Supervising AI: question the outputs you cannot produce yourself
The real promise of AI is that it lets you do work you could never do yourself. That is also its real management problem.
A consultant writing in Supercompanies describes spending five hours with Claude Code turning a workflow-mapping method into an agent. Claude designed data schemas and wrote code she says she could not have written and cannot fully understand. When it reported that it had "found an error we didn't notice before", her first reaction was "sounds good". An hour later she was stuck on a better thought: if that error only turned up by accident, how many more are in there that nobody has found?
She also makes a sharper point about the client meeting. She is the one who has to explain why one workflow scored higher than another, and "Claude told me" is not an answer.
Supervising someone who does a junior version of your job is easy. Supervising a specialist whose work you cannot do is hard. The usual fix is to learn enough of their craft to ask good questions, and that does not scale when the collaborator works across every discipline at once.
Her answer is simple. She cannot audit the schema, but she can interrogate the artifact. She had Claude build a visualisation of what it created and keeps it open, pushing on anything that does not sound right. How did this get scored this way? Why is this workflow ranked above that one? Where did that assumption come from? Each question surfaces a decision the machine made on her behalf. It is slow, and it does not give certainty.
The method is a familiar one: questioning an expert witness on the conclusion rather than the method. The organisations that retreat to using AI only for work they already understand will get modest efficiency and miss the main prize. The skill to build is disciplined scepticism about outputs you cannot produce yourself.
Source: Supercompanies, "What happens when we can't check AI's work?" (1 October 2026), https://lnkd.in/gVNsvFk4
#AIAdoption #Leadership #FutureOfWork