Paper deep dive
AI Adoption Across a Multinational Workforce: Sociotechnical Conditions for GenAI Acceptance in Human Resources
Dalia Ali, Maria JosĂŠ RodrĂguez VelĂĄzquez, Manoel Horta Ribeiro, Vera Liao, Orestis Papakyriakopoulos
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%
Last extracted: 6/21/2026, 1:17:42 AM
Summary
This paper investigates the sociotechnical conditions of Generative AI (GenAI) adoption within a multinational tech company's Human Resources (HR) department. By analyzing a transition from a legacy search system (STEVE) to a GenAI-enhanced system (PEOPLE TOOL), the researchers identify that adoption is not a simple migration but a complex process influenced by situational fit, search literacy, and trust calibration. The study highlights that adoption depends on the alignment between system design assumptions (e.g., language, role, tenure) and employee positionalities. Key findings suggest that trust is built through active behaviors like source-checking and comparison, and that organizations must treat organizational knowledge infrastructure as critical AI infrastructure to ensure inclusive and accountable deployment.
Entities (7)
Relation Signals (4)
PEOPLE TOOL â replaces â STEVE
confidence 100% ¡ transitioning from a legacy Human Resources (HR) search system to a GenAI-supported system
PEOPLE TOOL â uses â Generative AI (GenAI)
confidence 100% ¡ PEOPLE TOOL is a pseudonym for the newer GenAI-enhanced HR system.
Generative AI (GenAI) â influences â Adoption
confidence 90% ¡ adoption depended on the fit between the GenAI system's design assumptions and employees' work positionalities
Search Literacy â affects â Adoption
confidence 85% ¡ showing that adoption is influenced by factors such as situational fit, search literacy, and trust calibration.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Generative AI (GenAI) deployment in the workplace is accelerating rapidly. Nevertheless, questions of who adopts, who benefits, and who is left behind and why are still understudied. In this paper, we investigate these dynamics in the context of a multinational tech company transitioning from a legacy Human Resources (HR) search system to a GenAI-supported system, analyzing search log data, survey data (n=25), and ten semi-structured interviews. Our findings show that adoption depended on the fit between the GenAI system's design assumptions and employees' work positionalities (role, spoken language, tenure). Further, we find that employees' trust in GenAI answers was built through source-checking, comparison among systems, and seeking input from colleagues or HR when in doubt. Our contribution is twofold. First, we provide empirical evidence of workplace GenAI adoption during a live organizational transition, showing that adoption is influenced by factors such as situational fit, search literacy, and trust calibration. It is also further shaped by knowledge conditions such as the system's content quality, employee training, and guidance. Second, we translate these findings into design considerations for inclusive deployment and adoption in high-stakes environments such as HR. We argue that organizations should design systems considering the role and context-sensitive benefits they yield to different social groups. They also need to treat the organizational knowledge infrastructure as AI infrastructure to improve the accountability and usability of GenAI systems
Tags
Links
- Source: https://arxiv.org/abs/2606.17887v1
- Canonical: https://arxiv.org/abs/2606.17887v1
Trouble viewing inline? Open PDF directly â
Full Text
88,697 characters extracted from source content.
Expand or collapse full text
AI Adoption Across a Multinational Workforce: Sociotechnical Conditions for GenAI Acceptance in Human Resources Dalia Ali 1 * â , Maria Jos Ě e Rodr Ě Äąguez Vel Ě azquez 1,* , Manoel Horta Ribeiro 3 , Vera Liao 2 , Orestis Papakyriakopoulos 1 1 Technical University of Munich 2 University of Michigan 3 Princeton University Abstract Generative AI (GenAI) deployment in the workplace is accel- erating rapidly. Nevertheless, questions of who adopts, who benefits, and who is left behind and why are still understud- ied. In this paper, we investigate these dynamics in the context of a multinational tech company transitioning from a legacy Human Resources (HR) search system to a GenAI-supported system, analyzing search log data, survey data (n=25), and ten semi-structured interviews. Our findings show that adop- tion depended on the fit between the GenAI systemâs design assumptions and employeesâ work positionalities (role, spo- ken language, tenure). Further, we find that employeesâ trust in GenAI answers was built through source-checking, com- parison among systems, and seeking input from colleagues or HR when in doubt. Our contribution is twofold. First, we pro- vide empirical evidence of workplace GenAI adoption during a live organizational transition, showing that adoption is in- fluenced by factors such as situational fit, search literacy, and trust calibration. It is also further shaped by knowledge con- ditions such as the systemâs content quality, employee train- ing, and guidance. Second, we translate these findings into design considerations for inclusive deployment and adoption in high-stakes environments such as HR. We argue that or- ganizations should design systems considering the role- and context-sensitive benefits they yield to different social groups. They also need to treat the organizational knowledge infras- tructure as AI infrastructure to improve the accountability and usability of GenAI systems. 1 Introduction Artificial Intelligence (AI) is becoming a general-purpose layer of organizational infrastructure, transforming pro- cesses and decision-making (Enholm et al. 2022). Recent studies suggest that AI capabilities are improving rapidly, enabling systems to handle increasingly complex tasks with limited human intervention (Douglas and Verstyuk 2025; Kwa et al. 2025). Improvements in AI capabilities can there- fore affect organizations at multiple levels. At the task level, AI can increase productivity by automating routine cogni- tive work, supporting repetitive analytical tasks, and help- ing workers make faster or better-informed decisions (Bryn- jolfsson, Li, and Raymond 2025; Lee et al. 2022; Ali et al. * Equal contribution. â Corresponding author: dalia.ali@tum.de 2025). At the employee level, these systems may offer real- time assistance, feedback, and communication support, fa- cilitating learning and increasing job satisfaction, especially for less-experienced employees (Brynjolfsson, Li, and Ray- mond 2025; Bhargava, Bester, and Bolton 2021). Last, at the organizational level, AI can make information more widely available across teams and support coordination between de- partments, reducing some of the silos and expertise imbal- ances that often limit organizational learning and collabora- tion (DellâAcqua et al. 2025). Yet, adoption lags behind technological progress. Recent industry reports show a gap between AI access, exploration, and exploitation. While workforce access to AI technolo- gies is growing, few companies manage to move from the pilot to the production stage; 65% of senior leaders strug- gle to tie productivity gains directly to AI adoption (Insti- tute 2026; MIT Project NANDA 2025; Ernst & Young LLP 2025). Previous work shows that adoption is shaped by dif- ferent factors such as perceived usefulness and ease of use, social influence, data, technology infrastructure, organiza- tional culture, ethics, and regulation (Enholm et al. 2022; Venkatesh 2022; Davis 1989). However, these explanations do not fully capture the complex, interconnected, and dy- namic environments in which technology is implemented; further, they offer limited insight into the complex situated user experiences and their fit with available technologies (Lee, Ramasamy, and Subbarao 2025; Shachak, Kuziemsky, and Petersen 2019; Mogaji et al. 2024). To bridge this gap, we need to understand not only how employees use AI but also what drives different employ- eesâ decisions to adopt, reject, or selectively use AI dur- ing organizational transitions. Without such understanding, organizations risk failing to realize the value of their in- vestments, missing the benefits of increased efficiency and performance, and reproducing uneven patterns of adoption among employees (Enholm et al. 2022; Mlekus et al. 2020). From the employee perspective, poor GenAI design may mean that a system does not support the task at hand, pro- duces outputs that are difficult to evaluate or trust, or ex- cludes some employees due to a lack of language fit, acces- sibility, and work-context needs (Gu et al. 2024; Tankele- vitch et al. 2024; Liao and Sundar 2022; Weidinger et al. 2021; Budhwar et al. 2023; Sarkar et al. 2022). Since barri- arXiv:2606.17887v1 [cs.HC] 16 Jun 2026 ers to GenAI adoption already exacerbate existing inequali- ties across workers, inclusive design is needed to ensure that efficiency benefits are not concentrated among some em- ployees while others are left behind (Humlum and Vester- gaard 2025). To examine AI adoption from a user-centered perspective and understand why some employees benefit from GenAI systems while others are left behind, we study a real or- ganizational transition within a multinational tech company from an existing knowledge system for HR to one that in- corporates GenAI capabilities into its retrieval process. The new GenAI-based search system relies on HR documents from the internal organizational repository to enable em- ployees to access HR-related information faster and more accurately. Existing organizational leadershipâs willingness to implement an innovative AI solution and the availability of necessary technological and financial resources allow us to investigate actual sociotechnical factors that facilitate or impede adoption, rather than boundary conditions of tech- nological availability. Specifically, we raise the following re- search question: Why do users adopt or resist a GenAI-enhanced HR knowledge system during an organizational transition? By combining one year of the employeesâ search log data from an existing knowledge search system, two weeks of data from the new GenAI-based search system, an ex- ploratory survey, and ten semi-structured interviews with project team members and end-users, we show that AI adop- tion was not a simple migration from an old to a new system. Instead, it was characterized by selective use across parallel systems, situated fit across user roles, social contexts, and the technologies, userâs existing GenAI search literacy, how users calibrated their trust in the new system, and organiza- tional conditions such as content quality and user guidance. Our paper makes two main contributions: (1) we iden- tify sociotechnical mechanisms through which workplace GenAI produces unequal benefits, showing that when sys- tems are designed around assumptions of office-based ac- cess, digital confidence, organizational familiarity, and lan- guage fit, workers outside these assumptions are systemati- cally less well served despite formal access; (2) we translate the findings into design considerations for inclusive and ac- countable workplace GenAI systems. 