A two-phase longitudinal study across 350 frontline workers in a live operational environment — measuring what happens when AI governance moves from policy document to platform architecture.
The IMF estimates that 40% of global jobs are now exposed to AI disruption — yet only 26% of jobs in low-income countries face AI exposure, compared to 60% in advanced economies. This asymmetry does not mean developing economies are protected. It means they have less capacity to benefit and fewer governance frameworks equipped to respond.
Frontline workers in high-churn service sectors — hospitality, retail, food service — are simultaneously the most exposed and the least supported by reskilling infrastructure. They are also the population most often excluded from the design of AI systems meant to serve them.
This research was designed precisely for them.
IMF · Gen-AI and the Future of Work (Jan 2024)
Yet 170M new roles also projected — net gain depends entirely on governance and reskilling investment.
WEF Future of Jobs Report 2025
IMF (2024) — same source
"Technology scales fast. Trust scales slowly. And this asymmetry is the central governance challenge of our era."
— Goudhaman, Dahiya & Kumar (2026) · CERE 2026 IIM Indore · Paper ID 41Most AI governance frameworks are written at the level of principle. They articulate values — fairness, transparency, accountability — but leave a critical gap between declaration and deployment. This research was designed to close that gap.
Two years of platform-generated behavioural data from a real operational environment gave us something that policy documents rarely have: objective, longitudinal evidence about what actually happens when AI enters a frontline workplace.
No self-report data. No survey instruments. Behavioural trace only — completion records, assessment scores, time-on-task per module. This methodological choice was deliberate: to address common method bias (Podsakoff et al., 2003) and to ground every finding in what workers actually did, not what they said they would do.
What This Study Found in the Field
The study was designed to test what happens when AI governance shifts from a controlled doctoral pilot to a live operational environment at 2.5× scale — with reduced direct supervision and AI governance as the primary accountability layer.
AI Micro-Agent Functions — Three Constrained Modes Only
Timed completion notifications. No decision-making authority. No escalation capability.
Adaptive content sequencing based on learner progress. Algorithm adapts; human reviews outcomes.
Post-assessment feedback supporting correct responses. Supportive only — no punitive signals.
Constrained agency design follows Amershi et al. (2019) · doi.org/10.1145/3290605.3300233
The study introduced a new theoretical concept: Constrained Agency — AI decision-making authority deliberately restricted to specific, low-stakes task-level functions. All higher-order decisions remain exclusively with human managers.
All AI outputs subject to manager review before action
AI restricted to 3 task functions only — no scope creep
Exception-handling stays exclusively with human managers
Platform logs serve as accountability evidence — objective behavioural trace
No algorithm aversion detected in either phase — governance was the trust anchor
"Constrained Agency is not a limitation on AI. It is the condition under which AI earns the right to be trusted."
— Theoretical contribution, Goudhaman, Dahiya & Kumar (2026) · Building on Hemmer et al. (2025) and Amershi et al. (2019)When AI is constrained and governed, speed and learning depth are not in conflict — they are complementary. Workers who spent less time per module scored higher on assessments. This finding challenges the assumption that reduced time-on-task signals disengagement. In a governed AI environment, it signals learning fluency and consolidation. The Inverse Learning Paradox is a new theoretical contribution to the AI governance and learning science literature.
