5 Einstein tools across all roles; AI assisted selling Decision Auto. (DAR) 7.8 Einstein GPT; automated service resolution; ML forecasting AI Product/Revenue (APR) 8.2 AI is core product; $900M Data & AI ARR Org. AI Capability (OAC) 8.2 C suite AI mandate; Responsible AI framework published AITG Raw 8.07 IR Score 8.10 (8.07/9.96×108.07/9.96× 10) GeffG_eff 1.89 ADRI 2.1 Low; Moat =0.80=0.80 (switching costs, data network effects) IFS 0.92 Strong data readiness; minimal regulatory friction UQ ±0.38± 0.38 VCB Summary. Primary value pool: NRR expansion from AI native features ($4–8B addressable at 77 % gross margin). Secondary: SG&A efficiency (further 2–4 p possible). 5 year risk adjusted EV creation (ARR multiple 22×22×, IFS =0.92=0.92): $85–140B. Implementation cost: $5–9B (largely existing R&D). AITG-VD: 9.4–15.6×. Tier 1 (High Conviction). Key finding: exit multiple dominance. Salesforce demonstrates the most counterintuitive implication of the AITG framework. Its effective gap is the smallest in the sample (Geff=1.89G_eff=1.89) yet its Value Density is the highest (9.4–15.6×). The 22×22× ARR multiple means every $1 of incremental EBITDA creates $22 of enterprise value. Exit multiple leverage and industry ceiling jointly dominate raw gap size in the VD formula. S2.4 Ferguson Enterprises: Industrial Distribution, Mid Gap Ferguson is the largest U.S. distributor of plumbing and HVAC products (FY2025 revenue $30.8B, gross margin 30.7%) Ferguson (2025). Digital sales represent ∼ 7 % of U.S. revenue. AI powered procurement and predictive inventory have reduced average delivery times 15 %. FY2025 restructuring ($68M charges) targets $100M annualized savings. Digital native competitors (Amazon Business, Grainger’s digital platform) are actively compressing margins. Scores: AITG Raw 4.00 ∣ IR 5.52 ∣ GeffG_eff 3.24 ∣ ADRI 4.2 (Moderate) ∣ IFS 0.79. VCB Summary. 5 year risk adjusted EV creation (exit multiple 10×10×, IFS =0.79=0.79): $8–14B. Implementation cost: $2.5–4.5B. AITG-VD: 1.7–2.9×. Tier 2 (Monitor). Ferguson has a wide effective gap but the second lowest Value Density, driven by a relatively modest exit multiple and a wide gap requiring significant foundation investment before value pool capture can begin. S2.5 Rockwell Automation: Industrial AI Leader Rockwell (FY2025 revenue $8.3B, EBITDA margin 20.8%) occupies a unique dual position: deploying AI internally while embedding it in products sold to industrial customers Rockwell Automation (2024). FactoryTalk Design Studio’s AI Copilot generates ladder logic code; FactoryTalk Analytics VisionAI provides no code quality inspection; LogixAI runs analytics on ∼ 3–4M deployed PLC controllers. An NVIDIA partnership (March 2024) integrates machine vision, learning agents, and AMRs. Rockwell’s own State of Smart Manufacturing Survey finds 83 % of manufacturers expect to use GenAI Rockwell (2024b). Scores: AITG Raw 5.83 ∣ IR 8.05 ∣ GeffG_eff 1.41 ∣ ADRI 3.5 (Low Moderate; Moat =0.70=0.70 on installed PLC base) ∣ IFS 0.86. VCB Summary. 5 year risk adjusted EV creation (exit multiple 18×18× for industrial automation software premium, IFS =0.86=0.86): $9–16B. Implementation cost: $1.8–3.2B. AITG-VD: 2.5–4.4×. Tier 1. S2.6 Goldman Sachs: Investment Banking at the S Curve Inflection Goldman Sachs (FY2025 revenue $54.7B, ROTE 14.2%) operates in the highest IASS industry vertical in the framework (Investment Banking / Securities, IASS=∗10.39^*=10.39). GS has deployed AI across trading (quantitative risk models replacing analyst work), asset management (Marco, its internal LLM for research synthesis), and the consumer pivot through Marcus. Disclosed AI investment run rate exceeds $1.2B/yr, with 35 % of software engineers using AI code generation tools as of Q4 2025 Goldman Sachs (2025). With the AFC adjusted ceiling now at 10.39, the effective gap Geff=2.22G_eff=2.22 positions GS in the productive mid gap zone for its sector. Scores: AITG Raw 8.17 ∣ IR 7.86 ∣ GeffG_eff 2.22 ∣ ADRI 0.6 (Very Low; Moat =0.95=0.95 from proprietary data and regulatory moat) ∣ IFS 0.89. VCB Summary. 5 year risk adjusted EV creation (exit multiple 22×22× banking franchise premium, IFS =0.89=0.89): $30–40B. Implementation cost: $3–4B. AITG-VD: 8.5–13.8×. Tier 1 (HC). Goldman confirms the high IASS pattern: Geff=2.22G_eff=2.22 positions GS in the productive mid gap zone in the sector with the highest capability ceiling, generating near Salesforce VD at substantially larger absolute scale. S2.7 Wells Fargo: IFS Suppression in a High Ceiling Industry Wells Fargo (FY2025 revenue $82.3B, ROTCE 11.9%) operates under the same AFC adjusted IASS ceiling as JPMorgan and Zions (9.38) but its effective gap (Geff=3.38G_eff=3.38) and IFS (0.65) reflect a structurally impaired transformation capacity. The Federal Reserve consent order (lifted partially in 2024, full resolution pending) constrains risk system overhaul sequencing, delaying AI deployment in credit decisioning and fraud detection, the two highest value pools in banking. WFC has committed $3.1B to technology modernization but the organizational change capacity (OCC) dimension registers at 5.5/10 due to the compliance first governance overhead Wells Fargo (2025). Scores: AITG Raw 6.00 ∣ IR 6.40 ∣ GeffG_eff 3.38 ∣ ADRI 2.3 (Moderate) ∣ IFS 0.65. VCB Summary. 5 year risk adjusted EV creation (exit multiple 9×9×, IFS =0.65=0.65): $1.4–2.1B. Implementation cost: $0.9–1.2B. AITG-VD: 1.2–2.0×. Tier 2. Wells Fargo is the paper’s most important demonstration of IFS suppression: a wide gap in a high ceiling industry (Geff=3.38G_eff=3.38, IASS=∗9.38^*=9.38) that should theoretically produce high VD is compressed to near Zions returns purely by an IFS of 0.65, a regulatory constraint that directly caps execution velocity. S2.8 ServiceNow: Second SaaS Data Point Confirming the Sweet Spot ServiceNow (FY2025 revenue $12.5B, operating margin 27.1%) operates in Vertical SaaS (IASS=∗9.96^*=9.96), the same industry as Salesforce. Its Now Intelligence platform, GenAI powered workflow automation, and AI Agents for ITSM/HRSD position it squarely in the near apex zone (Geff=1.46G_eff=1.46). Unlike Salesforce, ServiceNow has not yet fully converted its AI capability into a comparable data network moat, placing its IFS at 0.88 vs. Salesforce’s 0.92 ServiceNow (2025). Scores: AITG Raw 8.50 ∣ IR 8.53 ∣ GeffG_eff 1.46 ∣ ADRI 1.8 (Low Moderate) ∣ IFS 0.88. VCB Summary. 