Keeping your experts is not enough

A study in The Lancet Gastroenterology & Hepatology followed experienced endoscopists, people with two thousand procedures or more behind them, after their clinics brought in AI detection tools. Within a few months, their detection rate on procedures done without the AI had dropped from 28.4% to 22.4%. Nobody lost a job and nobody did anything wrong. These were exactly the people we would all point to as the safe pair of hands, and the skill faded anyway. That is what I want to think about this week.

A few of the things I read last week:

• Experienced endoscopists who worked regularly with AI detection saw their unassisted detection rate fall from 28.4% to 22.4% within months (Lancet Gastroenterology & Hepatology).
• Ford put 900 AI cameras on its quality lines, then spent three years rehiring around 350 veteran engineers to train them, and topped JD Power's quality ranking for the first time since 2010 (Bloomberg / BBC).
• 57% of senior leaders say AI is already embedded in their core processes, while only 23% call their workforce ready, down six points in a year (Kyndryl 2026 People Readiness Report).
• Entry-level roles in the most AI-exposed occupations are shrinking 3.8% a year for 22-to-25-year-olds while the same occupations grow for people in their late thirties, so the experienced are carrying more of the work, not less (Stanford Digital Economy Lab & ADP).
• Klarna, which replaced around 700 customer-service staff with an AI assistant in 2023, was rehiring people by 2025 after quality dropped on the complex, sensitive cases (Forbes).

Almost every AI plan I have seen assumes the experienced people are a constant. I do not think we can assume that any more.

The experts did not notice it happening

I do not think the endoscopists in that study were careless. They did what all of us are being asked to do: adopt the tool and work with it every day. The tool watched the screen for them, and the part of them that used to do the watching got less practice. What unsettles me is not the size of the drop. It is that nothing about it was visible from the inside. Their assisted numbers stayed good. If their clinic ran a dashboard, every line on it was green.

The Kyndryl report reads differently once you hold it next to that. We can say to the decimal how embedded AI is in our processes, and leaders' confidence in their own workforce is falling while the embedding deepens. But neither number tells us what is happening to the skill inside the people we already trust. When we plan, we treat experience as a stock, something the organisation owns and keeps. I think it behaves more like fitness. It stays because it is used, and the daily use is exactly what the tool takes over. Meanwhile the Stanford numbers show us leaning harder on experienced people than ever, at the same moment their unassisted practice is quietly thinning out.

Aviation dealt with this on purpose

Aviation met this problem decades before us. Autopilot flies more smoothly than most humans, most of the time, and pilots stopped hand-flying. In 2013 the FAA put out a safety alert saying plainly that autoflight was eroding manual flying skills, and told airlines to build hand-flying back into ordinary operations. Not because autopilot was bad, but because systems fail, and meet situations they were never trained on, and the skill you need in that moment has to have been practised before it.

There is a fair argument against all of this, and I want to give it room. If the tool is always there, and performance with the tool is higher, why should we care that unaided performance slips? Nobody mourns our mental arithmetic since the calculator. And last week's headlines even said AI is not taking jobs after all; the Yale Budget Lab finds no real change in unemployment in the most exposed occupations. I think both things are true at once. The jobs stay, and the skills inside them can still fade, and nothing in a calm unemployment figure would show it. As for the calculator: Ford is the answer I would give. The cameras only worked once veteran engineers trained them, and their hardware chief said it without decoration: "It's only as good as the information you use to train it." Klarna found the same at the other end, when the complex, sensitive cases came back needing people. The tool does not remove the need for the skill. It moves the skill upstream, to training and checking the tool, and that skill has to live in someone.

If you work in L&D, HR, or transformation

We are good at adoption dashboards, and nobody has ever asked me for a skill-retention dashboard. I understand why; it is a strange thing to measure and an awkward thing to raise. But two things seem doable this week. First, find one place where your experienced people now work only through an AI tool, and ask when they last did that work unassisted. The endoscopy study says this is maintenance, not nostalgia. Second, pick one workflow and schedule regular unassisted practice into it, the way airlines schedule hand-flying: expected, routine, nothing heroic. And when we write readiness plans, let us include the people we currently call ready.

The provocation

If every AI tool in your organisation went quiet for a week, which skill would you discover had gone with it? Pick the one that worries you most, and find out this week when someone last practised it without the tool. That answer costs one conversation, and it might be the most useful capability number you collect this quarter.