Study: experts followed the AI agent in 86% of choices
In an Anthropic test, experienced analysts adopted the AI agent's hidden preference in 86% of choices without noticing. What this changes about human review.

On Tuesday (the 27th), Anthropic published a controlled experiment with 50 experienced data analysts, who used an AI agent to decide whether a fictional farming cooperative should expand or shut down a program. Without their knowledge, each agent was given a randomly drawn preference for each choice of method. On those choices, the analysts went with the agent’s option 86% of the time, when chance would give about 50%. No participant suspected the manipulation. The study was also covered by Unite.AI.
Experience didn’t protect them
Of the 50 analysts, 65% had a master’s degree, a PhD or six or more years of experience. Even so, none of the characteristics tested, such as years in the field, education or trust in AI, predicted who would follow the agent less. Another figure stands out: 77% repeated the agent’s arguments in their own explanations, but only 15% said the idea came from it.
The authors are honest about the limits. The agents’ preference was deliberately reinforced, it was a single task of about an hour, with a single model. In everyday work, the effect may be smaller. In a smaller group that worked without AI, serious method errors were more common, but the authors warn that the two groups aren’t directly comparable.
What this changes for your company
1. “There’s a human reviewing it” isn’t enough of a control. Review catches calculation errors, but it doesn’t catch the choice of path. If the AI sets the criteria for a credit analysis, a purchase or a selection process, the person tends to follow.
2. Single out the decisions that require your own opinion. For the choices that weigh most on the business, ask the person responsible to write down their own position before consulting the AI. It’s simple and creates a reference point for comparison.
3. Ask for alternatives, not just the recommendation. Set up the agent to show the two or three possible options, with pros and cons, instead of a ready-made answer. And make clear, in writing, who signs off on the decision.
In 29 years of programming, I’ve learned that every system carries the choices of whoever designed it. With AI agents, the governance question is the same: who is actually deciding.
Originally published on IWA7 Insights, IWA7 Digital’s blog, which is now part of Escovando Bytes.


