This week I spoke with a large company about AI adoption. The conversation started with the tools and ended, like almost every one this year, with the people.
Last Friday Ethan Mollick put a name to what was being discussed at that table. In The Overhang he uses the term for the gap between what AI models can already do and what almost nobody is doing with them. It is not a problem for the next release. It is capacity most companies have already paid for and still do not use.
The licence is not the adoption
The pattern repeats with few variations. A licence is bought for the whole workforce. A few months later, a handful of people generate almost all the usage, nearly always the most junior. And the experts who hold the business up, the ones who know where the expensive mistakes are, stop logging in.
The data points the same way. In a University of Konstanz survey of 1,105 employees, published last week, AI use at work barely moves from 35% to 38% in a year. And 45% of those who use it mostly do so with a tool their employer did not introduce. Adoption is happening, but under the radar and without shared judgement.
Why it stalls
It is almost never resistance to technology. It is usually three more prosaic things.
The purchase is budgeted and the redesign is taken for granted. Nobody decides which work changes, who reviews what the machine produces or who answers when it gets it wrong. Without those three answers everyone improvises their own, and most decide not to take the risk.
The people who know most use it least. And they are exactly the ones who can tell a good result from one that only sounds good. Mollick sums up what people bring in four advantages: deep knowledge, broad knowledge, the judgement to choose, and initiative. A company where AI is used only by those without the first two produces more, but gets things wrong faster.
The temptation to turn AI into a cut. The numbers always add up: so many roles, so much saved. What the numbers leave out is that the first roles to go are entry-level ones, and that is where the experts of five years from now were being trained. The company gains productivity this year and in 2030 is left with nobody able to review the machine's work.
Speed works against training courses
All of this would be manageable if the technology stood still. It does not. Models change every few months, and what they could not do in January they do in June. A course on which buttons to press expires before its second run.
What does not expire is knowing where the frontier is. The Harvard and BCG study that coined the idea of the jagged technological frontier measured it with 758 consultants. On the tasks AI does well, those using it completed 12.2% more work, 25.1% faster and at more than 40% higher quality. On a task that fell outside that frontier, those using AI were 19 points less likely to get it right than those who were not. And the frontier is not visible to the naked eye: two tasks that look equally hard can fall on different sides of it.
That frontier also moves every quarter. That is why what needs training is not the handling of a particular tool, but the decision, task by task, of what gets delegated, what gets reviewed and what is still done by hand.
Training the workforce is no longer optional, and it is not a prompting course
Since February 2025 the EU AI Act has required companies that use AI to attend to the AI literacy of their staff. The reform passed this summer lowered the bar (there is no longer a level to guarantee, only measures to take to develop it), but kept the obligation. The question is no longer whether to train, but how to do it so that it works.
And here is the most common mistake: treating AI training as software training. A tutorial teaches where the button is. It does not teach you to spot the plausible error in a contract, to decide whether an agent may touch customer data, or to defend before a committee why the junior profiles are not being cut. That is judgement, and judgement is trained by deciding with consequences, not by listening.
How we can help: training judgement with simulations
At Kudzu Partners we have spent ten years building business simulations, and this year we opened a full line on AI adoption. The idea is simple: put a leadership team in front of a company that is not theirs, with a problem that is, and let them decide.
The latest is The Overhang, built on Mollick's essay. A 142-person software company has had an AI licence for its whole workforce since January. 68% of the licences are used less than once a week, 11 people generate 71% of the usage, and the six experts who hold the business up have not logged in since March. Meanwhile, a 31-person competitor takes a client worth 1.5 million from them, and the investor proposes cutting 30% of the workforce. In two hours, the team has to find the idle capacity, redesign the work, choose between AI-generated results and defend its plan before the board. The capacity is already paid for: the problem is one of distribution and judgement, not licences.
Around it there are half a dozen more in the Eureka Express catalogue. The Jagged Frontier asks you to decide which tasks go to AI, which to people and which to both. SkillShield and JudgmentForge set out, five years ahead, the dilemma between productivity today and experts tomorrow. And in AI Adoption in Team you have to win eight sceptical people over to a 60-day adoption campaign.
For executive audiences we also have four simulations co-developed with IESE faculty: Big Bank, on rolling out AI in a bank with the FATE framework (fairness, accountability, transparency and ethics), now in an agentic version; GPT Workshop; Critical Care, Critical Data; and Change Star, on how to lead the AI transformation in the organisation.
And if your company's problem is not in the catalogue, that is fine: a new microsimulation, built on your case, is ready in under a week from start to finish. We say more about how we work on our Applied AI page.
Our proposal is to start small: a pilot with one group, one simulation and one concrete question. By the end, each participant has taken decisions that would take them months to take in their real job, and the company has something it almost never has: data on how its people decide when AI is in the room.
The AI you have bought already works. What is missing is people who know when to trust it, and that does not come with another licence: it is trained.
If you want to try The Overhang with your team, or tell us about a problem you cannot find in any catalogue, get in touch.