5 year risk adjusted EV creation (exit multiple 25×25× SaaS premium, IFS =0.88=0.88): $18–25B. Implementation cost: $1.5–2.5B. AITG-VD: 6.2–9.8×. Tier 1. ServiceNow confirms that Salesforce’s VD is not an outlier: both high IASS SaaS firms with medium gaps generate VD an order of magnitude above the industrial distribution cohort, validating the industry ceiling dominance finding. S2.9 Target: Retail at the Inflection Target (FY2025 revenue $109.1B, EBIT margin 5.3%) operates in E Commerce / Digital Retail (IASS=∗8.64^*=8.64), a sector where AI driven inventory management, dynamic pricing, and personalization are materially reducing out of stock rates and increasing basket size. Target’s AI powered supply chain system Drive-Up Plus and team member scheduling tools represent mature Wave 1 deployments Target (2025). At Geff=3.97G_eff=3.97, Target sits precisely in the sweet spot zone: past the expensive foundation build, approaching the S curve inflection in Wave 2 (personalization / recommendation) deployment. Scores: AITG Raw 4.67 ∣ IR 5.41 ∣ GeffG_eff 3.97 ∣ ADRI 3.2 (Moderate; Amazon pressure) ∣ IFS 0.82. VCB Summary. 5 year risk adjusted EV creation (exit multiple 12×12×, IFS =0.82=0.82): $11–17B. Implementation cost: $2–3B. AITG-VD: 3.8–5.6×. Tier 1. S2.10 CVS Health: Same Regulatory Ceiling, Different IFS: Healthcare Industry Test CVS Health (FY2025 revenue $371.9B, adjusted operating income $13.4B) operates in Healthcare Services (IASS=∗5.46^*=5.46), the same regulatory ceiling as HCA. The comparison isolates the IFS effect: CVS’s PBM business complexity and ongoing Aetna integration create organizational change capacity friction (OCC = 5.0/10) that suppresses IFS to 0.72 vs. HCA’s 0.77 CVS Health (2025). Both face the identical ψ=0.743ψ=0.743 RFF ceiling. Scores: AITG Raw 4.83 ∣ IR 8.85 ∣ GeffG_eff 0.63 ∣ ADRI 3.1 (Moderate) ∣ IFS 0.72. VCB Summary. 5 year risk adjusted EV creation (exit multiple 8×8× healthcare services, IFS =0.72=0.72): $7–10B. Implementation cost: $2.5–3.5B. AITG-VD: 2.1–3.3×. Tier 2. CVS vs. HCA is the paper’s cleanest within industry IFS test: same IASS∗, comparable GeffG_eff (0.63 vs. 1.13), but CVS’s lower IFS (0.72 vs. 0.77) and lower exit multiple produce materially weaker VD, confirming IFS’s role as a first order VD determinant within a regulated industry. S2.11 Palo Alto Networks: Near Apex in a High IASS Sector Palo Alto Networks (FY2025 revenue $8.0B, free cash flow margin 37%) operates in Cybersecurity (IASS=∗9.65^*=9.65), one of the highest ceiling sectors in the framework. PANW’s Precision AI platform, Cortex XSIAM, and AI Security Operations (SecOps) represent near complete gap closure (Geff=1.82G_eff=1.82), with AI deeply embedded across detection, response, and prevention Palo Alto Networks (2025). This creates a different VD dynamic than JPMorgan’s near apex position: PANW’s smaller scale (vs. JPM) means incremental AI value capture still generates meaningful absolute EV, while the high IASS ceiling preserves real returns even at low GeffG_eff. Scores: AITG Raw 7.83 ∣ IR 8.11 ∣ GeffG_eff 1.82 ∣ ADRI 1.2 (Low) ∣ IFS 0.93. VCB Summary. 5 year risk adjusted EV creation (exit multiple 22×22× cybersecurity software premium, IFS =0.93=0.93): $12–17B. Implementation cost: $1.1–1.5B. AITG-VD: 4.5–7.2×. Tier 1 (HC). PANW demonstrates that near apex position in a high IASS sector preserves meaningful VD (∼ 5.9× midpoint) unlike JPMorgan’s near apex position (∼ 1.05×). The IASS ceiling dominance effect: the same GeffG_eff regime generates radically different VD depending on the sector ceiling. S2.12 Ford Motor: Wide Gap, Low Ceiling, Low IFS: Industrial Constraint Case Ford (FY2025 revenue $185.0B, EBIT margin 3.0%) provides the framework’s clearest demonstration of the wide gap low capture pattern in Discrete Manufacturing (IASS=∗7.49^*=7.49). Ford’s AI investments (generative AI for vehicle design, predictive quality control, Blue Oval Intelligence in vehicle AI) are real but deployment is constrained by UAW labor agreements, legacy manufacturing system integration requirements, and a capital structure under EV transition pressure Ford Motor (2025). IFS registers at 0.65, identical to Wells Fargo, though for entirely different structural reasons: labor relations rather than a regulatory consent order. Scores: AITG Raw 4.67 ∣ IR 6.23 ∣ GeffG_eff 2.82 ∣ ADRI 4.8 (High; Tesla and EV native competitors) ∣ IFS 0.65. VCB Summary. 5 year risk adjusted EV creation (exit multiple 5×5× auto manufacturing, IFS =0.65=0.65): $1.3–2.5B. Implementation cost: $1.5–2.5B. AITG-VD: 1.1–1.8×. Tier 3. Ford is the widest gap, lowest VD case in the 14 company sample, illustrating the wide gap fallacy in its starkest form: Geff=2.82G_eff=2.82, yet VD barely above 1.0×. The combination of a moderate IASS ceiling (7.49), low IFS (0.65), and a compressed exit multiple (5×5×) produces near zero economic return on AI transformation cost. Appendix S3 ESM Part I: Extended Robustness and Sensitivity Analysis The main manuscript reports summary results from the robustness analysis (Main Paper Section 10, “Robustness Summary”). This section provides full methodological detail, equations, and tables. S3.1 Monte Carlo Weight Sensitivity Following OECD best practice OECD (2008), I test IASS ranking stability under random perturbations of the weight vector. I draw M=10,000M=10,000 weight vectors (m)w^(m) uniformly from the simplex subject to wd∈[wdbase−0.05,wdbase+0.05]w_d∈[w_d^base-0.05,\;w_d^base+0.05] for each dimension d. The average rank shift statistic is: R¯s=1I∑i=1I|Ribase−R¯i(m)| R_s= 1I _i=1^I |R_i^base- R_i^(m) | (47) For the five anchor industries under 10,000 draws with ±5%± 5\,\% weight perturbations, the mean absolute rank shift is R¯s=0.19 R_s=0.19 positions (out of 5). No pair of anchor industries exchanges rank with probability >5%>5\,\% under any weight draw. This confirms robustness: the rank orderings are determined by the data, not by weighting choices within reasonable bounds. S3.2 Sobol Sensitivity Analysis I apply first order and total effect Sobol sensitivity indices Saltelli (2002) to the VCB AITG-VD output. For Zions Bancorporation, the main effects are: Sd=VarXd[X∼d(AITG-VD∣Xd)]Var(AITG-VD)S_d= Var_X_d\! [E_X_ d(AITG -VD X_d) ]Var(AITG -VD) (48) Decomposition of Var(AITG-VD)Var(AITG -VD): exit multiple 50 %; capture rate assumption 44 %; implementation cost 1 %; AITG gap score <<1 %; IFS 5 %. This result confirms that exit multiple leverage and capture rate assumptions dominate raw gap size as sources of uncertainty in the output, precisely the pattern that explains Finding 1. For practical due diligence, the critical inputs to verify are exit comparable multiples and the realistic capture rate in the three largest value pools. S3.3 Normalization Stability I test substituting z score normalization for min max across all dimensions and find that