Microsoft published its 2026 Work Trend Index on 5 May 2026, the largest annual data set on how AI is actually used at work, drawn from 20,000 full-time knowledge workers across ten countries. The headline number boards have been quoting is the agent one: 15x year-on-year growth in active agents on Microsoft 365, rising to 18x inside large enterprises. The number worth quoting in the same meeting is more sobering. Only 26% of AI users say their leadership is clearly and consistently aligned on AI.
The gap between those two numbers is the story.
What Microsoft Actually Found
The 26% leadership-alignment figure is the first half of the diagnosis. The second is the manager one. When line managers actively model AI use in their day-to-day work, employees on those teams report a 17-point increase in the perceived value of the tools, a 22-point increase in critical thinking about AI use, and a 30-point increase in trust in agentic AI. Same tooling. Same training. Different manager behaviour. Different outcome.
Microsoft's own framing is that organisational factors, defined as culture, manager support and talent practices, account for roughly twice the reported AI impact compared with individual factors. Two-thirds of the variance sits in the organisational layer, one-third in the individual. The 19% of users Microsoft labels Frontier Professionals are simply the people whose organisation has built the second part. They use the same Copilot the rest of us use.
Context for the 26% figure: 66% of AI users say the tools have allowed them to spend more time on high-value work, and 58% say they are producing work they could not have done a year ago. The technology, on the user side, is delivering. The leadership and manager layer is the bottleneck most procurement decks have not budgeted for.
Why The Standard Diagnosis Misses
The default response to a low adoption number is procedural. Add an AI module to the LMS. Mandate a prompt-writing workshop. Bring in a vendor demo. None of those touch the moment a manager is asked, in the meeting they are running, whether they actually use the thing themselves.
Two findings from the wider research literature explain why awareness training so rarely shifts adoption. The first is Roediger and Karpicke (2006), in Psychological Science: being asked to retrieve and apply material increases long-term retention by around 50% compared with re-reading the same content. A two-hour Copilot module behaves like the re-reading condition. Eight weeks later, the manager who is meant to model the behaviour is operating on the demo they half-watched in March.
The second is Bandura's social-learning insight, given empirical legs in Edmondson's (1999) Administrative Science Quarterly work on psychological safety: people copy what their manager visibly does, and they take risks in proportion to whether the manager has visibly taken them first. The 30-point trust lift Microsoft measured is not a mystery. It is what happens when a senior person opens the tool in front of a junior one and says, on purpose, "this is how I use it, and this is where it went wrong last week."
What Most Organisations Do (And Why It Stays At 26%)
The default 2026 AI strategy is recognisable. A central enablement team is set up. A licence count is reported quarterly. A best-practice playbook is published on the intranet. A small group of enthusiasts becomes very productive, and the variance between teams widens. The board concludes that the rest of the workforce is "change-resistant".
The Microsoft data does not support that reading. The same workforce, under managers who were equipped to model the tool, moves 17 to 30 points. Resistance, in the data, is mostly absence of rehearsal at the layer above.
What Works
The organisations whose AI adoption, productivity and trust numbers have moved upward share three habits.
They equip the manager layer to model in public. The manager has to use the tool in the meeting, not in private, and has to narrate the use. The point is not virtuosity. The point is visibility, including the failures.
They put the conversation in front of managers in a low-stakes setting before the real one arrives. Building on academic behaviour-change work from UCL, Cambridge and Bocconi, our own research found that immersive role-play with professional actors was around 20% more effective than passive modalities such as slide-show or video e-learning at moving observed skill. The same study found participants overestimated their skill before measurement, the Dunning-Kruger pattern that AI rollout decks tend to flatter.
They measure the right thing. Not licence count. Not training completion. The honest test is whether the manager who runs the Tuesday meeting opens the tool in it, names what they used it for, and asks the team what it has been useful and useless for this week. If that conversation is not happening, the 26% is a forecast.
This is the work behind our leadership rehearsal labs and our immersive simulations. Professional actors stage the meeting, with the AI tool open, the awkwardness intact, and the leader narrating in front of their team for the first time in a setting where the cost of getting it wrong is rehearsal, not reputation.
The honest test for any organisation reading the Microsoft numbers this month is simple. In your last leadership meeting, did the most senior person in the room open the tool, use it visibly, and say what they used it for? If not, 26% is not a global statistic. It is the ceiling you have set for yourself.
Book a free 30-minute consultation → or read about our research-backed approach.