Small firms often resist generative AI not for lack of skill but because its outputs and value flows are hard to evaluate; that opacity shifts invisible verification work onto frontline staff and undermines managerial oversight, so apparent efficiency gains can fail to translate into verifiable benefits.

Main Finding

Resistance to generative AI (GenAI) in SMEs is not primarily a symptom of technical incapacity but of an "evaluability paradox": GenAI appears accessible and functional, yet the organisational mechanisms that distribute value, accountability and recognition remain opaque. That opacity produces unrecognised labour, eroded managerial visibility, and owners’ misread benefits. Resistance therefore functions as diagnostic feedback—expressed through calibrated engagement, selective reliance and identity-preserving practices—revealing where adoption fails to translate into verifiable value for workers and the organisation.

Key Points

• Evaluability paradox: GenAI feels usable but its outputs and effects are hard for actors to evaluate in terms of who gains value, who is accountable, and who receives recognition.
• Role-differentiated experiences:
  • Operational staff: experience unrecognised labour (hidden extra work, verification burdens).
  • Managers: experience eroded process visibility and uncertainty about where responsibility lies.
  • Owners/founders: perceive promised benefits but often misattribute or overclaim organisational gains.
• Forms of resistance:
  • Calibrated engagement (using tools selectively or limiting scope),
  • Selective reliance (double-checking, keeping human oversight for sensitive tasks),
  • Identity-preserving practices (retaining tasks tied to professional identity).
• Conditional productive engagement arises when organisations:
  • Redistribute system information (who the model did what and why),
  • Preserve relational deliberation (collective interpretation protocols),
  • Draw clear boundaries around human-led work.
• Practical manager checklist (summary):
  • Redistribute system information to operational staff,
  • Make value verifiable at the operational level (metrics, attribution),
  • Establish protocols for collective interpretation of outputs,
  • Formalise feedback loops from staff to decision-makers,
  • Define and communicate explicit boundaries for human-led tasks.
• Interpretation over suppression: treating resistance as feedback allows adaptation that can make GenAI adoption more durable, accountable and socially beneficial.
• Limitations: snapshot of Dutch SMEs during a formative adoption phase; comparative and longitudinal work needed.

Data & Methods

• Empirical base: 108 semi-structured interviews across 75 Dutch SMEs.
• Stratified sampling: interviews deliberately included founders/owners, managers and operational staff to triangulate perspectives and identify convergences, divergences and contests.
• Analytic approach: qualitative coding focused on how actors interpreted and experienced GenAI in everyday work; categorized sources of resistance and organisational responses.
• Temporal/contextual scope: captured early/formative GenAI adoption in Dutch small businesses; not longitudinal or cross-national.

Implications for AI Economics

• Value capture and measurement
  • Efficiency gains reported at organisational level can mask the accumulation of unrecognised labour at the operational level; GDP/productivity signals may overstate welfare gains if redistribution and recognition are not accounted for.
  • Economists should measure not only output changes but who captures value, who bears verification costs, and how time budgets shift across roles.
• Adoption dynamics and productivity estimates
  • Resistance is an information-rich indicator of misaligned incentives; treating it as such can explain slow or uneven productivity effects of GenAI in SMEs.
  • Standard diffusion models should incorporate organisational transparency and evaluability as factors shaping uptake and realized gains.
• Labour market and distributional effects
  • GenAI may shift burdens (verification, monitoring, contextualization) onto lower-paid operational staff even as owners/managers claim aggregate efficiency—implying potential increases in unpaid or invisible labour.
  • Policy and firm-level interventions are needed to prevent value-extraction without recognition or compensation.
• Managerial and policy interventions
  • Managers can increase realised benefits by (a) making model provenance and decision logic more transparent to workers, (b) creating shared interpretation protocols, (c) formalising feedback and attribution mechanisms, and (d) explicitly protecting tasks tied to professional identity.
  • Policy tools (guidance, incentives, reporting standards) that require accountable deployment (traceability, impact reporting, worker involvement) would better align technological gains with decent work (SDG‑8) and accountable innovation (SDG‑9).
• Research implications
  • Empirical studies of AI’s economic impact should combine quantitative measures with qualitative diagnostics of organisational information flows and recognition mechanisms.
  • Longitudinal and comparative work is needed to assess how managerial responses alter the distribution of benefits over time and across institutional contexts.

Overall: the paper reframes resistance not as a barrier to be overcome but as an organisational signal revealing where GenAI’s promised economic gains are unrealised at the level of the people doing the work. For AI economists, this foregrounds the importance of institutional arrangements, evaluability, and distributive measurement when estimating the true social and economic returns to GenAI adoption.