| Metric & Hypothesis | Pilot 2024 (n=100) | Scale 2025 (n=250) | Change & Interpretation |
|---|---|---|---|
| Completion Rate H1 — AI Adoption | 100% | 86.82% |
✓ H1 SUPPORTED
−13.18% — operational normalisation, not disengagement. Above the 85% institutional threshold. AI micro-agents sustained adoption despite 15× volume increase. |
| Assessment Score H2 — Learning Quality | 79.12 pts | 83.38 pts |
✓ H2 SUPPORTED
+4.26 pts — deeper learning at scale despite reduced supervision. AI reinforcement cues drove consolidation of correct responses. |
| Time-on-Task H3 — Learning Efficiency | 10.4 min | 5.98 min |
✓ H3 SUPPORTED
−4.42 min (−57%) — learning fluency and efficiency gain. Inverse correlation with quality = Inverse Learning Paradox. The Paradox reframes efficiency as a proxy for mastery, not disengagement. |
| Algorithm Aversion H4 — Trust in AI | None detected | None detected |
✓ H4 SUPPORTED
No aversion across either phase. Governance structures sustained continuous AI engagement. Managerial oversight was the trust anchor — not the AI's capability alone. |
All figures: SafetyCulture platform behavioural logs 2024–2025 · Toscano Restaurant Chain, India · Goudhaman, Dahiya & Kumar (2026) CERE IIM Indore Paper ID 41
"Developing economies are not 'behind.' They are designing adaptive, resilient, human-centred AI governance under real-world constraints — and the world must learn from them."
AI designed for workers across education levels, languages, and digital literacy — not just skilled users. This is not an accommodation. It is a design standard that raises the floor for everyone.
High-churn, diverse, multilingual frontline workforces as a design constraint — not an edge case. The governance framework must hold across this diversity.
Trustworthy AI built under real budget and infrastructure pressure — producing governance models that are replicable globally, not just in well-resourced environments.
Governance that bends to operational context rather than imposing uniform top-down templates. The pilot-to-scale playbook is replicable in healthcare, retail, and manufacturing.
Workers as active participants in trust-building — not passive subjects of AI transformation. This is not a soft principle. It is what the behavioural data shows drives adoption.
AI deployment stress-tested by real-world complexity, churn, and 15× scale increase — producing evidence about what governance architectures actually hold under pressure.
Each recommendation is grounded in a specific finding from this longitudinal study — not in abstract principle, but in what 350 frontline workers' behavioural data showed across two years of live deployment.
Embed review, escalation, and role bounding into deployment design from the beginning — not as audit, but as structure. Governance that is architectural cannot be switched off under operational pressure. Governance that lives only in a policy document will be.
Finding: governance architecture prevented algorithm aversion across both phases. Zero aversion detected.Completion rates alone are insufficient signals of AI adoption quality. Policymakers and organisations must track assessment quality, time-on-task, and aversion indicators in parallel. The Inverse Learning Paradox demonstrates why: a falling metric (time-on-task) masked a rising outcome (learning quality).
Finding: completion fell 13.18% while quality rose +4.26 pts — the opposite of what a completion-only lens would suggest.Frontline learners are stakeholders in governance, not end-users of a system designed without them. Agency over pacing, escalation rights, and visible accountability structures sustained engagement without producing aversion. This is structural trust-building — not soft policy.
Finding: no aversion in either phase, despite 15× training volume — workers remained active participants throughout.Restrict AI authority to specific, low-stakes task-level functions. All higher-order decisions — assignment, escalation, exceptions — must remain human-held. Role bounding is the operational difference between AI as assistive scaffolding and AI as threatening automation.
Finding: three constrained functions (reminders, pacing, reinforcement) sustained adoption at 86.82% across 250 workers.Governance must live in platform architecture, manager training, data infrastructure, and review cycles. This study's pilot-to-scale playbook is replicable across healthcare, retail, and manufacturing — anywhere frontline workers intersect with AI in high-stakes, low-resource environments.
Finding: SafetyCulture platform logs functioned as real-time governance audit trail — making accountability visible at scale."The future of AI governance will not be decided only in laboratories or policy rooms. It will be decided where people learn, work, struggle, adapt — and choose to trust."
The four foundations of trustworthy AI
Trustworthy AI is not built through automation alone.
Goudhaman, S., Dahiya, R., & Kumar, A. (2026). Designing Trustworthy AI: Human Oversight Lessons from India for Developing Economies. Proceedings of the 16th International Conference on Excellence in Research and Education (CERE 2026). IIM Indore. Paper ID 41.
References