IASS rankings shift by ≤0.40≤ 0.40 points in all cases and that no pair of anchor industries exchanges rank. This confirms that the IASS is not an artifact of the normalization choice. S3.4 Geometric vs. Linear Aggregation Substituting geometric aggregation for linear (fully compensatory) in the IASS produces IASS scores consistently 0.200.20–0.500.50 points lower for industries with uneven dimensional profiles (healthcare, construction) and negligibly different for industries with uniform profiles (SaaS). The geometric specification is more conservative and should be used when the analyst seeks a defensible lower bound, but both specifications produce the same rank ordering. S3.5 Inter Rater Reliability: ICC Protocol and Target Composite indicator frameworks are only as reliable as their inter rater agreement (Krippendorff, 2018). I specify the formal intraclass correlation coefficient (ICC) protocol required to certify the AITG scoring instrument. ICC specification. I target ICC(2,1) (two way mixed effects, absolute agreement, single rater), following Shrout and Fleiss (1979): ICC(2,1)=MSB−MSWMSB+(k−1)MSW+k(MSBA−MSW)/nICC(2,1)= MS_B-MS_WMS_B+(k-1)MS_W+k(MS_B^A-MS_W)/n (49) where MSBMS_B is between company variance, MSWMS_W is within company (between rater) variance, k is raters per company, and n is companies. Target: ICC(2,1) ≥0.85≥ 0.85 for each dimension and ≥0.88≥ 0.88 for the overall AITG composite. Proposed validation protocol. A formal reliability study would require: (1) 3–5 independent analyst teams; (2) 12–15 target companies spanning at least 3 industries and the full AITG range; (3) all teams working from identical public data packages; (4) ICC computed per dimension, per IFS factor, and for the composite. This study has not yet been conducted and constitutes the highest priority validation item in the research agenda. Pilot estimate. From two internal analyst applications of the framework on the original seven company cohort (Section 9), preliminary inter rater agreement (Kendall’s τ) for dimension level rank ordering was τ=0.82τ=0.82 for DIM, PAC, and OAC, and τ=0.68τ=0.68 for DAR and APR (the two dimensions most dependent on interpretive inference from earnings call language). These pilot estimates suggest that public data only scoring of the two most inference dependent dimensions will require rubric refinement before the ICC target is achieved. S3.6 Convergent Validity Test Convergent validity asks: do AITG scores correlate with independent external measures of the same construct? I test this against two observable proxies. Test 1: Stanford HAI AI Index corporate rankings. The Stanford Human Centered AI (HAI) 2025 AI Index Stanford HAI (2025) ranks Fortune 500 companies by AI investment intensity, measured by AI patent filings, AI job posting share, and AI research paper authorship. Extending to the fourteen company cohort, AITG rankings agree with the Stanford HAI rankings on direction for 12 of 14 companies. The two divergences share the same mechanism: HCA Healthcare and Ford Motor both rank higher in HAI due to extensive AI research output (clinical AI publications and EV/autonomous driving patents respectively), while AITG assigns lower scores because its RFF penalty and IFS architecture explicitly discount research activity that has not translated to deployment at scale under regulatory or labor constraints. This constitutes partial convergent validity across a fourteen company, eight industry sample, with theoretically explicable divergences. Test 2: Lightcast AI job posting concentration. The Lightcast AI Skill Demand Index measures the share of job postings in each industry requiring at least one AI specific skill. This is an independent measure of an industry’s AI adoption intensity, not used in IASS calibration for CADR (where Lightcast data is used for industry level benchmarks, not company level scoring). Extending to the fourteen company cohort, the rank correlation between AITG and company level Lightcast AI job posting share is rs=0.88r_s=0.88 (Spearman’s ρ, p<0.001p<0.001, n=14n=14), a strengthening of the seven company estimate (rs=0.81r_s=0.81, p<0.05p<0.05) as the sample expanded across industries. This represents strong convergent validity. Discriminant validity. AITG scores should not correlate with total revenue (because a large laggard should score below a small leader). Spearman correlation between AITG and revenue rank across the fourteen companies: rs=−0.28r_s=-0.28 (not significant, p=0.33p=0.33). The near zero correlation (with no size driven directionality) confirms that AITG is measuring transformation state, not firm size. CVS Health (largest by revenue at $372B) scores 4.83; Palo Alto Networks (smallest at $8B) scores 7.83. This is a necessary property for cross industry comparability. S3.7 Cross Sectional Discriminability A useful framework must discriminate meaningfully between companies in the same industry. I test this formally for the within industry pair (JPMorgan, AITG =8.22=8.22 vs. Zions, AITG =3.80=3.80). The AITG gap of 4.42 points must exceed the joint Uncertainty Quotient to be a statistically credible distinction: ΔAITG=4.42vs.UQJPM2+UQZION2=0.482+0.622=0.78 =4.42 . UQ_JPM^2+UQ_ZION^2= 0.48^2+0.62^2=0.78 (50) The signal to noise ratio is 4.42/0.78=5.74.42/0.78=5.7, well above the conventional threshold of 2.0 for a statistically credible distinction. This confirms that the JPM/Zions discrimination is not an artifact of measurement uncertainty for these two companies at Tier 1 (public data) scoring. For the cross industry comparison (Zions AITG =3.80=3.80 vs. UPS AITG =4.08=4.08), the raw gap is Δ=0.28 =0.28, compared to joint UQ =0.622+0.542=0.82= 0.62^2+0.54^2=0.82. The signal to noise ratio is 0.28/0.82=0.340.28/0.82=0.34: the two companies are not significantly distinguishable on raw AITG with public data alone. However, once IR scores are applied (Zions IR =4.05=4.05 vs. UPS IR =5.31=5.31), the normalized gap of 1.26 IR points exceeds the joint UQ on the IR scale (≈0.97≈ 0.97) at a signal to noise ratio of ≈ 1.3≈\,1.3. The improvement from S/N=0.34S/N=0.34 (raw) to S/N≈1.3S/N≈ 1.3 (IR normalized) is precisely the behavior the framework is designed to produce: raw AITG comparisons between different industries are unreliable; IR normalized comparisons are more discriminating. S3.8 Predictive Validity: Framework and Status Table S10: Predictive Validity Test Battery: Status and Proposed Design Test Null Hypothesis Proposed Design Status AITG predicts subsequent margin improvement H0H_0: No AITG margin correlation Retrospective: compute 2020 AITG from 2020 10-K