What changes about managing a team that uses AI?

The first change is the loss of a signal managers relied on without noticing. A well-structured document, a competent analysis or a clean piece of work used to be evidence that the person understood the problem, because producing it required understanding it. That inference no longer holds, and continuing to make it means assessing people on an output that is now cheap. What still distinguishes is the ability to explain the reasoning, defend a choice, and say what was considered and rejected, which is a conversation rather than an artefact.

The second is where the bottleneck sits. When production was the constraint, adding capacity meant adding producers. When production is fast and review is not, adding producers makes the queue longer. Teams that adopt these tools without addressing review find that throughput does not improve as expected and that quality drifts, because the review that used to be sufficient is now spread across more work. The uncomfortable resolution is that the arrival rate has to match the review capacity rather than the production capacity.

The third is the training problem. The tasks traditionally given to less experienced people — first drafts, initial research, the tedious groundwork — are exactly the tasks now most easily delegated, and they were the mechanism by which judgement developed. A team that removes them gains throughput and loses its pipeline, and the consequence arrives years later when the people who would have been reviewing have not built the judgement to do it. This has no cheap solution and it needs to be a deliberate decision rather than a default.

There is a fourth change that is about consistency. When everyone uses different tools, different prompts and different personal settings, the same task performed by two people produces genuinely different work, and the difference is invisible because it lives in configurations nobody can see. Shared standing context and shared templates address this, and the reason to bother is that unexplained inconsistency between people is expensive to manage.

The measurement question changes too. Output volume was a reasonable proxy for effort and is now a proxy for nothing, so any management practice resting on it needs revisiting. What remains meaningful is outcomes, the quality of judgement in hard cases, and whether the work someone produces holds up when examined. Those are harder to measure and were always the things that mattered.

Finally, the practical management task is different from the technology question. What people need is permission to use these tools openly, clarity about what may not go into them, and an expectation that they can explain what they produced. Those three are ordinary management and they resolve most of what appears to be a technology problem.