Human–AI Work
What must people still understand when AI does more of the work?
As AI takes on more professional work, leaders need to consider how people acquire the expertise to assess it.
Download PDFConsider a team reviewing an AI-assisted market assessment. The recommendation is persuasive and the figures appear consistent. Then someone asks whether customer interest would survive a higher price, or whether the projected margin includes the cost of serving the market. The document offers an answer. The person presenting it cannot explain how much confidence to place in it.
This can happen even when the analysis is broadly correct. A better output can obscure what its author has understood. As organisations introduce AI, they need to think about the work people will be able to do next year as well as the work they can complete today.
The benefit and the learning are different
A September 2026 working paper by David Autor and colleagues examines this distinction directly. In a three-month randomised trial involving 133 patent lawyers at eleven US firms, access to an AI drafting assistant improved the quality of their work. When participants later reviewed a patent application without AI, the improvement was concentrated among senior lawyers. Junior lawyers showed no average gain on that assessment, although individual results varied.1
The study offers evidence that AI-assisted practice can strengthen judgment. It also raises a question about who is equipped to learn from it. These findings come from one profession and a particular tool; they do not settle what happens across knowledge work. They do, however, give leaders a reason to distinguish improvements in the work delivered from improvements in the person’s understanding.
A smaller experiment reported by Anthropic earlier in 2026 examined developers learning an unfamiliar programming library. Those with AI assistance scored lower on an immediate comprehension test than those working without it: 50% against 67%. There was no statistically significant difference in completion time.2 Again, the finding concerns a particular learning task, not inevitable long-term deskilling.
For an organisation, the practical implication is to examine what employees are learning through AI use. The answer may differ between someone already fluent in the field and someone who is still acquiring its basic concepts. Treating both as equally capable because their finished work looks similar risks missing that difference.
What people need to understand
Reviewing an output requires more than knowing that AI can make mistakes. People need a working grasp of the subject, a sense of what good evidence looks like and an understanding of the consequences of being wrong. Without that foundation, an instruction to “check the answer” asks for a capability the person may not yet have.
In the market assessment, that means recognising that expressed interest is weaker evidence than a purchase, that a margin estimate depends on how costs are allocated, and that a finding from one customer group may not apply to another. It also means knowing when to bring in someone with deeper expertise. These are capabilities developed through experience with customers, analysis and colleagues, not simply through familiarity with an AI interface.
Develop expertise in the new work
Leaders need not preserve every old task for its educational value. Some junior work was repetitive without being particularly instructive. The useful question is which experiences helped people learn, and where those experiences will now occur.
Take the preparation of a client recommendation. A colleague might once have learned by building the analysis, discussing an awkward finding with a manager and revising the argument. If AI prepares the first draft, the learning can move into testing its assumptions, comparing alternative explanations and seeing how the recommendation holds up with the client. That requires more than adding a review step: the colleague needs to do enough of the reasoning to learn from the discussion.
A manager could ask for an initial view before the AI-assisted analysis, then discuss what changed and why. On selected pieces of work, a colleague could explain a rejected alternative or identify the evidence that would reverse the recommendation. AI can help generate alternatives and explanations; an experienced person can test whether those explanations fit the business. Such conversations let a manager see how a colleague is reasoning and where further experience would help.
Give managers room to teach
Expectations about speed matter here. If the time released by AI is immediately absorbed into higher delivery targets, there may be little room to discuss why a recommendation was weak or how an experienced colleague spotted the problem. Organisations should make a deliberate choice about how much of that capacity goes into developing their people.
This also changes what a good review looks like. Correcting a document may improve today’s deliverable. Explaining the correction, asking the colleague to reconsider the argument and following up on the result can help them handle the next case. Leaders should look for that development alongside the quality and speed of AI-assisted work.
The aim is a team that can take on more demanding work with AI and exercise sound judgment about the result. A useful test is whether colleagues are becoming better at explaining their reasoning, recognising uncertainty and knowing when to seek help. Those are signs of expertise that a polished document alone cannot provide.
Sources
1. Autor et al. (2026). Does AI assistance enhance or erode expertise? NBER working paper 35720. ↗
2. Anthropic (2026). How AI assistance impacts the formation of coding skills. ↗