Who decides when AI gives its opinion?

AI helps you decide better, as long as the decision stays yours. In an Anthropic study with 61 recruited analysts, 50 experienced ones used an AI agent that argued, assigned at random, for one of two options in four choices in a fictional case of an agricultural cooperative. They followed the agent in 86% of the choices.

And almost nobody noticed. Afterwards, 77% explained their choices with the agent's arguments, sometimes word for word, and only 15% mentioned the AI. In a company, this applies to pricing, credit or purchasing: the report goes out with the manager's signature, but the criterion may have come from the machine.

The way forward is to write down your choice before opening the conversation with the AI, and ask it for the best argument for each side, not just for its own. A decision that involves money or customers deserves this rule written into the process, saying who decides and at what point.

It costs a few minutes per decision, and the study has limits: a fictional case, a single model and a deliberately reinforced preference. Without the agent, 6 of 11 made serious errors, against 3 of 50 with the agent, in groups that are not comparable. Those who used the agent made fewer mistakes and were faster, but the group without AI was small and less experienced, so this doesn't prove it was the AI. Which decision this week was the manager's, and which was the tool's?

This text was produced with the support of AI.

Sources

Anthropic, 27 Oct 2026. https://www.anthropic.com/research/agents-judgment-calls