2 Background 2.1 AI Adoption in Organizations As AI capabilities continue to advance rapidly, adoption re- mains uneven across both organizations and workers (Hum- lum and Vestergaard 2025; Institute 2026; MIT Project NANDA 2025). Explaining this variation requires examin- ing adoption at multiple levels. At the organizational level, the Theory of Technology, Organization, and Environment (TOE) identifies technological, organizational, and environ- mental factors that shape whether organizations embrace new technologies (Tornatzky and Fleischer 1990; Baker 2011). However, TOE does not capture individual-level adoption decisions within organizations (Awa, Ojiabo, and Orokor 2017; Oliveira and Martins 2011). At the individual level, the Technology Acceptance Model (TAM) and the Unified Theory of Acceptance and Use of Technology (UTAUT) explain adoption through per- ceived usefulness, ease of use, performance expectancy, ef- fort expectancy, social influence, and facilitating conditions (Davis 1989; Venkatesh et al. 2003). These models have strong predictive power for traditional software, but were not designed for AI systems, which are probabilistic, opaque, and context-dependent in ways that fundamentally alter how users interact with technology (Venkatesh 2022). Extensions of these models to AI domain have taken into account issues related to trust, transparency, and perceived risk, recogniz- ing that users must cope with probabilistic outputs, limited explainability, and unpredictable behavior (Venkatesh 2022; Enholm et al. 2022; Wolfe et al. 2025). Nonetheless, the effi- cient use of GenAI technologies introduces additional inter- action requirements that cannot be entirely accommodated by the previous trust models or adoption models (Tankele- vitch et al. 2024; Zamfirescu-Pereira et al. 2023; Sarkar et al. 2022; Weisz et al. 2023). Research on AI in the workplace consistently shows that these systems do not impact all workers equally. Several studies document how GenAI reshapes knowledge work (Woodruff et al. 2024), and how algorithmic systems create failure loops that systematically undermine certain types of work (Kawakami et al. 2024). In HR research, most scrutiny has focused on algorithmic hiring (Raghavan et al. 2020; Sanchez-Monedero, Dencik, and Edwards 2020), and emo- tion AI in interviews has been shown to result in lower per- ceptions of justice among gender minorities (Ingber and An- dalibi 2025). Despite evidence that AI affects workers unequally, ex- isting adoption models still provide limited insight into how adoption unfolds through everyday interactions among users, technology, and organizational context (Shachak, Kuziemsky, and Petersen 2019; Mogaji et al. 2024). They do not adequately explain how users learn, adapt, de- velop workarounds, or incorporate AI systems into practice (Shachak, Kuziemsky, and Petersen 2019); nor how adop- tion is shaped by usersâ expertise, job role, task suitability, and organizational conditions (Bankins et al. 2024). 2.2 User-centered considerations of AI systems Addressing this gap requires moving beyond current adop- tion models and examining how users engage with AI sys- tems in practice. User-centered AI adoption depends on more than access. How well employees benefit from the AI system depends on their ability to interpret its outputs, build trust, learn a new interaction logic, and rely on the orga- nizational knowledge infrastructure underlying the system (Tankelevitch et al. 2024; Liao and Sundar 2022; Alavi, Lei- dner, and Mousavi 2024). Although these factors are largely studied when people use systems, prior research does not take them into consideration when studying adoption dy- namics. Rather, the focus lies more on defining isolated social and technical factors that shape adoption(Venkatesh 2022; Wolfe et al. 2025). Research on Human-AI interactions shows that the level of trust in AI systems is not a function of system accuracy; is also influenced by how users interact with the system (Liao and Sundar 2022). Liao and Sundar (2022) differentiate be- tween trustworthy AI, which is related to system accuracy and fairness, and human trust, which requires perception and communication. They propose the MATCH model, in which trust arises from the interplay among model features, sys- tem affordances, trustworthiness cues, and usersâ cognitive processing. (Bansal et al. 2021) also show that explanations, such as supporting information features that promote trust in AI, do not automatically produce appropriate reliance; they can increase usersâ acceptance of AI recommendations even when those recommendations are incorrect. This means that reliance on the AI output needs to be studied empirically, rather than assumed based on the systemâs accuracy. A second factor is the interaction competencies GenAI demands of its users. Unlike traditional information retrieval systems, GenAI requires users to prompt, evaluate outputs, reformulate queries, and decide whether answers are reli- able enough to act on (Tankelevitch et al. 2024; Zamfirescu- Pereira et al. 2023). The development of such competence is not straightforward, as employees must develop proficiency in prompting, output evaluation, and knowing when GenAI is appropriate (Xia et al. 2026). This is made more challeng- ing because workers use GenAI within existing tasks rather than as a standalone activity (Xia et al. 2026). Another factor to consider is the organizational knowl- edge that GenAI systems draw from. Information systems are only as effective as the knowledge supporting them (Alavi and Leidner 2001); and GenAI systems are no ex- ception, since they retrieve and synthesize content from underlying content such as articles, metadata, and local- ized content instead of generating answers independently (Alavi, Leidner, and Mousavi 2024). Recent work argues that GenAI may change how organizations manage knowl- edge by making retrieval and recombination more conversa- tional and automated (Alavi, Leidner, and Mousavi 2024). Collectively, this literature identifies user-centered fac- tors that shape how people use and evaluate GenAI sys- tems. Given this, we investigate how the above and addi- tional sociotechnical factors interact to shape adoption dur- ing a live HR organizational transition across a structurally diverse workforce, in which a GenAI system is introduced alongside a legacy system. 3 Methodology To understand how employees adopt or resist GenAI in orga- nizational knowledge work, we conducted a mixed-method single-case study at a multinational technology organiza- tion with more than 300,000 employees worldwide. This case centered on the organizationâs introduction of PEO- PLE TOOL 1 , an HR platform with an embedded GenAI On Search feature, alongside the continued availability of its legacy HR search system, STEVE 2 . PEOPLE TOOL operated as a retrieval-augmented system; employee queries returned either a GenAI-generated answer alongside source links or 1 PEOPLE TOOL is a pseudonym for the newer GenAI-enhanced HR system. 2 STEVE is a pseudonym for the legacy HR knowledge system. search result recommendations only, depending on whether the system had sufficient content to generate a response. We conducted a three-phase sequential design. First, we analyzed STEVE and PEOPLE TOOL search logs to identify behavioral patterns across query formulation, result avail- ability, click-through behavior, and GenAI answer exposure. Based on these patterns, we conducted an exploratory sur- vey with 25 employees across six countries to assess usage, usefulness, usability, trust, system preference, and employee reactions when the answer generated by GenAI was insuffi- cient. Afterward, we conducted semi-structured interviews with 10 employees from four countries to explore how em- ployees perceive the two systems and how personal, orga- nizational, and technological factors affect adoption paths. Since each method served a different purpose, the three data sources cover different country contexts and are treated as complementary rather than directly comparable country- level samples. This study received ethical approval from our institution, and all data were handled in accordance with GDPR requirements. 3.1 Search Log Analysis Search logs are system-level logs that capture information about employeesâ search activities. Each search log recorded details of queries entered by employees, the systems in which queries were searched, the search timing, country context, and system feedback, such as whether the system returned results or displayed a GenAI response. Search log analysis served as the pre-study phase of our sequential mixed-methods design. Its purpose was exploratory: to sur- face behavioral patterns in how employees queried each sys- tem, with findings directly shaping the design of the survey instrument and the interview protocol. We obtained search logs from both systems covering six countries: Egypt, the Netherlands, the Philippines, Pak- istan, Qatar, and Saudi Arabia. STEVE logs spanned ap- proximately one year (September 2024âAugust 2025). PEO- PLE TOOL logs covered a two-week window (July 24â August 7, 2025) due to a hard constraint imposed by the sys- tem providerâs data retention policy. For cross-system vol- ume comparisons, we restricted both datasets to the common observation window (July 24âAugust 2, 2025) to avoid con- flating differences in system use with differences in obser- vation period; the longer STEVE baseline was retained sep- arately to characterize historical query patterns and queries that returned âno resultâ prior to the introduction of PEOPLE TOOL (Appendix A). Non-English queries were translated into English by a na- tive speaker to allow topic-level comparisons across coun- tries. The analysis focused on query volume and distribution across the two systems, query phrasing and length, result availability, GenAI response exposure rates, and topic over- lap between the two systems. The logs record what employ- ees searched and the results obtained from each system; they do not provide information about why employees chose one system over the other, nor about how they perceived those results; those questions were addressed in the survey and in- terviews. germany Steve / Legacy HR search People Tool / GenAI- enhanced search Steve + People Tool search logs 2. Survey Details 1. Search Log Analysis 3.Interviews Employee EgyptNetherlands Philippines Qatar Pakistan Saudi Arabia Focus: query length, phrasing, results, clicks, GenAI exposure Germany USA India Austria Switzerland France n = 25 employees Measures: ease of use, helpfulness, trust, preference, source checking n = 10 employees GermanyUSASpainNorway Languages: English or Spanish Topics: system comparison, trust, adoption dynamics Analysis: reflexive thematic analysis Figure 1: Sequential mixed-methods study design. Each phase informed the next, with search log patterns shaping the survey and survey responses shaping the interview protocol. 