The hard part is you

Eighty per cent of the people McKinsey surveyed this year said AI has made them personally more productive. Only 37 per cent said it had shown up in their organisation's earnings, and that number hasn't moved in twelve months.

I think that gap is the most important number in AI right now. The individuals are getting faster and the organisations they work for are not, which means something between the person and the P&L is soaking up the capability before it arrives. BCG's numbers point the same way: of 1,250 companies, 5 per cent are getting substantial value from AI and 60 per cent are getting almost none, despite spending real money. They are buying the same models from the same handful of vendors. Whatever separates the 5 per cent from the 60, it sits in the organisation.

So the question I'd want every executive team asking isn't how to adopt AI faster. It's why their organisation can't absorb the capability it already has. In almost every case I see in my advisory work, the answer comes back to how leaders respond to the uncertainty the technology creates.

The reflex that feels like good management

When the environment gets less predictable, leaders reach for control. That isn't a character flaw. Organisational psychologists named it more than forty years ago: Barry Staw, Lance Sandelands and Jane Dutton called it threat-rigidity, the tendency of people and organisations under threat to narrow the information they take in, pull decisions towards the centre and fall back on the responses they know best. Decades of running stable, process-driven businesses taught a generation of executives that control and good management were the same thing.

So the response looks responsible. Another governance framework, another approval layer, a steering committee, a risk assessment before anyone tries anything, senior sign-off before a team can experiment. Each of those is defensible on its own, and together they build an organisation that is structurally unable to learn at the speed the technology is moving.

I call this the Control Trap. It's one of three I see constantly, alongside the Efficiency Trap and the Pilot Trap, and it's the hardest to see from inside because every decision that builds it looks like prudence.

I watched a client in a capital-heavy industry do exactly this. Their leadership team spent a long stretch on AI policy, governance and automation, and almost every conversation came back to one question: how much more efficient can this make us? What they missed was happening underneath them. Their people were getting better at using AI the whole time, and nobody at the top noticed, so all that growing capability was aimed at shaving cost off the way the business already worked. Meanwhile their customers were moving. The focus on control and cost was steering the business away from where those customers were heading, when the same people, given room to build their skills and a mandate to chase revenue growth and serve customers better, could have taken it there. I came away doubting they would get where they needed to go, because underneath all that governance sat a leader who didn't trust his people, or didn't believe they could ever become genuinely advanced. The controls were that belief, written down as policy.

I want to be careful here, because this argument gets misread as a case for less governance, and the evidence says otherwise. McKinsey's AI high performers, the roughly 6 per cent who attribute at least 5 per cent of EBIT to AI, are more likely than others to report negative consequences from AI, and they also work to protect against a wider range of AI risks. They are taking more risk and managing it better. What differs is the job the governance is doing. In the Control Trap it exists to grant permission, one request at a time. In the organisations pulling ahead it sets guardrails, and inside those guardrails the people closest to the problem are expected to move without asking.

You cannot Six Sigma your way into a learning organisation

For most of my career, advantage came from doing known things more efficiently. Scale, process discipline, execution rigour. Lean, Six Sigma, business process re-engineering and continuous improvement all rest on the same assumption: the work is knowable and repeatable, and it can be broken into steps you measure and refine.

AI breaks that assumption, because it changes what the work is as well as how fast it gets done. Most organisations are still pointing it at the old question anyway. Around 80 per cent of McKinsey's respondents set efficiency as an objective for their AI work. The high performers set growth and innovation goals as well, and they are nearly three times as likely to have redesigned their workflows from scratch. Even the efficiency case is running behind the hype: in 2025, 32 per cent of respondents expected AI to cut their headcount over the following year, and a year later 14 per cent said it had.

An automation lens gets you pilots and incremental savings. A redesign lens gets you structural change in how decisions are made, how knowledge moves and how the organisation learns. That choice decides where the money goes and what leaders spend their attention on.

Return on Learning

If optimisation is no longer the main source of advantage, the replacement, I believe, is learning velocity: how quickly an organisation can detect change, interpret it, act on it and rebuild how it operates around what it found.

I call the payoff Return on Learning, a deliberate jab at the ROI language that dominates AI business cases. Most of those cases measure the return on a technology purchase. Return on Learning measures the rate at which the organisation builds capability it didn't have before.

In the Intelligence-Centred Enterprise framework I co-created, that cycle has four moves. Sense means picking up weak signals across customers, markets, operations and people, including the ones your reporting doesn't show you. Reason is where data, context and judgement meet, and where AI is most useful as a thinking partner rather than a decision-maker. Act means teams closest to the signal move quickly within guardrails, and it is the move the Control Trap breaks first. Learn means the finding changes something durable: a process, a policy, a model, a mental model.

When that loop turns quickly, capability compounds. BCG describes its future-built firms reinvesting their AI returns into people and technology, which is a learning loop under another name, and the gap between them and everyone else is widening each year.

Fund the dip, or don't start

There's an uncomfortable part most frameworks skip: this gets worse before it gets better.

The best evidence comes from Kristina McElheran, Erik Brynjolfsson and colleagues, working with US Census Bureau data on tens of thousands of manufacturers. Firms that adopted AI saw productivity fall in the short run, by 1.33 percentage points on the raw comparison and by far more once the researchers corrected for the fact that optimistic firms adopt first. Over the following years the adopters outperformed their peers on both productivity and market share. The losses were concentrated in older, established firms, and management practices shaped who recovered. If you run a large, long-established organisation, you are the incumbent in that finding.

That data covers industrial AI between 2017 and 2021, before generative AI arrived, so read it as a pattern rather than a forecast. It is the same pattern Brynjolfsson and others have traced through earlier general purpose technologies: the investment in people, process and redesign comes first, and the measured return arrives later.

I call the bottom of that curve the Unlearning Valley. Teams are learning new tools while still delivering on old commitments, and the metrics dip. The leadership test is whether you fund the dip or retreat, and retreat is the instinctive move, because a quarterly reporting cycle reads a falling number as proof the investment isn't working. The leaders who get through it explain the curve to their boards before they enter it, and they measure different things while they're in it: how long it takes to get from a signal to a decision, and what share of experiments end up changing how the business runs.

I'm less certain than I'd like to be about how long the valley lasts in knowledge work. Nobody has good data on that yet, and anyone promising you two quarters is guessing.

Smarter, but none the wiser

Across twenty years of advisory work, in government, healthcare, education, utilities and financial services, the leaders who handle this well share one quality that has nothing to do with technical knowledge. They notice their own patterns. They catch themselves adding a governance layer when what the situation needed was permission to experiment.

That is metacognition, thinking about your own thinking, and I used to assume AI would make it easier. The research says otherwise. A team at Aalto University gave several hundred people a set of law school admission reasoning problems with ChatGPT's help. Their scores improved, but they overestimated those scores by more than they'd improved, and the most AI-literate participants were the worst judges of their own performance. The researchers titled the paper "AI makes you smarter but none the wiser".

That finding should worry every executive who describes themselves as an enthusiastic AI user, me included. Using AI and thinking with it are different activities. Thinking with it means turning it on your own reasoning: asking for the strongest case against the decision you've already made, then checking honestly whether you changed your mind or simply felt reassured. The leaders I see getting the most from AI have changed how they think, and they are building organisations designed to do the same.

So here is a question for your next executive meeting. Think of the last time someone in your organisation spotted something AI could change. How long did it take to get from that signal to a decision, and how many of the steps in between were there to learn something, and how many were there to grant permission?