data; regress on 2020–2024 EBITDA margin change; 50+ company panel Not yet run ADRI predicts competitive share loss H0H_0: ADRI uncorrelated with share loss High ADRI firms (>3.0>3.0) in Commerce, Logistics lose revenue share within 24 mo; IV: CADR shock via LLM release event Proposed IFS predicts transformation failure H0H_0: IFS uncorrelated with AI project cancellation Lightcast AI job posting drop as proxy for abandoned programs; low IFS firms show higher cancellation rate Proposed Cross industry VD ranking is ordinal consistent H0H_0: AITG-VD does not predict actual deal multiples paid PE transaction database: AI thesis deals vs. non thesis deals; AITG scores predicted from pre deal public data Proposed Current status. The margin prediction test (row 1 of Table S10) has now been partially executed as a retrospective backtest on the ten non financial companies in the fourteen firm cohort; results are reported in Section 10.1 of the main paper. AITG scores reconstructed from Q4 2021 public filings predict FY2021–FY2023 EBITDA margin changes with Spearman ρs=0.818 _s=0.818 (p<0.001p<0.001, n=10n=10); within sector directional accuracy is 4 of 4 non confounded pairs. The limitations of this preliminary test are noted explicitly in Section 10.1: n=14n=14 is illustrative rather than confirmatory, same team retrospective scoring introduces potential look back bias, three observations are confounded by non AI macro events, and endogeneity is not resolved. Tests 2–4 in Table S10 remain in the proposed/designed stage and are the highest priority empirical research program implied by this framework. The full validation agenda requires a multi year panel study as specified in the Research Agenda (main paper Section 11.2). S3.9 Engagement with Acemoglu (2024) Acemoglu (2024) estimates that under realistic scenarios, AI raises aggregate TFP by at most 0.66 % over ten years. This result is not in conflict with the AITG’s firm level value estimates. Three distinctions reconcile them: 1. Heterogeneity vs. average. Aggregate TFP is a population weighted average that includes non adopters, laggards, and failed implementations. Syverson (2011)’s finding of a 1.92× 90th/10th percentile TFP spread within industries implies that leading quartile firms capture substantially more than the aggregate average. AITG targets above median firms or identifies acquisition targets for above median capture. 2. Competitive advantage vs. aggregate surplus. AITG measures value creation relative to competitors, not the aggregate social surplus. A firm can capture significant competitive margin advantage even in a world where aggregate TFP gains are modest, if it gains share from laggards through lower costs, faster service, or superior product Autor et al. (2020). 3. Static capability ceiling. Acemoglu’s estimate is for the current technology class. The AFC mechanism captures upward revision of the opportunity as capabilities advance, a dynamic that a static macro model cannot accommodate. If the AI capability improvement trajectory observed in 2022–2024 continues, the addressable task surface expands materially beyond what current cost reduction estimates imply. Appendix S4 ESM Part IV: Implementation Appendices This part supports Main Paper Sections 7–8 (VCB and IFS) and provides practitioner implementation tools. Appendix S5 Public Data Availability Map A central practitioner question is: how much of an AITG score for a publicly traded company can be populated from freely available sources, without a diligence data room? This appendix provides a systematic answer by mapping each of the six company dimensions and five IFS factors to their primary public sources, estimating coverage achievable from public data alone, and specifying what requires primary diligence. S5.1 Dimension by Dimension Public Data Coverage Table S11 shows, for each AITG scoring component, (a) the primary public data sources, (b) the estimated fraction of the scoring range reliably determinable from those sources, and (c) what specifically requires primary diligence to resolve the remaining uncertainty. Table S11: AITG Scoring Component: Public Data Coverage Map Component Primary Public Sources Diligence Gap Public % Tier Company Dimension Scores DIM (Data Infra.) Tech budget disclosures (10-K MD&A); cloud provider announcements; IT vendor press releases; EDGAR risk factors mentioning legacy systems Actual data quality, schema maturity, MLOps pipeline state 50–65 % B/C PAC (Process Auto.) Earnings call AI deployment counts; investor day demos; vendor partnership announcements (e.g., nCino, Salesforce, SAP) Process level automation depth; proportion of workflows vs. surface coverage 55–70 % B/C WAR (Workforce Aug.) Headcount disclosures; AI platform adoption figures (when disclosed); Lightcast AI skill demand by employer; LinkedIn job postings with AI tool requirements Actual employee usage rate vs. license provisioning rate 45–60 % B/C DAR (Decision Auto.) SEC filings describing automated underwriting, pricing, or routing systems; vendor AI product deployment announcements; patent filings (Google Patents) Decision quality metrics; actual automation rate vs. human in loop proportion 40–55 % C APR (AI Revenue) Product documentation; AI feature release notes; investor day product demos; sell side analyst coverage of AI driven pricing; ARR growth in AI segments Revenue attribution by AI enabled feature; AI product margin vs. legacy 60–75 % B/C OAC (Org. Capability) CAIO/CDAO appointment announcements; AI governance policy filings; AI staff count estimates from LinkedIn; published AI ethics frameworks; Glassdoor/Blind AI culture signals Change management track record; prior transformation success rates; internal AI governance maturity 50–65 % B/C IFS Factors OCC (Org. Change Cap.) Prior transformation announcements and outcomes; attrition data (10-K headcount trends); workforce restructuring history; American Banker / HBR culture awards; union contract disclosures Frontline adoption resistance; change culture ground truth 40–55 % B/C DR (Data Readiness) Cloud migration progress (vendor announcements); data governance program disclosures; CIO/CDO tenure and mandate; tech audit findings (when disclosed); O*NET/BLS data infrastructure benchmarks by sector Actual data quality scores; schema fragmentation depth; real time pipeline health 35–50 % C VTR (Vendor/Tech Risk) Vendor partnership concentration (10-K); material AI dependency disclosures; SOC 2 Type I certifications; public breach/outage history; AI insurance disclosures Model dependency in production; proprietary vs. API accessed model split 65–80 % A/B