3.2 Survey Building on the initial analysis of the search log, we con- ducted an exploratory survey of employees who could use the PEOPLE TOOL and its built-in GenAI search functional- ity. Participants were recruited from six countries: Germany, the US, India, Austria, Switzerland, and France. This survey was designed to accomplish two tasks: collecting descriptive data on employeesâ experiences with STEVE and PEOPLE TOOL, and refining the interview guide for the upcoming in- terview phase. The survey collected contextual information, including country of work, job level, and tenure. Employees were re- quired to report on their usage frequency for each system, their top HR topics searched for, and their ratings on the sys- temsâ ease of use, helpfulness, and trustworthiness. Specific attention was paid to the use of the GenAI search within the PEOPLE TOOL, where respondents answered on how often they received an answer by the GenAI function, how helpful the answers were, how often they clicked on sources pre- sented after answering, and what action they would take in case the answer was insufficient. The respondents were free to report any problems they faced, their preferences between the two systems, topics that did not yield useful results, and areas for improvement. Finally, a call was made to those willing to participate in the interviews. The survey was answered by 25 employees. For the closed-ended questions, we conducted descriptive statistics using counts and percentages. For the open-ended questions, we analyzed responses to identify recurring concerns that might reappear in interviews. The survey results cannot be considered statistically representative due to their limited sample size and unequal representation. Instead, they serve as descriptive evidence for the interviews and help connect behavior identified in the logs with the process discussed in the interviews (Appendix B, Table 1, Limitations Section 6). Table 1: Key descriptive survey results used to contextualize the findings. MeasurePEOPLE TOOLSTEVE Daily or weekly use19/25 (76%)14/25 (56%) Preferred system14/25 (56%)3/25 (12%) Very/extremely helpful16/25 (64%)8/25 (32%) Easy/very easy to use16/25 (64%)16/25 (64%) Mean helpfulness3.763.12 Mean trust3.923.72 GenAI-specific items Never/rarely clicked sources in GenAI answers13/25 (52%) Would rephrase when GenAI answer was poor19/25 (76%) Note. Survey data are descriptive, not statistically representative. Preference responses were mixed or unclear for 7/25 (28%). Help- fulness and trust means are on 5-point scales (see Appendix B). 3.3 Interview Study Semi-structured interviews were the primary source of in- terpretive data in this research. While search logs and sur- veys provided information on behavior and contextual de- scriptions, interviews helped examine how people inter- preted STEVE, PEOPLE TOOL, and the GenAI search feature in practice. Interviewees were selected among those who agreed to participate in the interviews in addition to the sur- vey, as well as from employees involved in everyday use of PEOPLE TOOL. We conducted ten semi-structured interviews using an on- line videoconferencing tool. Each interview was recorded, lasted between 20 and 35 minutes, and was conducted in En- glish or Spanish, at the participantâs preference. At the be- ginning of each interview, participants were reminded of the purpose of the study, informed that participation was volun- tary, and asked to consent to the recording for transcription and analysis. The interview guide was organized around six key themes, which involved participantsâ background and famil- iarity with digital technologies; participant experience and comparison of STEVE and PEOPLE TOOL; perceived use- fulness and limitations of GenAI on Search; trust, source- checking, and information behavior; adoption dynamics and organizational context; and reflections on future improve- ment. Interview participants represented two broad perspec- tives. The first group comprised employees involved in strat- egy, rollout, knowledge management, system ownership, testing, or support for PEOPLE TOOL; however, these par- ticipants also used the system in their own work, and could therefore speak both as practitioners with implementation knowledge and as users seeking HR-related information. The second group comprised employees whose engagement with the systems was primarily as users seeking HR-related information. We use this distinction to contextualize partici- pantsâ accounts, but we do not conduct a systematic compar- ative analysis between these groups. Participants varied in tenure, ranging from less than 1 year to more than 15 years, and were distributed across four countries. Table 2 provides an overview of the interview participants. We generated the interview transcriptions automatically and manually checked them for accuracy. We anonymized them by removing all personal identifiers and assigned new participant IDs ranging from P1 to P10. Minor corrections were made to enhance readability without altering the con- tent bearing the message. To analyze the interview data, we used reflexive thematic analysis following Braun and Clarke (Braun and Clarke 2006). We first read the tran- scripts repeatedly to familiarize ourselves with the data, then generated descriptive codes capturing participantsâ accounts of system comparison, perceived usefulness, ease of use, trust, source checking, fallback practices, AI-search literacy, training and communication, organizational support, content quality, work context, language, and attitudes or fears to- wards AI. These initial codes were then iteratively grouped into more focused sub-themes. This analysis was conducted over multiple rounds of discussions among four researchers. Instead of consider- ing coding as a single classification process, the discus- sions helped us compare understandings, merge redundant codes, push against poorly formed themes, and link sub- codes to broader analytical themes that relate to the research question. The outcome of this discussion led us to the fi- nal themes presented in the Findings section: selective use across parallel systems, situated fit across workers and con- texts, learning a new GenAI search logic, reliance on check- ing and fallback, and knowledge conditions. 4 Findings By analyzing interviews, survey responses, and the search logs of both systems, we observed that adoption was not a linear process of migration from legacy search to GenAI- enhanced search. As both systems, PEOPLE TOOL and STEVE, were available, employees could switch between them, compare their outcomes, verify the responses, and return to alternative routes when necessary. Adoption was Table 2: Overview of interview participants (n = 10). IDGeneralized role contextTenure (yrs) Country Project team participants P1HR IT / digital transformation; application ownership 1â5Germany P2Transition project member5â15Germany P3HR IT intern; content migration and AI testing 1â5Germany P4Product ownership for GenAI and knowledge management 15+Germany P7HR support leadership5â15Spain P8Regional HR IT coordination5â15USA End-user participants P5Digitalization-related role in adjacent ServiceNow environment 5â15USA P6Compensation and HR tooling role 5â15USA P9HR systems project management and cloud migration 1â5Spain P10 Early-career P&O systems professional <1Norway Note. Role descriptions are intentionally generalized to preserve participant anonymity. shaped by the interaction between employees, the system, and the organization. In this section, we first show that adoption involved se- lective use across parallel systems rather than a complete migration to the new system. Then, we examine why this use varied across employees and contexts. Next, we ar- gue that effective use requires employees to learn a new GenAI search logic. This, in turn, raises questions of trust, as employees had to determine when an answer was reliable enough to act upon. Finally, we show that these everyday adoption practices were shaped by organizational knowl- edge conditions. 4.1 Selective Use Across Parallel Systems Since STEVE continued to function after the transition, em- ployees were not required to rely on a single route for HR information at all times. Instead, they chose among the avail- able routes depending on the situation. For instance, employ- ees would use PEOPLE TOOL if it provided speed, integra- tion, and clearer direction. They would use STEVE if the search process were familiar, the results were predictable and article-based, or if it were a route they already knew how to navigate. Source articles provided precise informa- tion when needed, and employees reached out to colleagues, managers, or the HR department for personal or complex an- swers. Thus, adoption was not a single decision point, but an ongoing practice. Survey responses reflected this parallel-use pattern. In- deed, 76% of the respondents (19/ 25) used PEOPLE TOOL either daily or weekly. However, STEVE was not discon- tinued as 56% of respondents (14/25) used STEVE daily or 1. Selective Use Across Parallel Systems What adoption looked like? Easy, quick, clear Employees chose among People Tool, Steve, source articles, and colleagues based on which route best served their HR needs. Openness to change Optional use Comparison to external AI 2. Situated Fit Across Workers and Contexts Why adoption varied? 3.Effective Use Required Learning a New GenAI Search Logic How users decided whether to rely on it? 4.Trust Through Checking, Comparison, and Fallback Fit varied by work type, device access, language, tenure, and organizational familiarity; formal access did not equal practical access. Work type Experience level Age / life stage Need Cultural and language factors What users had to do? Users had to phrase queries, reformulate failed searches, interpret summaries, and decide when to move from GenAI answers to sources. AI search skill Tool difficulty Individual learning Gradual familiarization Reliance was built through source- checking, comparison with Steve, and fallback to colleagues or HR when uncertain. Consistency First impression Fears and attitudes Content quality Guidance and training 5. Knowledge Conditions for Reliable GenAI Use GenAI reliance depended on content quality (structure, tagging, localization) and guidance (how to search, verify, and escalate) What made answers reliable or fragile? Knowledge conditions underpin adoption Figure 2: Adoption unfolded through selective use across parallel systems, varied by situational fit, required learning a new GenAI search logic, and depended on trust-building through checking, comparison, and fallback. Knowledge conditions under- pinned the whole process. weekly. Further, regarding preferences, 56% of respondents preferred PEOPLE TOOL, whereas 12% preferred STEVE. It appears that although PEOPLE TOOL was widely used, the legacy system was still used as part of the routine HR information-seeking workflow. Interviews also elucidate the underlying logic by showing that switching to a new system was worthwhile. Employees were not asking only whether PEOPLE TOOL was new or AI-powered, they were interested in its advantages relative to familiar information-search methods. One participant ex- plained his concerns as follows: âOh man, now thereâs another HR system (PEOPLE TOOL) I have to use. You know, Iâm comfortable with this one. Why do we have to change again?â (P5) However, this response cannot be regarded as a sign of general resistance to AI. Learning a new route takes time and effort that employees simply do not have to devote to a task of lower priority than their main work. Since a familiar route could be accessed with a single click, the new route needed to provide sufficient value. If PEOPLE TOOL offered clear answers and useful information, employees could find a valid reason to use it; but when it added more navigation and cognitive burden without a clear benefit, familiar routes remained the easier choice. Therefore, the simultaneous availability of PEOPLE TOOL and STEVE meant that adoption was not a single migration decision from the legacy to the new GenAI system, but an ongoing practice of choosing among routes. This raises a question: what made PEOPLE TOOL the preferred route for some employees but not others? 