Insights
Why AI Adoption Stalls in the Middle
Ask executives what is slowing their AI programme down and you hear about tooling, data quality, procurement. The honest answer is usually simpler: nobody in the middle of the organisation has been told what happens to their people if it works.
Two pieces of 2026 evidence make this hard to dismiss. In a pre-registered experiment with 2,000 managers in the US and UK, the Institute for Fiscal Studies showed managers short videos about AI. Managers who saw evidence of AI’s labour-displacing potential reported intent to adopt and advocate for AI that was 0.4 to 0.5 standard deviations lower, and pulled back their hiring intentions too. Managers who saw evidence of productivity gains did not move at all.
The productivity case persuaded nobody
The productivity case, in other words, persuaded nobody. The threat case persuaded everybody — to stop. Separately, work using the Gallup Workforce Panel (more than 30,000 US employees, 2023 to early 2026) found that employees who say their organisation has a clear AI strategy are roughly 27 percentage points more likely to use AI frequently. Frequent use without that clarity shows near-zero association with engagement and a positive association with burnout. Strategic clarity was concentrated in places with developmental feedback and trust in leadership — which suggests people infer strategy from how their managers behave, not from what the policy document says. Put those together and you get an uncomfortable diagnosis. Adoption is not primarily a technical problem or a training problem. It is a problem of what people believe the technology means for them. And the belief is formed locally, by a manager who has done the arithmetic and does not like the answer. You cannot fix that with enthusiasm.
Commitments that are costly to break
You can only fix it with commitments that are costly to break. Three seem to matter most.
1. A stated reinvestment rule. Before deployment, say in writing where recovered hours go. Not “efficiency” — a destination. Twenty per cent of recovered time to customer contact, quality review, or team thinking, with the rest returned to capacity. A rule can be audited. An intention cannot.
2. Headcount decoupling, with a horizon. Say plainly whether this deployment is linked to a staffing decision, and for how long. Vagueness is not neutral. Vagueness is read as the worst case, and the manager protects the team by slowing the project down.
3. Contribution in the record, not just output. If the only thing performance systems can see is throughput, then AI making throughput cheap makes people look redundant. So the record has to widen: the question that stopped a bad plan, the judgement call, the knowledge someone wrote down so a colleague would not have to rediscover it.
What happens to their people if it works
None of this requires a new platform. It requires one page, written before the tool arrives rather than after the resistance appears. What strikes me most in the IFS result is the asymmetry. The upside argument moved no one; the downside argument moved everyone. They are not confused about the productivity numbers. They are waiting to hear what the productivity is for. Perhaps the first deliverable of any AI programme should not be a pilot. Perhaps it should be a sentence a manager can say out loud to their team without flinching.