CRS (Competitive Resp.) Lightcast CADR by industry; industry press on competitor AI deployments; CB Insights AI deal data; earnings call competitor references; BTOS adoption rates by sector Competitor specific adoption velocity; pricing response timeline 70–85 % A/B REG (Regulatory Exp.) NIST AI RMF sector classification; EU AI Act high risk list; HIPAA applicability; OCC/Fed model risk guidance; FINRA regulatory notices; FDA SaMD guidance documents Company specific regulatory exam history; model risk management scores 75–90 % A/B Financial Baseline (for VCB) Revenue, EBITDA EDGAR iXBRL 10-K/10-Q filings; Bloomberg/Refinitiv – 100 % A Labor/Rev. Ratio BLS OEWS sector benchmarks; 10-K headcount × median wage Company specific compensation mix 85 % A/B Working Capital EDGAR iXBRL balance sheet – 100 % A Exit Multiple Public comparable company analysis; capital IQ; PitchBook Deal structure; NTM vs. LTM basis 90 % A/B WACC Damodaran sector betas; risk free rate; sector capital structure Company specific debt structure for private targets 90 % A/B % = fraction of scoring range reliably determinable from public sources; remainder requires diligence. S5.2 Aggregate Public Data Coverage by Company Type Table S12: Estimated AITG Scoring Coverage: Public Data vs. Diligence Grade Company Type Public Only % UQ Penalty Primary Gap Large cap public (S&P 500) 60–75 % +0.15+0.15–0.250.25 Depth behind headline metrics Mid cap public ($1–10B) 45–60 % +0.20+0.20–0.350.35 Limited investor day disclosure Small cap public (<<$1B) 30–45 % +0.30+0.30–0.500.50 Minimal AI specific disclosure Private (PE diligence) 10–20 % pre D +0.45+0.45–0.600.60 Requires full data room + survey UQ Penalty = incremental Uncertainty Quotient contribution from public data only scoring. S5.3 Fourteen Company Public Data Audit For the fourteen companies scored in Section 9, I retroactively classify the evidence tier of each dimension score: Table S13: Fourteen Company Evidence Tier Audit (Tier A–D per Appendix S6) Company DIM PAC WAR DAR APR OAC OCC DR VTR CRS Original cohort JPMorgan Chase A A A A/B A/B A A A A A Zions Bancorp. C C C/D C D C C C/D B B UPS B B B B/C B B B B/C A A HCA Healthcare B B B B/C B/C B B C A A Salesforce A A A A A A A A A A Ferguson Ent. C C C C C C C C B A Rockwell Auto. B B B B B B B B A A Extended cohort Goldman Sachs A A A A A/B A A A A A Wells Fargo A B B B B A B B A A ServiceNow A A A A A A A A A A Target B B B B B B B B A A CVS Health B B B C B B B C A A Palo Alto Ntwks A A A A A A A A A A Ford Motor B B B B B B B C A A A = public structured data; B = public unstructured/alt. data; C = indirect inference; D = absence of evidence. Extended cohort uses Tier 1 (public only) scoring throughout; all seven are S&P 500 constituents with extensive 10-K/investor day AI disclosure. The fourteen company evidence tier audit reveals two systematic patterns that practitioners should internalize before applying the framework. Pattern 1: Disclosure quality tracks market cap and AI centrality, not industry. The seven extended cohort companies are all S&P 500 constituents with AI as a disclosed strategic priority, and accordingly achieve near complete Tier A/B coverage across all ten dimensions. Goldman Sachs, Salesforce, ServiceNow, and Palo Alto Networks reach Tier A on virtually every dimension because their AI deployment activity is commercially material enough to generate structured investor day disclosure, product documentation, and regulatory filing specificity. This pattern holds regardless of industry: PANW (cybersecurity) and GS (investment banking) both achieve the same coverage quality as Salesforce (SaaS) because all three treat AI capability as a product facing disclosure obligation. Pattern 2: The hardest Tier C/D cells cluster in two consistent locations. First, DAR (Decision Automation Rate) consistently drops to B/C across industries because the proportion of decisions that are fully automated vs. human in loop is operationally sensitive and rarely disclosed. Second, DR (Data Readiness) drops to B/C for companies undergoing major integration events (CVS post Aetna, WFC under consent order, HCA across 186 hospitals) because data fragmentation at scale cannot be assessed from outside. The Zions and Ferguson scores retain the highest Uncertainty Quotients (±0.62± 0.62 and ±0.55± 0.55 respectively) in the full sample, driven by Tier C/D reliance on absence of evidence inference. No extended cohort company requires this inference because all seven voluntarily disclose sufficient AI specific activity. Practitioner implication. A private equity firm conducting diligence on a Zions equivalent regional bank or Ferguson equivalent industrial distributor should plan for 8–10 weeks of primary data collection (IT infrastructure audit, management survey administration, data platform assessment, and vendor contract review) to reduce the UQ to diligence grade levels. For the extended cohort, public only scoring achieves 65–80 % coverage (Table S12), reducing primary diligence to targeted gap filling rather than full scope collection. The scoring template in the companion implementation document operationalizes both collection protocols. Appendix S6 Data Tier Classification AITG scoring employs four evidence tiers that directly determine the Uncertainty Quotient: Table S14: Data Evidence Tier Definitions Tier Source Type Examples Δ A Public structured data EDGAR XBRL financials, BLS OEWS, O*NET 0.000.00 B Public unstructured / alt. data Earnings transcripts, Lightcast, job postings +0.08+0.08 C Primary diligence (verified) Management survey with evidence attachments +0.18+0.18 D Primary diligence (self reported) Unverified management assertion +0.30+0.30 Appendix S7 25 Question Management Survey Instrument The following standardized 25 question survey instrument enables practitioners to generate the six AITG dimension scores and IFS adjustment factors for any firm. Each question uses a single select 0–4 response scale mapped to fixed score anchors (1, 3, 5, 7, 9). For any response ≥3≥ 3 (score ≥7≥ 7), the assessor must attach corroborating evidence (SEC filings, vendor contracts, internal dashboards, or third party benchmarks). Without attached evidence, the response is capped at 2 (score =5=5) and flagged as Tier D (see Table S14). Scoring protocol. Each dimension’s raw score is the average of its constituent question scores. AITGrawf_f^raw is the arithmetic mean of the six dimension scores (Equation 51). Dimension 1: Data Infrastructure Maturity (DIM) Q1. Cloud data platform adoption. What share of enterprise data resides on a governed cloud platform (data lake, lakehouse, or warehouse)? 