4.2 Situated Fit Across Workers and Contexts The reasons behind selective adoption were rooted in situ- ational fit. We define situational fit as the degree to which the toolâs offerings match usersâ actual work conditions, needs, and contexts. PEOPLE TOOL was designed for lap- top and computer access, assumed organizational familiar- ity, and prioritized some languages over others, resulting in a fit that varied considerably across the workforce. A recurring concern throughout the interviews was that PEOPLE TOOL was better suited to office-based employees who regularly had access to laptops, were accustomed to the companyâs internal knowledge system, and could frame their HR inquiries in organizational terminology. For shift-based, factory workers, and shop-floor employees, this was not the case. Participants suggested that these worker groups relied on their managers, team assistants, local HR staff, or infor- mal channels. For instance, one participant noted that blue- collar workers may not be familiar with organizational HR terminology, making it difficult to even formulate a ques- tion âwhen youâre a blue-collar, youâre not as exposed to that side of things and so even knowing what to askâ (P6). Another participant suggested that PEOPLE TOOL was ânot idealâ for factory workers (P1). These responses suggest that although the system was technically available to all em- ployees, its design for office-based, digitally familiar, and or- ganizationally embedded users could lead to exclusion, with employees with diverse work conditions at a disadvantage in using the tool. Thus, making PEOPLE TOOL formally available to em- ployees did not mean that all employees could use it equally in practice. Some employees lacked regular access to a suit- able device âworkers who go on site to work...they some- times donât even have a laptopâ (P5); others did not know how to phrase HR questions in organizational terminology (P6), and others continued to rely on managers, local HR staff, or colleagues as their main route to information. These employees could disappear from system usage logs or ag- gregate adoption figures, even if the tool was technically available to them. As a result, aggregate adoption figures therefore overstate inclusion, as a high query volume from office-based employees masks the absence of shift-based, blue-collar, and multilingual workers who never meaning- fully engaged with the system. Tenure and life stage played an important role in shap- ing situational fit as well. Because new recruits were still learning the organizationâs internal processes and the loca- tion of HR documentation, they could gain the most from PEOPLE TOOL. On the other hand, tenured employees used the PEOPLE TOOL less, not because they resisted it, but be- cause their information uncertainty had declined, and they already knew the companyâs policies, the personnel respon- sible for addressing HR concerns, and the routine mecha- nisms to resolve issues. As one participant explained: âThe older you are, the less changes you have in life. I would say the kids are gone, house is built, car is there. What shall you ask HR? Maybe the younger people have more interactions with HR because they are still building their lives.â (P4) The GenAI toolâs situated fit with employees was also ev- ident from the logs. First, search behavior was distinct by country. For instance, in the Netherlands, users searched for policy and process-related topics, in Saudi Arabia and Qatar, searches focused on documents and certificates, in Egypt, on benefits and insurance, and in Pakistan, on travel claims. System-level indicators were distinct as well, for instance, in Pakistan, click-through was the lowest across all countries at 18.5%, suggesting employees were less likely to move from search results to further interaction with the system. These patterns show that a global HR search system encountered diverse local task environments, with employees searching for HR artifacts using distinct terminologies and addressing diverse requirements. Language was another factor that impacted adoption. If a system failed to accommodate the userâs language or did not reflect local HR terminologies, its effectiveness suffered despite its technical capabilities. In multinational rollouts, language support is typically added incrementally, with non- English and non-European languages often deprioritized or absent altogether; one participant emphasized that âif you donât have the languages, then people sometimes donât use itâ (P4). Together, the interview and log data suggest that a one-size-fits-all design for GenAI is ineffective in multina- tional HR knowledge work. The same system provided con- venience to some users while creating friction or exclusion for others. Thus, situational fit shaped the selective use of PEOPLE TOOL. When the toolâs offerings aligned with usersâ device access, language, work environment, life stage, and HR in- formation needs, it gained greater traction. Otherwise, em- ployees relied on legacy systems or local sources that they were already familiar with. This raises the next question, even when the tool aligned with a userâs context, what did it take for employees to effectively use GenAI HR search? 4.3 Effective Use Required Learning a New GenAI Search Logic Access to PEOPLE TOOL did not automatically translate into effective use. Even when system fit a userâs context, employ- ees still had to learn a new way of searching that differed fundamentally from STEVE. This involved phrasing ques- tions, adding context, rephrasing failed queries, interpreting summaries, and knowing when to move from a GenAI re- sponse to source articles or human support. Thus, adoption was based not only on what the system offered but also on whether employees could translate their HR needs into ques- tions the system could answer; and as the survey data show, these two dimensions usefulness and ease did not always align. This distinction helps explain why usefulness and ease did not necessarily occur in parallel. The survey showed that PEOPLE TOOL was rated as âveryâ or âextremelyâ useful by 64% of respondents (16/25), compared with 32% (8/25) for STEVE. However, both systems were rated as âeasyâ or âvery easyâ by the same proportion of respondents, 64% (16/25). In other words, employees recognized the added value of PEOPLE TOOL GenAI search, but realized that value required learning a new interaction logic, including how to ask, interpret, and adjust their searches. The survey further illustrates this interaction effort. When GenAI failed to provide a satisfactory answer, 76% of the re- spondents (19/25) reported they would rephrase their ques- tion. This demonstrates that users viewed poor answers as something that could be improved by asking differently, rather than a system failure. By doing so, employees partici- pated in creating the response, making the answer dependent not only on the system itself, but also on employeesâ knowl- edge of what to ask and how. The remaining 24% did not rephrase, suggesting that some employees disengaged en- tirely when the system failed to provide a useful answer. The logs show that this new search logic had not yet fully replaced older keyword-search habits. We measured query length in terms of tokens, defined as units similar to words in a search request. The average query length in- creased only slightly, from 1.87 tokens in STEVE to 2.11 tokens in PEOPLE TOOL. Similarly, the share of short one- to two-token queries stayed practically unchanged, 83.4% in STEVE and 83.3% in PEOPLE TOOL. This suggests that em- ployees continued to use traditional keyword-based search practices even in the GenAI-enabled environment, although short keyword queries may also have remained effective given the systemâs retrieval-augmented architecture. The in- terviews explain this, as a participant described the change as a shift in search logic: âOnce they [the user] have understood that the logic changes, that the way of searching and finding is dif- ferent, I think there is greater acceptanceâ (P7) Effective use, therefore, required more than familiarity with the system; employees had to develop a new form of search competence, including learning not only where to search, but how to search with AI, how to recover when the answer was poor, and how to judge whether an answer was enough for their HR question. The survey evidence supports this; while 64% of respondents rated PEOPLE TOOL as use- ful, the same proportion rated both systems as equally easy to use, suggesting that perceived value and interaction com- petence are distinct dimensions of adoption. This compe- tence was not about learning where to search, but about de- veloping an ongoing practice of prompting, evaluating, and deciding. This leads to the next issue: even when employ- ees learned how to ask, how employees decided whether a GenAI answer was reliable enough to use? 4.4 Users Built Trust by Checking, Comparing, and Fallback Receiving GenAI answer was not sufficient for adoption. Employees also had to decide whether that answer was ac- curate enough to act on. In HR, acting on incorrect or partial information can have serious implications for salary, leave, and personal circumstances. Therefore, adoption depended on employeesâ ability to move from obtaining an answer from the GenAI tool to deciding whether to rely on it. This reliance was not based on blind acceptance but was cali- brated through checks, comparisons, and fallback. The survey suggests a gap between usefulness and trust. Mean helpfulness was higher for PEOPLE TOOL than STEVE: 3.76 compared with 3.12. By contrast, the differ- ence in mean trust was more modest, 3.92 compared with 3.72. This suggests that employees may perceive greater value in PEOPLE TOOL while remaining cautious about whether its answers are complete, current, and applicable to their situation. This caution also appeared in the interviews as one participant stated: âI will not sell my house because of an AI an- swer. Thereâs always a double net. Thereâs always checks and balances, talking to managers, talking to colleagues, talking to my wife, consulting someone, friends who made this already. So itâs not just the an- swer in [People Tool] leading to the decision. Itâs al- ways a mix of information and then you come to your own conclusion.â (P4) Source-checking was one way of managing this uncer- tainty, although it was not used consistently. In the sur- vey, 52% of respondents (13/25) reported that they never or rarely clicked the sources shown in GenAI-generated an- swers, while 48% (12/25) clicked them sometimes or often. This difference should not be viewed simply as a split be- tween trusting and distrusting users; sometimes, employees did not click sources because the GenAI answer was suf- ficient for their needs and required no further verification. Others coped with this uncertainty in different ways, such as inspecting sources, relying on plausibility, restructuring their question, comparing the answer with what is known from other systems, and seeking help from others. At the same time, trust was not always actively calibrated. Some employees appeared to lend credibility to PEOPLE TOOL because it was an internal company system. As one participant explained: âI think people innately trust things internally in [the Company] that itâs going to give them the correct answer, and I think people are trusting AI too much.