The Seventeen Percent Problem
McKinsey's 2025 workplace AI report found a 3x gap between what leaders think employees are doing with AI and what employees are actually doing. That gap is the real crisis.
There’s a number buried in McKinsey’s January 2025 “Superagency in the Workplace” report that I keep coming back to, because it names a problem that I’ve watched play out in every organisation I’ve worked with or built over the past two years.
The report surveyed 3,613 employees and 238 C-suite executives across six countries. They asked both groups how many employees were using generative AI for at least 30% of their daily work.
C-suite leaders estimated: 4%.
Employees reported: 13%.
A 3x gap. Leaders don’t even know what their people are already doing. They’re managing in the dark — making AI strategy decisions, setting AI budgets, designing AI training programs — based on a picture of reality that is three times too conservative.
That gap is the real problem. Not whether AI will take jobs. Not whether your company has the right large language model. The gap between what leadership thinks is happening and what is actually happening is where strategies go to die.
The adoption shape everyone gets wrong
When most leaders think about AI adoption in their organisations, they imagine a clean bell curve. Early adopters at the left tail, resistors at the right, majority in the middle waiting for clear direction.
The actual shape is more fractured than that.
What’s really happening: a subset of employees — call it 10 to 15%, consistent with McKinsey’s numbers — are using AI deeply, daily, in ways that are materially changing their output. They learned on their own. They found tools, experimented, built personal workflows. They are, right now, producing at a meaningfully different level than their peers.
Another subset — probably 25 to 30% based on Gallup’s Q3 2025 data showing 45% of US employees use AI “at least a few times a year” — are dipping in occasionally. They’ve used ChatGPT to draft an email or summarise a document. They know it exists. They don’t have a systematic practice.
The remaining majority either haven’t engaged substantively, or are actively sceptical, or work in roles where access or context hasn’t made adoption obvious.
This isn’t a bell curve. It’s a bimodal distribution becoming trimodal, with the early deep-adopters pulling away from the rest faster than the rest are catching up. That has enormous implications for team dynamics, performance management, and career trajectories — none of which most companies have thought through.
What the training gap looks like from the inside
McKinsey found that 48% of employees rank training as the most important factor for successful AI adoption. And nearly half report receiving minimal or no training. Those two facts together are a complete explanation for why most organisations are moving far slower on AI than the technology would allow.
I’ve lived this from both sides. When I’m building a team and trying to push them into new capability, the training problem is real. It’s not that people are resistant. It’s that they’re being asked to adopt tools without context — without understanding why the tool is relevant to their specific role, without examples from work that looks like their own work, without a low-stakes environment to experiment in.
The tools themselves don’t make this easier. AI products have a habit of being impressive in demos and confusing in practice. The gap between “this can do anything” (the demo impression) and “this keeps getting my context wrong” (the actual experience) creates a disillusionment cycle that is completely predictable and almost completely preventable — if you do real onboarding instead of just dropping a licence on people.
Microsoft’s Work Trend Index 2025, which surveyed 31,000 workers across 31 countries, found something that cuts directly to this: when managers actively modelled AI use, employees reported a 17-point lift in reported AI value, a 22-point lift in critical thinking about their AI use, and a 30-point lift in trust in agentic AI. That’s not a small effect. That’s the difference between an organisation that’s actually making progress and one that’s paying licence fees for tools nobody uses.
The number one lever for AI adoption isn’t the tool. It’s the manager.
The perception gap at the top
One of the things that surprised me most when I dug into McKinsey’s report: the gap isn’t just between leaders and employees on current adoption. It’s on future expectations too.
47% of employees believe that AI will replace 30% or more of their work within a year. Leaders believe this about roughly half as many employees. Employees expect AI to transform their work faster than their managers do.
This is the opposite of the story that usually gets told. The common narrative is: employees are scared of AI, leaders need to manage change carefully, people are resistant. The McKinsey data tells a different story. Employees who are actually using the tools have a more realistic — and in many cases more optimistic — picture of what AI can do in their specific roles than the leaders above them.
The resistance narrative is a management excuse, not an empirical finding. In my experience, it’s also often a projection — leaders who haven’t personally engaged deeply with AI tools project their own discomfort onto the teams they’re supposed to be developing.
I’ve had the specific experience of asking a senior leader at a mid-size company what tools their team was using, and getting a very confident answer — “we use X tool for this, Y tool for that” — and then talking to the people three levels down and discovering that approximately four of them had built entirely parallel workflows using tools the leader had never heard of and definitely hadn’t authorised. The actual AI capability in the organisation was underground, invisible to leadership, and doing real work.
That’s not an exception. That’s the median story in most organisations right now.
Closing the gap: what I’ve actually tried
I’m not going to list a bunch of frameworks. What I’ll do is tell you what I’ve tried, what worked, and what didn’t.
What didn’t work: Announcing that AI is now a priority. Sending links to articles. Creating an “AI committee.” Running a single all-hands training with a vendor. None of these moved the actual distribution of capability in the team in any meaningful way.
What worked, partially: Identifying the existing underground adopters — the people who were already using AI tools daily — and making them visible. Giving them a platform to share what they were doing and why. This had two effects: it signalled to the rest of the team that early AI adoption was valued rather than punished (some people had been hiding it, assuming it would be seen as cutting corners), and it gave the rest of the team concrete examples from work that looked like their work, not a vendor demo.
What worked well: Role-specific workflow mapping. Not “here are AI tools” but “here are the three things you do most frequently in your role, and here is exactly how AI changes each of them.” This requires time and individual context, but the adoption rate after this kind of session is dramatically higher than after any general training.
The manager lever: The 30-point lift in AI trust that Microsoft found when managers model AI use is real in my experience. When I personally share my actual AI workflows — including the failures, the weird outputs, the prompts that work — the team calibrates faster and more accurately than when they’re figuring it out alone. Modelling imperfect use is more effective than performing expert use. Nobody trusts a demo that goes perfectly. Everyone recognises a workflow that goes well except for that one thing that keeps being annoying.
The window
Here’s the thing about the McKinsey gap that feels urgent to me. The 13% who are using AI for 30%+ of their daily work right now are not waiting for their organisations to catch up. They are compounding. Every week they use AI productively, they get better at it — better prompts, better judgment about when to use it and when not to, better intuitions about model limitations.
The gap between them and the rest of the organisation grows every month that nothing structural changes. And at some point, that gap produces outcomes that are hard to close: better output, faster output, different quality of thinking. The early adopters aren’t just ahead on tools. They’re ahead on capability in a way that tool adoption alone can’t catch up to.
The leadership perception gap — thinking 4% when it’s 13% — means that most organisations are not even registering that this divergence is happening inside their own walls.
That’s the seventeen percent problem. Not that seventeen percent of people are using AI. That the gap between what’s real and what leadership perceives is approximately that order of magnitude — and in that gap, the future of the organisation’s capability is quietly being decided.
