Articles · 17 November 2026 · Lucy Pitt
AI Is Doing the Work Your Future Leaders Learned On
Judgement is built in the struggle. So what happens when we automate the struggle?

Think back to the first proper job you had, and the work nobody wanted. The first drafts that came back covered in red pen, the spreadsheet you rebuilt three times because the numbers would not reconcile, the case notes you wrote up while someone who had done it a thousand times read over your shoulder and said “yes, but why?” None of it was glamorous, and most of it was slow, but it is almost certainly where you learned to tell a good answer from a plausible one.
That unglamorous layer of work is exactly the layer AI is best at. It drafts, summarises, reconciles and tidies faster than any graduate, and it does not need a coffee break. So organisations are, quite reasonably, handing it over. The question very few of them are asking is what that layer was quietly doing for their people while it was being done by hand.
Having watched a lot of AI rollouts up close, our answer is that it was building judgement. And the research arriving this year suggests we are right to be worried about what happens when it goes.
Better work is not the same as better people
The most useful piece of evidence comes from an unlikely place: patent law. A team led by the MIT economist David Autor ran a three-month randomised trial with 133 patent lawyers across eleven US firms. Some were given an AI drafting tool and some were not. Unsurprisingly, the AI group produced better drafts, by about a third of a standard deviation, mostly because they turned in fewer weak ones. The gains looked largest for the juniors, which is exactly the result every vendor deck would put on slide two.
Then came the interesting bit. At the end of the three months, everyone was asked to review and correct a patent application without AI. The senior lawyers who had been using the tool did better than their peers. The juniors who had been using it did no better on average than juniors who had never touched it. Their results simply spread out, with fewer middling scores and more at both ends.
The authors’ conclusion is the sentence I would like every L&D team to pin up this month: “Foundational expertise may be a prerequisite for extracting durable skill from AI-assisted practice.” In plain English, people who already knew the route could use AI as a shortcut and came back sharper. People who were handed the shortcut before they knew the route got to the destination, but did not learn the way.
A note of caution: it is a working paper not yet peer reviewed, most of the authors work at Google, which built the tool, and it covers one profession in one country. But it matches what we see in the room, and a second study backs it up.
The 48 per cent that turned into minus 17
Researchers from Wharton, publishing in PNAS last year, worked with nearly 1,000 high school maths students in Turkey. One group had unrestricted access to a standard AI chatbot during practice. Their practice scores rose 48 per cent against students working without it, which would delight any headteacher until the exam. When the AI was taken away, the same students scored 17 per cent below the control group. They had not just failed to gain; they had gone backwards.
A third group used a version of the same tool designed as a tutor, which gave hints rather than answers and nudged them to do the thinking themselves. Their practice scores rose even more, and when the tool was removed, they performed about the same as students who had never used it. The harm had largely disappeared.
Same technology, opposite outcomes. The difference was not the model; it was the design of the learning around it. If you only take one thing from this article, take that, because it moves the conversation away from “should our people use AI?”, which is settled, and towards “what are we asking it to do for them, and what are we still asking them to do for themselves?”
Critical thinking is quietly changing shape
Lawyers and teenagers may not be your workforce, so here is a third piece that is about all of us. Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers who use generative AI at least weekly. They found that the more confidence people had in the AI, the less critical thinking they reported applying to its output. The more confidence people had in their own ability to do the task, the more they checked, questioned and pushed back.
They also noticed that critical thinking does not vanish so much as move. Effort shifts from gathering information to verifying it, and from solving the problem to stewarding someone else’s solution. That is a perfectly good shift for an experienced professional who knows what good looks like. It is a much shakier one for somebody in their first two years, who has been asked to check work they have never actually done.
Put the three studies side by side and the pattern is hard to miss. AI raises the floor of output almost immediately, and for experienced people it can raise the ceiling too. For people still building their expertise, it can produce work that looks expert without producing the expert. As the barrister Dr Peter Fields put it recently in Business Reporter, “AI can help someone produce an answer without helping them understand it.”
Why this is a leadership problem, not a training one
Here is where it stops being an academic point. Most organisations we work with have a talent pipeline that depends on juniors becoming seniors. In housing, the surveyor or the income officer learns by doing hundreds of cases. In professional services, the trainee learns by drafting what the partner rewrites. In construction, the site manager learns by getting the programme wrong in small ways before they are trusted with large ones. Judgement is not taught in a module. It is built in that loop of attempt, feedback and correction, repeated until the person can feel when something is off before they can explain why.
If AI takes over the attempt and the feedback goes to the tool rather than the person, the loop breaks without anyone noticing. Output stays high, possibly higher. The dashboards look lovely. And five years from now you discover that your next layer of managers can produce a beautiful report but cannot tell you whether it is right, which is a little like discovering your sat nav has been doing the driving and nobody in the car can read a map.
None of this argues for banning AI from junior roles or keeping inefficient tasks out of nostalgia. It argues for being deliberate about where the thinking happens.
What we would do on Monday
Drawing on the research and on what we see working in our own Champions programmes, four things are worth doing now.
First, try “human first, AI second” on the work that builds judgement. Fields suggests asking trainees to form their own view, compare it with the AI’s, investigate where they disagree and explain which answer they trust and why. It takes a little longer, which is the point, because the comparison is where the learning happens.
Second, test understanding separately from output. If the only thing you assess is the finished document, you will reward the people who are best at prompting rather than the people who are best at thinking. Ask juniors to talk through a piece of work without the tool in front of them now and again.
Third, protect the time AI saves and spend some of it on supervision. The temptation is to bank every saved hour as more throughput. Reinvest part of it in the conversations where an experienced colleague asks “yes, but why?”, because that question is the cheapest leadership development programme you will ever run.
Fourth, make it safe to challenge the machine. Critical thinking needs a climate where a junior can say “hang on, is this right?” about an AI output, or about a senior’s AI output, without looking slow or awkward. In our PRISM work this is the Climate lever, and it is often the one where adoption quietly stalls.
The question to take into 2027
AI is going to do more of the foundational work, and it should. The organisations that come out ahead will not be the ones with the most licences or the fastest drafts. They will be the ones that noticed what the foundational work was teaching, and found a deliberate way to keep teaching it.
So before your next AI rollout decision, ask one simple question of every role it touches: where in this job does someone learn to tell a good answer from a plausible one, and what will we do to protect it?
If you would like to talk about how this looks in your organisation, contact us to discuss your organisation’s PRISM read.
Sources
Autor, D. et al. (2026) Does AI Assistance Enhance or Erode Expertise? Evidence from a Three-Month Field Experiment in Patent Drafting. NBER Working Paper 35720. Link
Bastani, H. et al. (2025) Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122. Link
Lee, H.-P. et al. (2025) The Impact of Generative AI on Critical Thinking. CHI 2025, Microsoft Research and Carnegie Mellon University.
Fields, P. (2026) How businesses can develop junior talent when AI takes over foundational tasks. Business Reporter. Link
Where does your organisation actually sit?
PRISM answers with evidence rather than opinion.
Explore PRISM