â (P6). Yet, the organization placed responsibility for interpreting outputs back onto employees as one participant pointed out: âWe always tell the users, âYou can use AI, but the risk is on you.â How can I force them to make use if they know that they have the responsibility?â (P4). Employees sometimes trust PEOPLE TOOL GenAI responses not be- cause they have reviewed their source or assessed their com- pleteness, but because they were delivered through an ap- proved organizational platform; this, in turn, creates the risk that fluent but incorrect answers would be accepted without sufficient verification. The interviews also identified the continued need for a fallback. A GenAI summary could be useful for quick ori- entation, particularly when the question was general or low- risk. But when questions are specific, complex, or high-risk, employees find greater satisfaction in seeking reassurance from source documents, STEVE, peers, supervisors, or HR. Hence, trust was distributed across the answer itself and the wider set of information routes surrounding it. Trust was also shaped by prior experiences and first im- pressions. Some employees approached GenAI with curios- ity, while others began with skepticism, privacy concerns, or worries about unreliable HR guidance. These starting points did not determine adoption on their own, but they influenced how employees interpreted their initial interactions with the system. Repeated use could revise initial expectations, but only when the system behaved consistently enough to sup- port confidence. When GenAI answers were wrong or irrel- evant, employees would disengage quickly. When answers appeared plausible but remained uncertain, employees had to decide whether to verify, rephrase, or ask someone else. In general, reliance on PEOPLE TOOLâS GenAI search was built through repeated use rather than assumed a pri- ori. Employees delegated some information work to PEO- PLE TOOL, but also checked, compared, or returned to other routes when uncertain. This raises another question: what made GenAIâs answers reliable or unreliable in the first place? 4.5 Knowledge as a Condition for Reliable GenAI Use The previous analysis showed that employees needed mech- anisms to validate, compare, and trace answers back to dif- ferent sources when GenAI-produced answers are ambigu- ous. Thus, there was an even deeper issue about the trust- worthiness of PEOPLE TOOL that was not dependent on the GenAI feature alone. Employees accessed HR information through GenAI, which in turn depended on the quality of the underlying HR articles, metadata, tagging, taxonomy, lo- calization, training, communication, and escalation routes. When these were well-structured, this allowed the genera- tion of useful guidance via the GenAI process; however, un- coordinated input data, outdated articles, poor tagging, and vague explanations made it difficult for the GenAI tool to compensate for these issues. This is supported by survey data, which shows that while the PEOPLE TOOL proved to be more effective than STEVE (mean helpfulness of the tool has risen from 3.12 to 3.76), trust increased only moderately (up from 3.72 to 3.92). In other words, PEOPLE TOOL made HR information easier to find, but employees remained uncertain about whether the answers were complete, current, and applicable to their spe- cific situation. Employee responses focused on the quality of information architecture more than on the AI technology itself, as one employee noted: âIn PEOPLE TOOL the information is not that user- friendly and as each item might be under a different article, itâs hard to find what you want.â (Survey re- spondent) Another employee linked poor search results to weak tag- ging, âNot every document or article has the tags that I would search for and therefore they are not shown in the search resultsâ (Survey respondent). These responses illus- trate why GenAI reliability cannot be separated from content quality. Employees sometimes attributed poor answers to the AI when the underlying problem was poorly organized, out- dated, or localized HR content. Training and guidance were also part of these organi- zational conditions, not just rollout activities. Employees needed to know more than that PEOPLE TOOL existed; they needed to understand what had changed, how to search, when GenAI answers would appear, when to check sources, and when to escalate to human support. Interviewees sug- gested that support was most useful when provided close to the moment of use, through examples, short guidance, Q&A sessions, office hours, prompts, and clear contact routes. One-off pre-launch communication was less effective be- cause it could be forgotten before employees actually needed the tool. Therefore, content quality, guidance, and training were not background conditions for GenAI adoption; they deter- mined whether the answers employees received were reli- able enough to act on. PEOPLE TOOL could make HR search faster and more convenient, but only when the underlying content was structured, maintained, localized, clearly com- municated, and employees were supported at the point of use. 5 Discussion 5.1 Who Benefits from Workplace GenAI? Recent research on AI fairness and sociotechnical design argues that potential harms arising from AI cannot be un- derstood only by examining model behavior; the social and organizational context in which these systems are deployed must also be considered (Selbst et al. 2019; Costanza-Chock 2020). Similarly, work in workplace HCI has shown how al- gorithms affect workersâ experiences within organizations and the management settings and infrastructure of work- places (Cheon and Erickson 2025). Research on digital in- clusion has also highlighted the mistake of considering tech- nological access alone is sufficient for understanding in- clusion and emphasized the roles of skills, social context, conditions of use, and the ability to translate technolo- gies into concrete gains (Warschauer 2003; van Dijk 2005; Helsper 2021). Our findings extend this argument to work- place GenAI deployment: making a GenAI system available across an organization does not mean that all employees are equally positioned to benefit from it. In our study, formal availability of the GenAI HR sys- tem did not translate equally into useful HR support. As highlighted in Section 4.2, employeesâ ability to take ad- vantage of the system depended on its alignment with work processes, linguistic preferences, HR-related information re- quirements, and access to support channels. This raises an equity issue that remains challenging to operationalize. When the system is designed for office-based, digitally con- fident, and English-speaking users, how will organizations recognize which workers are at a disadvantage? This ques- tion is difficult to answer through aggregate usage metrics alone, because those who face the greatest friction may be characterized by their absence, inactivity, and infrequent use. A system may therefore appear broadly adopted while still providing more reliable and actionable support to some groups and exclude the others (Selbst et al. 2019; Humlum and Vestergaard 2025). The relevant question is not only who can access the system, but who can obtain useful, trust- worthy, and actionable support from it in everyday work. Design recommendations. Organizations should evalu- ate their GenAI deployment in the workplace by considering role-sensitive and context-sensitive benefit measures. This includes examining which groups are missing from usage logs, which users receive search results but not GenAI an- swers, which queries fail or require repeated reformulation, who continues to rely on local intermediaries, and which groups do not receive useful or actionable support. Then, the less visible employee segment should be included continu- ously in testing, feedback, and redesign, rather than being treated as edge cases after deployment. 5.2 Who Carries the Risk of Getting It Wrong? Previous studies on human-AI interaction show that effec- tive use of AI involves calibrated reliance, avoiding both over-reliance on incorrect outputs and under-reliance on use- ful ones (Lee and See 2004; Bansal et al. 2021). The fac- tors that impact trust are not limited to system efficiency or performance but also include information communicated about the AI system, such as explanations, source visibility, and institutional credibility (Liao and Sundar 2022). With GenAI specifically, users face additional metacognitive de- mands, including prompting, evaluating outputs, and decid- ing whether an answer is reliable enough to act on (Tankele- vitch et al. 2024). However, in reality, as illustrated by our findings (Section 4.4), reliance extended beyond the user- system interaction; we observed that employees were mov- ing between the GenAI system PEOPLE TOOL, legacy sys- tems STEVE, documents, coworkers, and HR personnel, de- pending on task priority and uncertainty. This raises a ques- tion worth discussing: When an employee navigates this network for HR information and still acts on incorrect in- formation, who is responsible? The system designer, the or- ganization that deployed it, or the employee who used it? This question is consequential in HR contexts, where em- ployees may act on AI-generated information about high- stakes decisions such as pay, leave, benefits, or employ- ment conditions (Budhwar et al. 2023; Dasaklis et al. 2025). When the AI-generated answers appear fluent and deliv- ered through an approved system by the organization, in- stitutional trust can replace the careful verification by the employee (Knowles and Richards 2021; Liao and Sundar 2022). Without clear information on source legitimacy, es- tablished escalation procedures, and robust accountability structures within the organization, responsibility for accu- racy rests with employees rather than with the organization that deployed the system. Design recommendations. At a system design level, workplace GenAI systems should make provenance, uncer- tainty, and fallbacks apparent as default features. Source URLs need to be seen as an accountability architecture, not simply a UI-based feature. Also, the GenAI answers them- selves must provide information about their origin, verifia- bility, and whether human intervention is recommended. At the organizational level, it is important to maintain existing human channels for assistance where possible, rather than replacing them with AI. It is also critical for organizations to establish clear lines of accountability when AI systems are involved in decision-making in high-risk areas, so that the extent to which the system can and cannot be relied upon is clear, and interpretive risk does not fall on individual work- ers by default. 