AI as a crutch or a coach: introducing AI so people get better
AI almost always lifts performance right away. The question is what happens to people's skills when you switch it off — and when AI is wrong in a convincing voice.
When a company introduces an AI tool, the first results are almost always good: reports get written faster, requests get resolved faster, documents get processed faster. One talk at this year's AI Summit in Barcelona asked a more uncomfortable question: are people getting better along the way — or just more dependent?
That question isn't academic. A company that, two years from now, has a team unable to spot a mistake without the tool carries a hidden risk that shows up in no report — until the tool gets something wrong, becomes more expensive or stops working.
Performance rises. Capability may not.
The difference is subtle but important. AI can work as a crutch or as a coach.
- Crutch: AI does the thinking. Performance rises, capability doesn't. Better with AI, worse without it than before.
- Coach: AI improves the practice. Performance rises, and so does capability. Better with AI — and later better without it, too.
This is not just theory. Research by Bastani and colleagues (2025) on high-school students practising maths found exactly this pattern. Students with unrestricted access to ChatGPT did markedly better while practising, but on the exam without AI they did worse than those who never used it. A version designed as a tutor — giving guidance instead of finished solutions — largely avoided that harm. The difference was not the model, but how it was used.
The mechanism is familiar to anyone who ever copied someone's homework. When you get a finished answer, you skip exactly the effort you learn from: the attempt, the mistake, the correction. The result on paper is better, and less stays in your head. Adults at work are no exception.
When AI is wrong, what the human knows matters most
Another widely cited study, Dell'Acqua and colleagues (2023), involved 758 consultants at BCG. On tasks within what the model does well, consultants using AI were faster and produced results of roughly 40% higher quality. But on a task just outside those limits — one that looked no different — those with AI were about 19 percentage points more likely to reach the wrong conclusion.
The authors call it the “jagged frontier”: a model is excellent at some tasks and weak at neighbouring ones, and the boundary is invisible from the outside. Two tasks can look exactly the same — the same kind of document, the same kind of question — and one lies within what the model does reliably while the other falls outside it. AI helps when it knows and hurts when it sounds right but is wrong.
The only defence is a person who understands the work well enough to notice when something is off. And here the circle closes: if AI has worked as a crutch for years, that is exactly the ability that has weakened.
Centaurs and cyborgs
The same study described two ways the most successful consultants worked with AI. Some were “centaurs”: they split the work clearly — this part I do, this part I hand to the model — and chose what went where. Others were “cyborgs”: they alternated with the model constantly within the same task, checking, asking for another version, correcting.
Both approaches share one thing: the person stays active. They accept a result not because it arrived, but because they checked it. That is a practical definition of AI as a coach.
Juniors are the most exposed
An experienced accountant who uses AI to suggest how to post an entry immediately sees when the suggestion makes no sense, because they have done that work a thousand times. A junior who started working alongside AI has no such reference. If the tool hands them finished answers from day one, they learn to click “confirm”, not to understand why something is right.
This is a serious question for any company: where will experienced people come from in five years, if AI now does the routine work people used to learn on? The answer is not to ban the tool for juniors, but to design it so that it teaches them — explains why, asks before it tells, shows them where they went wrong.
Same tool, two designs
What does the difference between a crutch and a coach look like in a real tool? Take an AI that assigns bank statement lines to ledger accounts — work that bookkeeping teams do every day.
Designed as a crutch: the tool assigns every line itself and shows a green “done” tick. The user doesn't see why anything went where it did, and confident and uncertain decisions look the same. The work is fast, but nobody can explain any more why a payment ended up in that account — and nobody notices when the tool gets the same type of payment wrong for weeks in a row.
Designed as a coach: the tool groups lines by confidence. Those that match a known pattern it posts with a short reason (“same payee and description as the previous 14 payments”). Uncertain ones it sets aside and first asks the user to choose the account, and only then shows its own suggestion. Every correction is logged, and once a month the team looks at the list of corrections: where the tool goes wrong, which rule needs extending, what a junior hasn't mastered yet.
The second design may be a second or two slower per line. But after six months the company has both a faster process and a team that understands it — and the list of corrections becomes the best training material for new hires.
Cognitive capital
The talk called this cognitive capital: the human capacity for judgment, expertise, critical thinking, creativity, adaptability and learning. Every workflow we bring AI into is also an environment in which people either develop or lose those capacities. So alongside “what can AI do?”, two more questions belong on the table:
- What should people still practise themselves?
- Where must human judgment remain accountable?
What this looks like in practice
In AI solutions for processing documents, data or requests, this translates into a few simple design rules:
- AI proposes, a human confirms — especially where mistakes have consequences: bookkeeping, pricing, contracts.
- Show the reasoning, not just the result. When AI extracts data from an invoice, the user sees where each value came from and can check it in a second.
- Flag uncertainty. A field the model read from a poor scan or an unusual format should look different from one it is sure about.
- Leave people the tasks that build expertise. Automate repetition, not judgment.
- Teach the team where the model is weak. A few examples where AI fails are worth more than ten where it shines.
- Measure capability, not just speed. Can the team still do the work when the tool is unavailable?
None of these rules noticeably slows the work down. A check that takes two seconds because the reasoning is right there is not a cost — it is the reason people still understand what is going on.
Questions for management
Before rolling an AI tool out to the whole team, it is worth answering a few questions:
- Would the team notice if the tool made a mistake that looked convincing?
- Which part of the work do people no longer do themselves — and was that a conscious decision, or did it just happen?
- How do juniors on the team learn the work the tool now does for them?
- What would happen to deadlines if the tool were unavailable for a week?
The closing message of the talk was short: verify, override, take responsibility. AI that leaves people more capable is worth more than AI that only makes them faster.