Technical Summary: AI Adoption Across a Multinational Workforce

Problem Statement

Despite the rapid acceleration of Generative AI (GenAI) deployment in workplaces, there is a significant gap in understanding who adopts these systems, who benefits, and who is left behind. Existing adoption models (e.g., TAM, UTAUT, TOE) often fail to capture the complex, situated user experiences within dynamic organizational transitions. Specifically, they offer limited insight into how employees learn, adapt, and develop workarounds when integrating probabilistic, opaque AI systems into existing workflows. In high-stakes domains like Human Resources (HR), poor design or misalignment can lead to systems that exclude specific social groups, produce untrustworthy outputs, or concentrate efficiency benefits among a privileged subset of workers while leaving others behind.

Methodology

The study employs a mixed-method, single-case study design within a multinational technology organization (300,000+ employees) transitioning from a legacy HR search system ("STEVE") to a GenAI-enhanced system ("PEOPLE TOOL"). The research utilized a three-phase sequential design:

1. Search Log Analysis:

   • Data: One year of logs from the legacy system (STEVE) and two weeks of logs from the new GenAI system (PEOPLE TOOL).
   • Scope: Data covered six countries (Egypt, Netherlands, Philippines, Pakistan, Qatar, Saudi Arabia).
   • Focus: Query volume, phrasing, length, result availability, click-through rates, and GenAI exposure. Non-English queries were translated for cross-country topic comparison.
   • Constraint: Cross-system volume comparisons were restricted to a common observation window (July 24–August 2, 2025) to ensure validity.

2. Exploratory Survey:

   • Participants: 25 employees across six countries (Germany, US, India, Austria, Switzerland, France).
   • Measures: Usage frequency, perceived usefulness, ease of use, trust, system preference, and behaviors regarding source checking and query reformulation.
   • Role: Descriptive evidence to contextualize findings and refine interview protocols; not statistically representative.

3. Semi-Structured Interviews:

   • Participants: 10 employees from four countries (Germany, USA, Spain, Norway), comprising both project team members (implementation/practitioners) and end-users.
   • Analysis: Reflexive thematic analysis following Braun and Clarke (2006) to identify patterns in adoption dynamics, trust calibration, and organizational conditions.

Key Findings

1. Adoption as Selective Use Across Parallel Systems

Adoption was not a linear migration from legacy to AI. Instead, employees engaged in selective use, switching between PEOPLE TOOL, STEVE, source articles, and human colleagues based on situational needs.

• Drivers: Employees chose PEOPLE TOOL for speed, integration, and clarity. They reverted to STEVE for familiarity, predictable article-based results, or when the new system added cognitive burden without clear benefit.
• Implication: Aggregate adoption metrics can mask the continued reliance on legacy systems or informal channels.

2. Situated Fit Determines Utility

The utility of the GenAI system was highly dependent on the "situational fit" between the tool's design assumptions and the employee's work positionality (role, language, tenure, device access).

• Exclusionary Design: The system was optimized for office-based, digitally confident, English-speaking users with organizational familiarity.
• Disadvantaged Groups: Shift-based, factory, and blue-collar workers often lacked device access or the organizational terminology to formulate queries, leading to systematic exclusion despite formal access.
• Tenure Effects: New recruits benefited most due to high information uncertainty, while tenured employees relied less on the tool as they already possessed internal knowledge and established resolution pathways.
• Language & Context: Local search behaviors varied significantly by country (e.g., policy searches in the Netherlands vs. certificate searches in Saudi Arabia), and language barriers significantly reduced effectiveness.

3. Learning a New GenAI Search Logic

Effective use required employees to develop a new form of search literacy distinct from traditional keyword searching.

• Skill Gap: Users had to learn to phrase questions, add context, rephrase failed queries, and interpret summaries.
• Behavioral Evidence: While 76% of survey respondents reported rephrasing questions when GenAI answers were poor, log data showed that query length and token counts remained largely unchanged (average 1.87 to 2.11 tokens), indicating that many users continued using traditional keyword habits rather than adapting to the conversational logic of GenAI.
• Competence: Adoption depended on the ability to translate HR needs into questions the system could answer, a skill not automatically conferred by system access.

4. Trust Calibration Through Checking and Fallback

Trust in GenAI was not blind; it was calibrated through source-checking, comparison, and fallback mechanisms.

• Verification: 52% of survey respondents rarely or never clicked the source links provided in GenAI answers, while others used them to verify plausibility.
• Institutional Trust vs. Verification: Some employees trusted the system simply because it was an internal corporate tool, potentially leading to over-reliance on fluent but incorrect answers.
• Risk Management: In high-stakes HR scenarios (pay, leave, benefits), employees frequently used GenAI for orientation but relied on legacy systems, peers, or HR staff for final verification. Trust was distributed across the answer and the wider information ecosystem.

5. Knowledge Conditions as Infrastructure

The reliability of GenAI outputs was fundamentally tied to the quality of the underlying organizational knowledge infrastructure.

• Content Quality: Poor tagging, outdated articles, vague taxonomies, and lack of localization directly resulted in unhelpful or incorrect GenAI answers.
• Guidance: One-off training was insufficient. Effective adoption required continuous, just-in-time guidance on how to search, verify, and escalate issues.
• Interdependence: GenAI reliability cannot be separated from content quality; the AI is only as effective as the knowledge it retrieves and synthesizes.

Key Contributions

1. Empirical Evidence of Unequal Benefits: The paper provides empirical evidence that workplace GenAI adoption produces unequal benefits. It demonstrates that when systems are designed around assumptions of office access, digital confidence, and language fit, workers outside these assumptions are systematically underserved, even when they have formal access.
2. Design Considerations for Inclusive Deployment: The study translates findings into design considerations for high-stakes environments. It argues that organizations must:
   • Design systems that account for role- and context-sensitive benefits for different social groups.
   • Treat the organizational knowledge infrastructure (content quality, tagging, localization) as a core component of AI infrastructure to improve accountability and usability.
   • Implement mechanisms for trust calibration, making provenance, uncertainty, and fallbacks visible by default.

Significance and Claims

The paper claims that access does not equal adoption or benefit. It challenges the notion that deploying a GenAI system automatically yields efficiency gains for the entire workforce. Instead, it posits that adoption is a sociotechnical process shaped by the interplay of user positionality, search literacy, trust calibration, and the quality of the underlying knowledge base.

The authors argue that to achieve equitable deployment, organizations must move beyond aggregate usage metrics and investigate who is missing from the logs, who requires repeated reformulation, and who relies on intermediaries. Furthermore, they highlight the critical issue of accountability: in the absence of clear escalation procedures and source transparency, the risk of acting on incorrect AI-generated information falls disproportionately on individual employees rather than the organization. The study concludes that responsible GenAI deployment requires preparing the organization around the AI (via knowledge management and continuous training) just as much as deploying the AI itself.