5.3 Knowledge Readiness as AI Infrastructure As stated in (Section 4.5), participants did not view PEO- PLE TOOL as a stand-alone AI model. Users interacted with PEOPLE TOOL through HR artifacts, including arti- cles, metadata, tagging, taxonomy, localization, source ref- erences, training guides, communication, and escalations. When these elements were well structured, GenAI could make HR information faster and easier to access. On the contrary, when some of these artifacts were absent, out-of- date, poorly labeled, or difficult to navigate, the GenAI sys- tem would generate incomplete answers or answers that did not address employeesâ needs. This means that knowledge management becomes a crit- ical part of accountability for workplace GenAI implemen- tations (Alavi, Leidner, and Mousavi 2024). In high-stakes domains, content quality is more than a back-office imple- mentation issue; it determines whether the information sum- marized by AI systems is complete, accurate, up-to-date, and reliable for action (Tankelevitch et al. 2024). Similarly, train- ing and communications are not limited to mere rollout con- siderations. Employees need to be able to frame their ques- tions, interpret the results generated, check their sources, and raise uncertainties. Without such considerations, companies may place AI-related uncertainties on individualsâ shoul- ders. Design recommendations. Knowledge infrastructure should be considered part of the core AI infrastructure. This includes keeping HR content up to date, assigning owner- ship of articles, improving tagging and taxonomies, local- izing terminology, auditing unsuccessful queries, and estab- lishing feedback loops between poor AI answers and knowl- edge owners. Training must be continuous and timely with respect to its use. This involves using brief instructions, ex- amples, rewording help, Q&A, and readily available help lines. In corporate GenAI applications, responsible deploy- ment requires preparing the organization around the AI, not just for the AI itself. 6 Limitations This study has several limitations. First, it is based on a single organizational case. This limits statistical generaliz- ability, but the case is valuable because it captures GenAI adoption during a live workplace transition, where employ- ees could compare the new system with an existing legacy system. Second, the number of survey respondents (n = 25) was relatively small and unevenly distributed across nation- ality, job roles, and work experience within the company. For that reason, the survey data were treated as descriptive only and triangulated with data from interviews and search logs to enable interpretation. Third, the logs captured dif- ferent usage periods for both systems. While STEVEâs logs capture a longer span of searches, PEOPLE TOOL logs had to be confined due to ServiceNowâs retention policies. To ad- dress this limitation, we restricted cross-system comparisons to the same observation window and used the longer STEVE logs only as historical context. Finally, some worker groups central to situational fit, such as shop-floor, shift-based, or less digitally embedded employees, were not systematically sampled as direct participants. Some interviewees held roles that exposed them to these work contexts, so their accounts provide informed yet indirect evidence of uneven fit. Fu- ture studies would do well to include such groups directly to study whether workplace GenAI systems reproduce or re- duce inequalities across different labor conditions. 7 Conclusion In this study, we examine how employees across a multina- tional workforce adopt, resist, and selectively use a GenAI- enhanced HR knowledge system during a live organizational transition from a legacy HR knowledge system to a new GenAI-enhanced system. Through our analysis of both sys- temsâ search logs, a survey, and ten semi-structured inter- views, we show that several factors shaped the adoption process, including situational fit, search literacy, trust cali- bration, and deployment conditions, such as content qual- ity, user guidance, and accountability structures. Drawing on these findings, we argue that equitable workplace GenAI de- ployment cannot be assessed solely by access to the system. Our results raise questions about how organizations should define adoption success, who is responsible for wrong AI- generated HR information, and what conditions are neces- sary for GenAI to be deployed reliably and equitably in high-stakes domains. 8 Researcher Positionality Statement Our research team combines expertise in human-computer interaction (HCI), responsible and sociotechnical AI, orga- nizational studies, computational social science, and human resource management, spanning both the Global North and the South. The team members include PhD and MSc stu- dents, faculty, an industry researcher, and an HR practi- tioner from five countries. One of our team members has a prior professional relationship with the organization, which facilitated data access; interpretive decisions were made collectively and without organizational involvement. Dis- agreements during the thematic analysis stage were resolved through iterative discussion among all researchers. 9 Ethical Statement This study received ethical approval from our institution prior to data collection, and all data were handled in ac- cordance with GDPR requirements. All participation in the survey and interviews was voluntary with no financial in- centive, and informed consent was obtained from all inter- view participants before recording. All identifying informa- tion was removed from transcripts and participants were as- signed anonymous IDs. Search log data were provided by the organization and contains no personally identifiable in- formation. The organization reviewed and approved the data sharing arrangements prior to analysis. References Alavi, M.; and Leidner, D. E. 2001. Review: Knowledge Management and Knowledge Management Systems: Con- ceptual Foundations and Research Issues. MIS Quarterly, 25(1): 107â136. Alavi, M.; Leidner, D. E.; and Mousavi, R. 2024. A Knowl- edge Management Perspective of Generative Artificial Intel- ligence. Journal of the Association for Information Systems, 25(1): 1â12. Ali, D.; Ahmed, M.; Wang, H.; Khan, A.; Jordan, N. P. A.; Kim, S. S.; Muchhala, M. D.; Merkle, A. K.; and Papakyr- iakopoulos, O. 2025. AI Adoption Across Mission-Driven Organizations. arXiv preprint arXiv:2510.03868. Awa, H. O.; Ojiabo, O. U.; and Orokor, L. E. 2017. In- tegrated technology-organization-environment (TOE) tax- onomies for technology adoption. Journal of Enterprise In- formation Management, 30(6): 893â921. Baker, J. 2011. The technologyâorganizationâenvironment framework, 231â245. Springer. Bankins, S.; Ocampo, A. C.; Marrone, M.; Restubog, S. L. D.; and Woo, S. E. 2024. A multilevel review of artificial intelligence in organizations: Implications for organizational behavior research and practice. Journal of Organizational Behavior, 45(2): 159â182. Bansal, G.; Wu, T.; Zhou, J.; Fok, R.; Nushi, B.; Kamar, E.; Ribeiro, M. T.; and Weld, D. 2021. Does the Whole Exceed Its Parts? The Effect of AI Explanations on Complementary Team Performance. In Proceedings of the CHI Conference on Human Factors in Computing Systems. Bhargava, A.; Bester, M.; and Bolton, L. 2021. Employ- eesâ Perceptions of the Implementation of Robotics, Artifi- cial Intelligence, and Automation on Job Satisfaction, Job Security, and Employability. Journal of Technology in Be- havioral Science, 6: 106â113. Braun, V.; and Clarke, V. 2006. Using Thematic Analysis in Psychology. Qualitative Research in Psychology, 3(2): 77â101. Brynjolfsson, E.; Li, D.; and Raymond, L. 2025. Generative AI at Work. The Quarterly Journal of Economics, 140(2): 889â942. Budhwar, P.; Chowdhury, S.; Wood, G.; Aguinis, H.; Bam- ber, G. J.; Beltran, J. R.; Boselie, P.; Cooke, F. L.; Decker, S.; DeNisi, A.; Dey, P. K.; Guest, D.; Knoblich, A. J.; Ma- lik, A.; Paauwe, J.; Papagiannidis, S.; Patel, C.; Pereira, V.; Ren, S.; Rogelberg, S.; Saunders, M. N. K.; Tung, R. L.; and Varma, A. 2023. Human Resource Management in the Age of Generative Artificial Intelligence: Perspectives and Research Directions on ChatGPT. Human Resource Man- agement Journal, 33(3): 606â659. Cheon, E.; and Erickson, I. 2025. Fulfillment of the Work Games: Warehouse Workersâ Experiences with Algorith- mic Management. Proceedings of the ACM on Human- Computer Interaction, 9. Costanza-Chock, S. 2020. Design Justice: Community-Led Practices to Build the Worlds We Need. Cambridge, MA: The MIT Press. Dasaklis, T. K.; Giannopoulos, P. G.; Koutras, D.; Malamas, V.; and Chountalas, P. T. 2025. Large Language Models in Human Resource Management: A Systematic Literature Review of Applications, Open Issues and Future Research Directions. Preprint submitted to Elsevier. Davis, F. D. 1989. Perceived usefulness, perceived ease of use, and user acceptance of information technology. MIS Quarterly, 13(3): 319â340. DellâAcqua, F.; Ayoubi, C.; Lifshitz, H.; Sadun, R.; Mol- lick, E.; Mollick, L.; Han, Y.; Goldman, J.; Nair, H.; Taub, S.; and Lakhani, K. R. 2025. The Cybernetic Teammate: A Field Experiment on Generative AI Reshaping Teamwork and Expertise. Technical Report w33641, NBER. Douglas, M. R.; and Verstyuk, S. 2025. Progress in Ar- tificial Intelligence and its Determinants. arXiv preprint arXiv:2501.17894. Enholm, I. M.; Papagiannidis, E.; Mikalef, P.; and Krogstie, J. 2022. Artificial Intelligence and Business Value: a Litera- ture Review. Information Systems Frontiers, 24: 1709â1734. Ernst & Young LLP. 2025. The Dividend Age: How AI Is Turning Promise into Payoff. Us ai pulse survey: Wave 4, EY. Survey of 500 US senior business leaders. Gu, K.; Grunde-McLaughlin, M.; McNutt, A.; Heer, J.; and Althoff, T. 2024. How Do Data Analysts Respond to AI Assistance? A Wizard-of-Oz Study. In Proceedings of the CHI Conference on Human Factors in Computing Systems, CHI â24. ACM. Helsper, E. J. 2021. The Digital Disconnect: The Social Causes and Consequences of Digital Inequalities. London: SAGE Publications. Humlum, A.; and Vestergaard, E. 2025. The unequal adop- tion of ChatGPT exacerbates existing inequalities among workers. Proceedings of the National Academy of Sciences, 122(1): e2414972121. Ingber, A. S.; and Andalibi, N. 2025. Emotion AI in Job In- terviews: Injustice, Emotional Labor, Identity, and Privacy. In Proceedings of the ACM Conference on Fairness, Ac- countability, and Transparency (FAccT). Institute, D. A. 2026. State of AI in the Enterprise: The Untapped Edge. Technical report, Deloitte AI Institute. Kawakami, A.; Taylor, J.; Fox, S.; Zhu, H.; and Holstein, K. 2024. AI Failure Loops in Feminized Labor: Understanding the Interplay of Workplace AI and Occupational Devalua- tion. In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society (AIES). Knowles, B.; and Richards, J. T. 2021. The Sanction of Au- thority: Promoting Public Trust in AI. In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 262â271. Kwa, T.; West, B.; Becker, J.; Deng, A.; Garcia, K.; Hasin, M.; Jawhar, S.; Kinniment, M.; Rush, N.; Von Arx, S.; Bloom, R.; Broadley, T.; Du, H.; Goodrich, B.; Jurkovic, N.; Miles, L. H.; Nix, S.; Lin, T.; Parikh, N.; Rein, D.; Sato, L. J. K.; Wijk, H.; Ziegler, D. M.; Barnes, E.; and Chan, L. 2025. Measuring AI Ability to Complete Long Tasks. arXiv preprint arXiv:2503.14499. Lee, A. T.; Ramasamy, R. K.; and Subbarao, A. 2025. Un- derstanding Psychosocial Barriers to Healthcare Technol- ogy Adoption: A Review of TAM and UTAUT Frameworks. Healthcare, 13(3): 250. Lee, J. D.; and See, K. A. 2004. Trust in Automation: De- signing for Appropriate Reliance. Human Factors, 46(1): 50â80. Lee, Y. S.; Kim, T.; Choi, S.; and Kim, W. 2022. When does AI pay off? AI-adoption intensity, complementary in- vestments, and R&D strategy. Technovation, 118: 102590. Liao, Q. V.; and Sundar, S. S. 2022. Designing for Respon- sible Trust in AI Systems: A Communication Perspective. In Proceedings of the 2022 ACM conference on fairness, ac- countability, and transparency. ACM. MIT Project NANDA. 