0 No cloud data platform; siloed legacy databases only [Score 1] 1 Initial migration underway; <<25% of data on cloud [Score 3] 2 25–60% on governed cloud platform; partial data catalog [Score 5] 3 60–90% on cloud; unified catalog with lineage tracking [Score 7] 4 >>90% cloud native; automated data quality >>99% [Score 9] Q2. Data integration and accessibility. How unified and accessible is enterprise data for cross functional analytics and ML feature engineering? 0 Data locked in departmental silos; no shared schema [Score 1] 1 Some shared reporting tables; manual ETL processes [Score 3] 2 Central data warehouse with governed access; basic feature store [Score 5] 3 Real time streaming pipelines; ML feature store in production [Score 7] 4 Enterprise knowledge graph; self serve analytics across all BUs [Score 9] Q3. Data governance and quality. To what extent are formal data governance policies (ownership, quality SLAs, privacy controls) enforced across the organization? 0 No formal data governance; ad hoc quality checks [Score 1] 1 Basic data dictionary exists; governance limited to finance [Score 3] 2 Enterprise wide governance framework; data stewards assigned [Score 5] 3 Automated quality monitoring; PII/PHI controls audited quarterly [Score 7] 4 Continuous data observability platform; regulatory compliance automated [Score 9] Q4. AI/ML infrastructure readiness. What compute and MLOps infrastructure is available for training, deploying, and monitoring AI models? 0 No GPU/TPU access; models run on analyst laptops [Score 1] 1 Cloud compute available on request; no CI/CD for models [Score 3] 2 Managed ML platform (e.g., SageMaker, Vertex); basic model registry [Score 5] 3 Full MLOps: automated retraining, A/B serving, drift detection [Score 7] 4 Multi cloud ML platform; model versioning, lineage, and compliance at scale [Score 9] Dimension 2: Process Automation Coverage (PAC) Q5. Breadth of process automation. What percentage of repetitive, rules based business processes are automated (RPA, workflow engines, or intelligent automation)? 0 <<5% of processes automated; predominantly manual workflows [Score 1] 1 5–15% automated; RPA pilots in back office functions [Score 3] 2 15–40% automated; cross functional RPA deployment [Score 5] 3 40–65% intelligent automation; AI augmented process orchestration [Score 7] 4 >>65% end to end automation; self optimizing process loops [Score 9] Q6. Complexity of automated tasks. What is the most complex type of task routinely handled by automation (structured, semi structured, or unstructured)? 0 Only simple data entry and file transfers [Score 1] 1 Structured rules: invoice matching, report generation [Score 3] 2 Semi structured: document extraction, email classification [Score 5] 3 Unstructured: NLP driven customer interaction, image analysis [Score 7] 4 Multi step agentic workflows: autonomous exception handling [Score 9] Q7. Automation monitoring and optimization. How are automated processes monitored, measured, and improved over time? 0 No monitoring; failures discovered manually [Score 1] 1 Basic error logs reviewed weekly [Score 3] 2 Centralized automation dashboard; SLA tracking [Score 5] 3 Real time process mining; automated bottleneck identification [Score 7] 4 Closed loop optimization: automation adjusts parameters autonomously [Score 9] Q8. Cross functional automation integration. To what degree are automated workflows integrated across departments (finance, operations, HR, customer service)? 0 Automation isolated to single department [Score 1] 1 Two departments share some automated handoffs [Score 3] 2 3–4 departments with integrated automation; shared orchestration layer [Score 5] 3 Enterprise wide automation fabric; API connected across all functions [Score 7] 4 Fully integrated digital twin of operations; autonomous cross BU optimization [Score 9] Dimension 3: Workforce AI Augmentation Rate (WAR) Q9. AI tool adoption breadth. What percentage of employees regularly use AI powered tools (copilots, assistants, recommendation engines) in their daily work? 0 <<5% of employees use any AI tool [Score 1] 1 5–15% with basic AI tools (e.g., email copilot) [Score 3] 2 15–40% augmented; formal adoption metrics tracked [Score 5] 3 40–65% AI augmented; integrated into standard workflows [Score 7] 4 >>65% augmented; org wide AI fluency benchmarked annually [Score 9] Q10. AI training and upskilling investment. What structured AI literacy and upskilling programs exist for the workforce? 0 No AI training programs; learning is ad hoc [Score 1] 1 Optional online courses; <<10% completion rate [Score 3] 2 Mandatory AI literacy modules; role specific tracks [Score 5] 3 Continuous AI upskilling tied to performance reviews; certification paths [Score 7] 4 AI academy with external partnerships; fluency measured and rewarded [Score 9] Q11. Human–AI collaboration design. Are workflows explicitly designed for human–AI collaboration (not just human or machine)? 0 No deliberate human–AI workflow design [Score 1] 1 AI outputs reviewed by humans; no feedback loop [Score 3] 2 Defined human in the loop checkpoints for AI recommendations [Score 5] 3 Co designed workflows: humans set objectives, AI executes and escalates [Score 7] 4 Adaptive teaming: AI reallocates tasks based on human cognitive load [Score 9] Q12. Change management for AI adoption. How effectively does the organization manage resistance and cultural barriers to AI adoption? 0 Significant resistance; no change management program [Score 1] 1 Executive communications about AI strategy; limited follow through [Score 3] 2 Dedicated change management team; champion network in key BUs [Score 5] 3 Structured incentives for AI adoption; measurable culture shift [Score 7] 4 AI first culture embedded; innovation time allocated across all roles [Score 9] Dimension 4: Decision Automation Rate (DAR) Q13. Share of decisions AI assisted or automated. What fraction of recurring operational and strategic decisions are AI recommended or fully automated? 0 All recurring decisions are fully human [Score 1] 1 <<10% of decisions AI assisted (e.g., pricing alerts) [Score 3] 2 10–30% AI recommended; human approval required [Score 5] 3 30–55% automated with human override capability [Score 7] 4 >>55% fully automated; exception based human review only [Score 9] Q14. Decision model governance. How are AI decision models validated, monitored, and governed? 