Three Questions to Ask Before Your Next AI Rollout
Many AI adoption programmes are designed backwards. They start with the technology, build a training programme around it, track usage at 30 and 60 days, and declare success when the numbers look plausible.
The research is unambiguous on what this produces. McKinsey surveyed over 10,000 executives this year and found that less than 20% of organisations deploying AI are seeing meaningful operational impact. The gap is not a technology problem. It is a behaviour design problem. And behaviour problems require different questions before you build anything. Here are three that are worth working through before your next rollout.
1. What does embedded adoption actually look like for this role?
Not use AI. Use AI to do what, at what point in the workflow, to what end? There is a meaningful difference between occasional use and embedded adoption: the point at which the new behaviour has become the automatic default, as unremarkable as opening email.
Many adoption programmes are too focused on usage alone. Before any rollout, define what embedded adoption looks like for each role in scope. For example, how can a credit analyst use AI tools to extract, collect, and source information, analyse financial information and visualise data? The specificity matters because it tells you what you are actually measuring for, and whether the conditions in the environment make that behaviour realistic. If that target is unclear, the measurement framework is not pointed at it. And if you are not measuring for embedded behaviour, you are almost certainly not designing for it either.
2. Where is the new behaviour harder than the old one, under real conditions, not ideal ones?
When a new behaviour requires more cognitive effort than the existing one, the existing one wins, particularly under time pressure, cognitive load, or ambiguity. The familiar workflow is automatic. The AI-supported workflow requires active thought until it becomes habitual. Under pressure, people revert to what they know.
This is predictable. And it is addressable, but only if you look for it before launch rather than after you see reversion in the data. For each role in scope, map the specific moments where the new behaviour is harder than the old one. What makes the AI-supported workflow uncertain or risky? What happens when the output is wrong or incomplete, and is the recovery path clear? What peer norms or management signals still make the old way feel safer?
Deloitte’s 2026 Human Capital Trends report gives a useful example from a European telecommunications company. When the organisation simply added an AI “expert” into customer service without redesigning roles or workflows, it saw only a 5% productivity lift. But when it redesigned the human-AI interaction itself — including workflows, trust thresholds, escalation paths and training — productivity increased by 30%. The difference was whether the new behaviour had been made clear, safe and usable inside the flow of work.
3. What does this change ask people to believe about their own expertise?
This is the question that determines whether experienced people genuinely integrate AI or settle into minimal viable use.
In knowledge work, professional identity and credibility are tied to expertise and judgment. AI adoption framed as efficiency — doing more, faster, with less effort — implicitly positions individual expertise as less central to performance. The response is not overt pushback. It is quiet protection of the high-judgment work.
Before launch, ask how this change will land for someone whose credibility rests on doing the work well. Does the framing amplify what they bring, or suggest they should bring less of it? Are the incentives and management signals still rewarding expertise demonstration over AI-augmented performance? If the system rewards the old behaviour while the training asks for a new one, the training will not win.
The answers to these questions will differ by organisation, by function, and by role. But they reliably surface the conditions that determine whether an AI adoption programme reaches embedded behavioural change or stalls at intention.
By Houra Amin, a behaviour-led learning strategist and Director of Blue Jay Learning.