2025. The GenAI Divide: State of AI in Business 2025. Technical report, Massachusetts Institute of Technology. Preliminary findings. Mlekus, L.; Bentler, D.; Paruzel, A.; Kato-Beiderwieden, A.-L.; and Maier, G. W. 2020. How to raise technology acceptance: user experience characteristics as technology- inherent determinants. Gruppe. Interaktion. Organisation. Zeitschrift f Ě ur Angewandte Organisationspsychologie, 51: 273â283. Mogaji, E.; Viglia, G.; Srivastava, P.; and Dwivedi, Y. K. 2024. Is it the end of the technology acceptance model in the era of generative artificial intelligence? International Journal of Contemporary Hospitality Management, 36(10): 3324â3339. Oliveira, T.; and Martins, M. d. R. F. d. O. 2011. Literature review of information technology adoption models at firm level. The Electronic Journal Information Systems Evalua- tion, 14(1): 110â121. Raghavan, M.; Barocas, S.; Kleinberg, J.; and Levy, K. 2020. Mitigating Bias in Algorithmic Hiring: Evaluating Claims and Practices. In Proceedings of the Conference on Fairness, Accountability, and Transparency (FAccT), 469â481. Sanchez-Monedero, J.; Dencik, L.; and Edwards, L. 2020. What Does It Mean to âSolveâ the Problem of Discrimi- nation in Hiring? Social, Technical and Legal Perspectives from the UK on Automated Hiring Systems. In Proceedings of the Conference on Fairness, Accountability, and Trans- parency (FAccT), 458â468. Sarkar, A.; Gordon, A. D.; Negreanu, C.; Poelitz, C.; Raga- van, S. S.; and Zorn, B. 2022. What is it like to program with artificial intelligence? arXiv preprint arXiv:2208.06213. Selbst, A. D.; boyd, d.; Friedler, S. A.; Venkatasubramanian, S.; and Vertesi, J. 2019. Fairness and Abstraction in So- ciotechnical Systems. In Proceedings of the Conference on Fairness, Accountability, and Transparency (FAccT), 59â68. Shachak, A.; Kuziemsky, C.; and Petersen, C. 2019. Be- yond TAM and UTAUT: Future directions for HIT imple- mentation research. Journal of Biomedical Informatics, 100: 103315. Tankelevitch, L.; Kewenig, V.; Simkute, A.; Scott, A. E.; Sarkar, A.; Sellen, A.; and Rintel, S. 2024. The Metacog- nitive Demands and Opportunities of Generative AI. In Pro- ceedings of the CHI Conference on Human Factors in Com- puting Systems, CHI â24. ACM. Tornatzky, L. G.; and Fleischer, M. 1990. The Processes of Technological Innovation. Lexington, MA: Lexington Books. van Dijk, J. A. G. M. 2005. The Deepening Divide: Inequal- ity in the Information Society. London: SAGE Publications. Venkatesh, V. 2022. Adoption and use of AI tools: a re- search agenda grounded in UTAUT. Annals of Operations Research, 308(1): 641â652. Venkatesh, V.; Morris, M. G.; Davis, G. B.; and Davis, F. D. 2003. User acceptance of information technology: Toward a unified view. MIS Quarterly, 27(3): 425â478. Warschauer, M. 2003. Technology and Social Inclusion: Re- thinking the Digital Divide. Cambridge, MA: MIT Press. Weidinger, L.; Mellor, J.; Rauh, M.; Griffin, C.; Uesato, J.; Huang, P.-S.; Cheng, M.; Glaese, M.; Balle, B.; Kasirzadeh, A.; Kenton, Z.; Brown, S.; Hawkins, W.; Stepleton, T.; Biles, C.; Birhane, A.; Haas, J.; Rimell, L.; Hendricks, L. A.; Isaac, W.; Legassick, S.; Irving, G.; and Gabriel, I. 2021. Ethical and Social Risks of Harm from Language Models. arXiv preprint arXiv:2112.04359. Weisz, J. D.; Muller, M.; He, J.; and Houde, S. 2023. To- ward general design principles for generative AI applica- tions. arXiv preprint arXiv:2301.05578. Wolfe, D.; Price, M.; Choe, A.; Kidd, F.; and Wagner, H. 2025.Revisiting UTAUT for the Age of AI: Un- derstanding Employees AI Adoption and Usage Patterns Through an Extended UTAUT Framework. arXiv preprint arXiv:2510.15142. Woodruff, A.; Shelby, R.; Kelley, P. G.; Rousso-Schindler, S.; Smith-Loud, J.; and Wilcox, L. 2024. How Knowledge Workers Think Generative AI Will (Not) Transform Their Industries. In Proceedings of the CHI Conference on Human Factors in Computing Systems. Xia, Q. N.; Constantinides, M.; Sarkar, A.; Brumby, D. P.; and Cox, A. L. 2026.âIf Youâre Very Clever, No One Knows Youâve Used Itâ: The Social Dynamics of Develop- ing Generative AI Literacy in the Workplace. arXiv preprint arXiv:2602.01386. Accepted at CHIWORK 2026. Zamfirescu-Pereira, J. D.; Wong, R. Y.; Hartmann, B.; and Yang, Q. 2023. Why Johnny canât prompt: how non-AI ex- perts try (and fail) to design LLM prompts. In Proceedings of the 2023 CHI conference on human factors in computing systems, 1â21. A Descriptive Search Log Analysis Descriptive statistics on the results of the search log anal- ysis are provided in this appendix. The search log analysis helped formulate the survey and interview guide by examin- ing how employees performed searches within the old sys- tem, STEVE, and the new GenAI-assisted system, PEOPLE TOOL. It is essential to understand that these results reflect search behaviors, access to search results, clicking patterns, and exposure to GenAI responses, but not measures of trust, satisfaction, motivation, or adoption intention. The two sys- tems cover different spans of time, which means that any comparisons between them are purely descriptive and direc- tional. A.1 Log Dataset Overview Table 3 summarizes the two log datasets used in the analy- sis. While the legacy system holds a more extensive history, the new system holds only a limited history due to platform retention restrictions. Table 3: Overview of log datasets. ItemDescription Legacy systemSTEVE Newer systemPEOPLE TOOL STEVE windowSept. 2024âAug. 2025 PEOPLE TOOL windowJuly 24âAug. 7, 2025 ContextsEG, NL, PH, PK, QA, SA STEVE size17,040 query rows; 19,562 weighted searches PEOPLE TOOL size1,226 query rows PEOPLE TOOL GenAI-displayed subset343 query rows STEVE fieldsSearch term, country, language, date, search count, result count PEOPLE TOOL fieldsQuery, country, language, times- tamp, has-results flag, click rank, GenAI answer displayed flag Note: The GenAI answer displayed flag means that a GenAI-generated answer appeared in the interface. It does not indicate that the user read, trusted, or acted on the answer. Country codes: EG = Egypt, NL = Netherlands, PH = Philippines, PK = Pakistan, QA = Qatar, SA = Saudi Arabia A.2 Query Behavior Across Systems Query forms for the two systems are depicted in Figures 3 and 4. These findings provide evidence for the claims made in the paper regarding how the users carried forward their legacy behavior in terms of keyword-based searching de- spite having access to a GenAI-powered search engine. It is noteworthy that there was not much of an increase in query length in PEOPLE TOOL and short queries consisting of just one or two tokens were the norm in both cases. Figure 4 illustrates the same trend from a different per- spective. The proportion of one- or two-token searches re- mained virtually unchanged for both systems. It is note- worthy that while GenAI-based searches might demand more precise phrasing, a significant number of people still searched with short, keyword-type searches. StevePeople Tool 0 1 2 1.87 2.11 Avg. tokens Figure 3: Average query length in STEVE and PEOPLE TOOL. StevePeople Tool 0 20 40 60 80 100 83.4 83.3 Queries (%) Figure 4: Share of one- to two-token queries in STEVE and PEOPLE TOOL. A.3 Result Availability and GenAI Exposure Figure 5 distinguishes between three types of interaction phases in PEOPLE TOOL, which include whether there was a response from the search query, whether one of the re- sponses was clicked on, and whether there was a response generated by GenAI. The reason for this distinction is that the use of PEOPLE TOOL does not automatically imply the use of GenAI. ResultsClickGenAI 0 20 40 60 80 100 90.9 58.2 28 Queries (%) Figure 5: PEOPLE TOOL result availability, click behavior, and GenAI answer exposure. Figure 6 compares click behavior when a GenAI an- swer was displayed and when it was not. Searches with a displayed GenAI answer had higher click-through than searches without one. This should be interpreted only as be- havioral engagement. It does not show whether users trusted the answer, found it useful, or acted on it. No GenAIGenAI shown 0 20 40 60 80 100 53 71.6 Any click (%) Figure 6: Click behavior by GenAI answer exposure in PEO- PLE TOOL. A.4 Country and Test-Context Variation Table 4 summarizes PEOPLE TOOL log indicators by coun- try or test context. These differences help contextualize the theme of situated fit. However, they should not be interpreted as cultural differences by themselves. Variation may reflect local HR processes, rollout conditions, content availability, language, sample size, or task mix. Table 4: PEOPLE TOOL indicators by country or test context. Ctx.NResultsGenAIClick (%)(%)(%) EG20585.46.849.3 NL67893.743.765.9 PH2692.342.361.5 PK2777.814.818.5 QA5994.910.257.6 SA23188.35.247.6 Note: Results = searches with returned results; GenAI = searches where a GenAI answer was displayed; Click = searches with any clicked result. Table 5 summarizes dominant search intents observed across country or test contexts. These patterns show that users were not bringing a single generic HR search need to the system. Instead, queries reflected local HR practices and task needs, including policy navigation, formal documents, benefits, insurance, and travel claims. B Descriptive Survey Analysis This appendix reports descriptive survey information used to contextualize the interview findings. The survey was ex- ploratory and was completed by 25 employees with access to PEOPLE TOOL and its GenAI search feature. The results should not be interpreted as statistically representative of the wider organization. Instead, they provide descriptive support for the qualitative analysis by showing patterns in reported use, perceived ease, helpfulness, trust, source-checking, and responses to poor GenAI answers. Table 5: Dominant search intents by country or test context. Ctx.Dominant observed search intents NLPolicy and process navigation, including handbook, salary, leave, and Workday-related queries. SAFormal document and certificate-related queries, including letters and salary certificates. QAHR documents, benefits, medical, passport details, and employee pro- file queries. EGBenefits and insurance-related searches. PKProcedural travel-claim and reimbursement queries. PHSmall sample of short, concrete queries with limited repeated pat- terns. Note: These patterns are descriptive. They should not be interpreted as direct evidence of national culture or user attitudes. Table 6: Overview of survey sections and item types. SectionSurvey items ContextReported work or test country, employment level, and tenure. Usage patternsFrequency of using STEVE and PEOPLE TOOL; common search topics in each system. Systemcompari- son Ease of use, helpfulness, trust, unhelpful or irrelevant re- sults, and system preference. GenAI-specific useFrequency of receiving a GenAI-generated answer; help- fulness of GenAI summaries; source-clicking behavior; response when GenAI does not provide a good answer. Future use and im- provements Open-text suggestions, additional comments, and invita- tion to participate in follow-up interviews. B.1 Survey Respondent Context The survey sample included respondents from seven work- places. Table 7 reports the normalized country or test- location distribution. These distributions are included for transparency only; the small and uneven subgroup sizes mean that country-level differences should not be interpreted as statistically reliable effects. Table 8 reports respondentsâ employment levels. The sample was skewed towards pro- fessional and specialist roles (See Limitations section). Ta- ble 9 reports tenure distribution. The sample included both newer and long-tenured employees, supporting descriptive interpretation of organizational familiarity, but not inferen- tial subgroup claims. Table 7: Survey respondents by normalized work or test lo- cation. LocationN% Germany1248 United States624 India312 Austria14 Switzerland14 France14 Mexico14 Table 8: Survey respondents by employment level. Employment levelN% Professional / Specialist1664 Intern / Entry-level employee28 Contractor / Temporary staff28 Team Lead / Manager28 Senior Manager / Director / Executive14 Other employment category.14 Prefer not to say14 Table 9: Survey respondents by tenure. TenureN% Less than 1 year28 1â3 years936 4â7 years624 8â15 years416 More than 15 years416 B.2 System Use and Preference Figure 7 summarizes the reported frequency of use and sys- tem preference. PEOPLE TOOL was used weekly or daily by more respondents than STEVE, but the continued use of both systems supports the interpretation that adoption was selec- tive rather than a clean replacement. Figure 8 shows that a majority clearly preferred PEOPLE TOOL, but a substantial minority preferred STEVE or gave mixed or unclear prefer- ences. This supports the paperâs framing of parallel-system use and selective adoption. StevePeople Tool 0 20 40 60 80 100 56 76 Respondents (%) Figure 7: Respondentsâ weekly or daily use of each system. 