0 No model governance; outputs unchecked [Score 1] 1 Basic accuracy tracking; annual model review [Score 3] 2 Model risk framework; bias testing before deployment [Score 5] 3 Continuous monitoring: drift, fairness, and explainability dashboards [Score 7] 4 Board level model risk committee; automated retraining triggers [Score 9] Q15. Real time decision capability. Can AI systems make decisions in real time (sub second) for time critical operations? 0 All decisions batch processed or manual [Score 1] 1 Near real time dashboards; humans act on alerts [Score 3] 2 Real time scoring for 1–2 use cases (e.g., fraud detection) [Score 5] 3 Real time AI across multiple domains (pricing, routing, risk) [Score 7] 4 Autonomous real time decisions across operations; adaptive policies [Score 9] Q16. Decision automation scope. Across how many functional areas (finance, supply chain, marketing, risk, HR) are AI driven decisions deployed? 0 None; all decisions manual across functions [Score 1] 1 1 functional area (typically finance or marketing) [Score 3] 2 2–3 functional areas with production AI decision systems [Score 5] 3 4–5 areas; integrated decision layer across operations [Score 7] 4 All major functions; enterprise wide AI decision fabric [Score 9] Dimension 5: AI Product/Revenue Integration (APR) Q17. AI revenue attribution. What share of total revenue is directly attributable to AI powered products, features, or pricing optimization? 0 No measurable AI revenue attribution [Score 1] 1 <<5% revenue from AI enhanced features [Score 3] 2 5–15% revenue AI attributable; tracked in financial reporting [Score 5] 3 15–35% revenue from AI driven products or dynamic pricing [Score 7] 4 >>35% revenue from AI native products and services [Score 9] Q18. AI product development pipeline. How mature is the pipeline for developing and launching AI powered products or features? 0 No AI product roadmap; R&D is non AI [Score 1] 1 1–2 AI product experiments; no shipping cadence [Score 3] 2 AI features in product roadmap; quarterly release cycle [Score 5] 3 Dedicated AI product team; rapid experimentation (monthly launches) [Score 7] 4 AI first product strategy; continuous deployment with A/B testing [Score 9] Q19. Customer facing AI capabilities. What AI powered customer experiences are deployed at scale (personalization, conversational AI, predictive service)? 0 No customer facing AI [Score 1] 1 Basic chatbot or recommendation widget; low adoption [Score 3] 2 Personalization engine or intelligent search in production [Score 5] 3 Multi channel AI CX: conversational AI, predictive service, dynamic UX [Score 7] 4 Autonomous AI agents handling end to end customer journeys [Score 9] Q20. AI monetization strategy. Does the firm have an explicit strategy for monetizing AI capabilities (pricing power, new revenue streams, platform effects)? 0 No AI monetization strategy [Score 1] 1 AI used for internal cost savings only [Score 3] 2 AI pricing premium on existing products; early revenue tracking [Score 5] 3 AI as a service offering or platform with third party integrations [Score 7] 4 AI platform economics: data network effects driving >>20% margin expansion [Score 9] Dimension 6: Organizational AI Capability (OAC) Q21. AI leadership and governance structure. What level of dedicated AI leadership exists within the organization? 0 No dedicated AI leadership; projects run ad hoc by IT [Score 1] 1 AI team lead or director within IT or analytics [Score 3] 2 VP level AI leader; C suite AI sponsor identified [Score 5] 3 Chief AI Officer or equivalent; published enterprise AI strategy [Score 7] 4 Board level AI committee; AI governance integrated into risk framework [Score 9] Q22. AI talent density. What is the density and caliber of AI/ML engineering talent relative to the firm’s technology workforce? 0 No dedicated AI/ML engineers [Score 1] 1 Small team (<<10); reliant on external consultants [Score 3] 2 Established AI center of excellence; 10–50 ML engineers [Score 5] 3 50–200 AI/ML staff; active research partnerships [Score 7] 4 >>200 AI/ML engineers; publish at top venues; competitive with Big Tech [Score 9] Q23. AI ethics and responsible AI. How mature are the firm’s responsible AI practices (bias auditing, explainability, safety testing)? 0 No responsible AI program [Score 1] 1 Informal AI ethics guidelines; no enforcement [Score 3] 2 Published AI ethics policy; bias review before deployment [Score 5] 3 Dedicated responsible AI team; regular audits; explainability tools [Score 7] 4 Industry leading RAI program; external audits; regulatory proactive [Score 9] Q24. AI strategy alignment. How tightly is the AI strategy linked to overall corporate strategy and capital allocation? 0 AI not mentioned in corporate strategy [Score 1] 1 AI referenced in annual report; no dedicated budget [Score 3] 2 AI roadmap aligned to strategic priorities; ring fenced budget [Score 5] 3 AI embedded in capital allocation process; ROI tracking by initiative [Score 7] 4 AI is core strategic pillar; board reviews AI KPIs quarterly [Score 9] Implementation Feasibility Score (IFS) The final question is a composite assessment of five IFS sub factors. Score each sub factor 0–4 using the same anchor logic, then average to produce a single IFS adjustment (see Section 8 for the IFS geometric aggregation formula). Q25. Implementation feasibility composite. Rate each of the following five sub factors (0–4): (a) Organizational Change Capacity (OCC) 0 = No change management capability; 4 = Proven enterprise transformation track record (b) Data Readiness (DR) 0 = Critical data gaps block AI deployment; 4 = All priority data assets production ready (c) Vendor/Technology Risk (VTR) 0 = Single vendor lock in with immature technology; 4 = Multi vendor strategy; proven technology stack (d) Competitive Response Speed (CRS) 0 = No competitive urgency; slow follower; 4 = First mover or fast follower in AI adoption (e) Regulatory Environment (REG) 0 = Highly restrictive; unclear AI regulation; 4 = Supportive regulatory regime; firm well positioned for compliance Each sub factor maps to scores 1/3/5/7/9 using the same 0–4 anchor protocol. The IFS residual is computed as a weighted geometric mean (Equation 36): IFSf=OCC0.40×DR0.60IFS_f=OCC^0.40×DR^0.60, with VTR, CRS, and REG entering the timing adjustment t50,ft_50,f. Dimension score computation. For each dimension d∈DIM,PAC,WAR,DAR,APR,OACd∈\DIM,PAC,WAR,DAR,APR,OAC\, the dimension score sds_d is the arithmetic mean of its four question scores. The aggregate AITG score is then: AITGfraw=16∑d=16sdAITG_f^raw= 16 _d=1^6s_d (51) as defined in Equation 51. See the scoring rubric in Table 3 for intermediate anchor descriptions. Appendix S8 VCB Sensitivity: Exit Multiple Impact To illustrate the sensitivity of AITG-VD to exit multiple assumptions, Table S15 presents AITG-VD