0204060 Mixed / unclear Prefer STEVE Prefer PEOPLE TOOL Respondents (%) Figure 8: System preference reported in the survey. B.3 Perceived Ease, Helpfulness, and Trust Figure 9 compares three system ratings. The key pattern is that PEOPLE TOOL was perceived as more helpful, while ease of use and trust changed less substantially. This sup- ports the interpretation that perceived usefulness alone did not remove the need for trust calibration, search literacy, and organizational support. Ease Helpfulness Trust 0 20 40 60 80 100 64 32 68 6464 72 Respondents (%) STEVE PEOPLE TOOL Figure 9: Survey ratings for ease, helpfulness, and trust. Ease = easy or very easy; helpfulness = very or extremely helpful; trust = mostly or completely trusted. B.4 GenAI Answer Exposure, Helpfulness, and Verification Figure 10 summarizes respondentsâ reported experience of GenAI answers in PEOPLE TOOL. Most respondents re- ported receiving GenAI answers often or always, but fewer rated GenAI summaries as very or extremely helpful. Fig- ure 11 shows how often respondents clicked sources dis- played in GenAI-generated answers. Source-checking was divided: 52% never or rarely clicked sources, while 48% clicked them sometimes or often. This should be interpreted as uneven verification behavior, not as a direct measure of trust. 020406080100 Very/extremely helpful Often/always received 44 64 Respondents (%) Figure 10: Reported GenAI answer exposure and summary helpfulness in PEOPLE TOOL. B.5 Responses to Poor GenAI Answers Figure 12 reports what respondents said they would do when GenAI did not provide a good answer. Rephrasing was the most common response, indicating that users actively repair failed interactions by changing how they ask. Escalation to human or other information channels was less frequent but remains important for understanding fallback practices. B.6 Common Search Topics Table 10 reports the most common standardized search top- ics selected by respondents for STEVE and PEOPLE TOOL. Never Rarely Sometimes Often 0 10 20 30 24 28 2424 Respondents (%) Figure 11: Frequency of clicking sources displayed in GenAI-generated answers. 020406080100 Rephrase Stop and try later Contact HR Ask manager or colleague Other / fallback 76 20 12 12 8 Respondents (%) Figure 12: Reported responses when GenAI did not provide a good answer. Responses were multi-select, so percentages do not sum to 100. These topic distributions show that respondents used both systems for practical HR needs, including policies, leave, payroll, sickness, travel, overtime, and benefits. Because re- spondents could select multiple topics, counts are not mutu- ally exclusive. Table 10: Top standardized search topics reported for each system. Search topicStevePeople Tool Employee handbook / HR policies1314 Annual leave / holiday policies810 Salary slip / payroll information107 Sick leave requests / policies47 Travel expenses / reimbursements65 Overtime policy / work hours65 Bonus / incentives information55 Maternity or paternity leave54 IT access34 Note: Topic responses were multi-select. Counts therefore indicate how many respondents selected each topic and may sum to more than the number of respondents. Table 6 summarizes the structure of the survey instrument. The questionnaire compared the legacy system, STEVE, with the newer GenAI-enhanced platform, PEOPLE TOOL, and included both closed-ended and open-text questions. GenAI- specific questions asked respondents about answer exposure, summary helpfulness, source-clicking, and response strate- gies when the GenAI answer was not useful. C Interview Study Materials This appendix provides the interview materials used in the qualitative phase of the study. Semi-structured interviews were the primary interpretive data source and were used to examine how employees made sense of STEVE, PEO- PLE TOOL, and the GenAI search feature in practice. The interview guide was designed to move from participantsâ general background and transition experiences to more fo- cused questions about usefulness, trust, adoption dynamics, and organizational support. The interview guide shown below is the English version used in the main study. Minor adjustments in wording and follow-up prompts were made during interviews to main- tain conversational flow and to probe relevant examples in greater depth, but the overall thematic structure remained stable across participants. Table 11: Anonymized survey instrument. Survey instrument: Comparing STEVE and PEOPLE TOOL GenAI Search Survey sectionItemsQuestions and response format Survey context and system overview IntroThe survey introduced the study as research on a GenAI-supported HR search feature in a large multinational organization. It explained that responses were anonymous, used for research and improvement, and expected to take approximately 10 minutes. The original researcherâs name, institution, company name, department, email address, form URL, and screenshots were removed in this anonymized version. ⢠STEVE: legacy HR query-based search and database system, matching employee questions with pre-scripted answers and HR documents. ⢠PEOPLE TOOL: GenAI-enhanced HR knowledge and support hub, offering source-backed answers based on human-reviewed organizational content. I. Context and demographics1â3 ⢠1. Work or test country. Open text. ⢠2. Current position or employment level. Options: intern / entry-level employee; professional / specialist; team lead / manager; senior manager / director / executive; contractor / temporary staff; prefer not to say; other. ⢠3. Organizational tenure. Options: less than 1 year; 1â3 years; 4â7 years; 8â15 years; more than 15 years; prefer not to say. I. Usage patterns and system comparison 4â15 ⢠4. Frequency of using STEVE and PEOPLE TOOL. Scale: never; rarely (less than monthly); occasionally (monthly); frequently (weekly); very frequently (daily). ⢠5â6. HR topics searched most often in STEVE and PEOPLE TOOL. Multi-select options: employee handbook / HR policies; annual leave and holiday policies; salary slip / payroll information; sick leave requests and policies; travel expense and reimbursements; maternity or paternity leave; overtime policy and work hours; bonus / incentives information; IT access; other. ⢠7. Ease of use for STEVE and PEOPLE TOOL. Scale: very difficult; difficult; neutral; easy; very easy. ⢠8. Difficulties experienced using either system. Open text. ⢠9. Helpfulness of STEVE and PEOPLE TOOL for completing tasks. Scale: not helpful at all; slightly helpful; moderately helpful; very helpful; extremely helpful. ⢠10. Elaboration on unhelpful experiences. Open text. ⢠11. Trust in search results from STEVE and PEOPLE TOOL. Scale: do not trust at all; slightly trust; moderately trust; mostly trust; trust completely. ⢠12. Elaboration on lack of trust. Open text. ⢠13. Frequency of unhelpful or irrelevant answers. Scale: never (less than 1/10); rarely (1/10); sometimes (3/10); half the time (5/10); often (7/10); almost always (9/10). ⢠14. Topics that returned unhelpful or irrelevant results. Open text. ⢠15. Preferred system and reason. Open text. I. PEOPLE TOOL GenAI search-specific questions 16â19 ⢠16. Frequency of receiving a GenAI-generated answer when searching in PEOPLE TOOL. Scale: always; often; sometimes; rarely; never. ⢠17. Helpfulness of GenAI summaries in PEOPLE TOOL. Scale: extremely helpful; very helpful; somewhat helpful; slightly helpful; not helpful at all. ⢠18. Frequency of clicking sources displayed in GenAI-generated answers. Scale: always; often; sometimes; rarely; never. ⢠19. Response when a GenAI-generated answer is not good. Multi-select options: I always get good GenAI-generated answers; rephrase the question; stop and try again later; contact HR directly; ask my manager or colleague; other. IV. Future use and improve- ments 20â22 ⢠20. Top two or three things that would make PEOPLE TOOL more useful. Open text. ⢠21. Additional comments. Open text. ⢠22. Willingness to join a short confidential follow-up interview; contact information stored separately and excluded from analysis. Open text. Note. This table reports the anonymized survey instrument. Original organization names, system names, researcher names, institutional affiliations, email addresses, form URLs, and screenshots were removed or replaced with anonymized labels. Table 12: Semi-structured interview guide. Interview guideline v.02 (English) Interview sectionApprox. timeGuiding prompts I. Warm-up2â3 min ⢠Role and tenure. ⢠Comfort level with digital tools at work. I. Transition experiences: STEVEâ PEOPLE TOOL 5â6 min ⢠Prior experience with STEVE. ⢠Experience with PEOPLE TOOL and perceived differences. ⢠Moments where one system felt easier or more useful than the other. I. Usefulness and limita- tions of GenAI 6 min ⢠Situations where PEOPLE TOOL was particularly helpful. ⢠Situations where PEOPLE TOOL was unhelpful and how the participant responded. ⢠Types of HR questions best suited to PEOPLE TOOL versus human support, such as colleagues or HR. IV. Trust and information behaviors 4â5 min ⢠Factors influencing trust or hesitation in PEOPLE TOOL responses. ⢠Whether participants checked source articles or relied on summaries, and why. V. Adoption dynamics and social context 6â7 min ⢠Observed differences in adoption by tenure, role, or country. ⢠Influence of managers, HR, or colleagues on PEOPLE TOOL use. ⢠Smoothness of the STEVEâ PEOPLE TOOL transition. ⢠Fit of PEOPLE TOOL for shop-floor versus office employees. ⢠Desired organizational support or training to ease the transition. VI. Closing reflections4â5 min ⢠Key factor influencing adoption or avoidance of PEOPLE TOOL. ⢠Message to organizational leadership about employeesâ experiences with AI in HR. Note. The guide was used flexibly during semi-structured interviews; follow-up questions were adapted to participantsâ roles, examples, and prior responses.