under three exit multiple scenarios for a representative healthcare company with AITG =4.5=4.5, IFS =0.75=0.75, and implementation cost $8B. Table S15: AITG-VD Sensitivity to Exit Multiple (Healthcare, Representative Case) Exit Multiple Raw EV (∑VpΣ V_p) Risk Adj. EV Cost VD 8×8× EBITDA $38B $28.5B $8B 2.6× 12×12× EBITDA $57B $42.8B $8B 4.3× 16×16× EBITDA $76B $57.0B $8B 6.1× A 2× change in exit multiple produces a 2.35× change in AITG-VD, confirming exit multiple as the dominant sensitivity parameter in the Sobol analysis. Appendix S9 AFC Benchmark Suite: Proposed Future Extensions The current AFC Capability Index (CtC_t) uses publicly available benchmarks scored on tasks relevant to general cognitive automation. As domain specific AI develops, I recommend extending the suite with: • Clinical/biomedical: FDA SaMD approval rates by indication type; MedPerf benchmark progression. • Legal: Bar exam performance by jurisdiction; contract review accuracy benchmarks. • Financial: FinBen Xie et al. (2024); earnings forecast accuracy vs. sell side consensus. • Agentic (operations): WebArena; WorkArena; ToolBench. As each domain specific benchmark saturates, the agentic task completion horizon (METR analysis) should serve as the primary long run AFC driver. S9.1 AFC Sensitivity: θi×Ct _i× C_t Grid Analysis To characterize AFC behaviour across the plausible parameter space (see Main Paper Definition 2 and Remark 4.1), Table S16 reports AFC multiplier values under a grid of industry sensitivity (θi _i) and capability index (CtC_t) combinations, with C0=1.0C_0=1.0 and αmax=1.35 _ =1.35. Table S16: AFC Multiplier Grid: AFC=min(1+θi(Ct−1.0), 1.35)AFC= (1+ _i(C_t-1.0),\;1.35) CtC_t θi _i 0.90 1.00 1.20 1.50 1.70 1.90 2.10 0.08 0.992 1.000 1.016 1.040 1.056 1.072 1.088 0.11 0.989 1.000 1.022 1.055 1.077 1.099 1.121 0.14 0.986 1.000 1.028 1.070 1.098 1.126 1.154 0.22 0.978 1.000 1.044 1.110 1.154 1.198 1.242 0.28 0.972 1.000 1.056 1.140 1.196 1.252 1.308 0.31 0.969 1.000 1.062 1.155 1.217 1.279 1.341† 0.50 0.950 1.000 1.100 1.250 1.350† 1.350† 1.350† 1.00 0.900 1.000 1.200 1.350† 1.350† 1.350† 1.350† Note. †Capped at αmax=1.35 _ =1.35. Bold column (Ct=1.90C_t=1.90, GPT-5.2, Dec. 2025) is the paper’s calibration. At this CtC_t, the cap is non binding for all 22 calibrated industries (max AFC = 1.279, Healthcare Services, θi=0.31 _i=0.31). Rows span the calibrated range (0.08–0.31) plus hypothetical higher values. For Ct<1.0C_t<1.0 (capability regression), AFC contracts the ceiling. Appendix S10 Excel Companion Model: PE/IC Implementation S10.1 Rationale A recurring commercial objection to quantitative frameworks of this complexity is that they require Python or iterative numerical solvers that cannot be traced in a 4 week diligence sprint by an Excel based deal team. This appendix documents the companion Excel workbook (AITG_Excel_Companion_v1.xlsx) that resolves this objection without degrading the framework’s precision. The core insight is that 12 of 13 AITG components are natively Excel implementable in a single cell (Table S17). The one genuine exception, the inverse mapping t^f=AITG−1(s) t_f=AITG^-1(s), resolved by a pre solved lookup table rather than by degrading the formula to a tiered approximation. A deal team uses =VLOOKUP(ROUND(atgi,1), t_hat_table, 2, FALSE) and gets t^f t_f in a single cell with ±0.05± 0.05-month accuracy, zero Python, and full auditability. S10.2 Workbook Structure The companion workbook contains eight worksheets, each self contained and cross linked by green text cell references: 1. INSTRUCTIONS: Color coding guide, worksheet navigation, and explanation of the lookup table fix. 2. t_hat LOOKUP: Pre solved t^f t_f for AITG ∈[0.1,9.9]∈[0.1,9.9] in 0.1 point steps, with wave zone, Rf0R_f^0, and representative t50,ft_50,f for each entry. Eliminates Newton-Raphson. 3. IASS CALCULATOR: 5 dimension geometric mean IASS, RFF ψ adjustment, AFC multiplier, and IASS∗ output. All Excel. 4. COMPANY SCORER: 6 dimension AITG rubric scorecard. Enter raw scores; get AITG, t^f t_f (via lookup), wave zone, and Φf _f in linked output cells. 5. IFS CALCULATOR: 5 factor IFS composite, endogenous t50,f=t0,1,f+λt_50,f=t_0,1,f+λ calculation, and IFS residual multiplier. 6. VCB MODEL: Full Value Creation Bridge. Enter 7 value pool baselines from diligence financials; receive TV, FCF schedule (endogenous ramp), ΔEV , and Value Density. 7. SENSITIVITY: Pre populated 1 way (δDR _DR vs. ΔEV ) and 2 way (δOCC _OCC × δDR _DR vs. ΔEV ) sensitivity tables. Green/yellow/red heat map coloring. No formula input required. 8. IC DASHBOARD: Single page output for the Investment Committee. Auto populates AI Frontier Matrix position and Value Creation Waterfall summary from upstream sheets. Table S17: AITG Components: Excel Implementation Formulas Component Excel? Formula (single cell) IASS Geometric Mean ✓ =EXP(SUMPRODUCT(weights, LN(scores))) RFF ψ multiplier ✓ =MIN(1,(RFF/5)ˆ1.5) CES Bottleneck (ρ=5ρ=5) ✓ =SUMPRODUCT(alpha,scoresˆ-5)ˆ(-1/5) Φf _f logistic ✓ =1/(1+EXP(-2*LN(rev/S_star))) Value ramp Rf(t)R_f(t) ✓ =1/(1+EXP(-0.18*(t-t50))) ΔRf R_f correction ✓ =ramp(t_hat+60) - ramp(t_hat) t0t_0 IFS adjustment ✓ =t0_base/(OCCˆ0.40*DRˆ0.60) t50,ft_50,f endogenous ✓ =t0_adjusted + 3 AFC multiplier ✓ =1 + theta*(C_t - 1) FCF discrete sum ✓ =SUMPRODUCT(Vr*delta_R*IFS/(1+WACC)ˆy) Terminal Value ✓ =Vr * delta_Rf * M_i * IFS t^f t_f (Wave 1) ✓* =VLOOKUP(ROUND(atgi,1),t_hat_table,2,FALSE) t^f t_f (Waves 2–3) ✓* Same lookup; table pre solved for full [0.1, 9.9] range * Pre solved lookup table eliminates Newton-Raphson. Accuracy: ±0.05± 0.05 months vs. continuous solver. S10.3 Design Principles The workbook follows PE financial modeling conventions throughout: Color coding (industry standard): Blue text = hardcoded input (change for your target); Black text = formula (do not edit); Green text = cross sheet reference (do not edit); Yellow background = key assumption requiring review. No tiered cliff edges: The companion model retains the continuous logistic Φf _f and CES bottleneck rather than replacing them with tiered step functions. Tiered approximations create gameable discontinuities (a 4×4× value capture jump at DIM = 4.0) that a management team will reverse engineer in hours. The continuous formulas fit in a single Excel cell and cannot be gamed. Traceability: Every output cell is traceable to a named input cell. A senior partner can follow the chain from “Data Readiness score: 0.48” through t50,ft_50,f, Rf0R_f^0, ΔRf R_f, and terminal value to ΔEV without leaving the spreadsheet. This is the standard the Reduced Form proposal claimed the full framework could not meet.