A new Anthropic study found experts followed an AI's hidden lean 86% of the time. What it means for your team, and five habits that keep decisions yours.
A study published by Anthropic on 27 October 2026 found that professional analysts working with an AI agent went along with the agent's preferred option 86% of the time, even though that preference had been set by a coin flip. None of them suspected it. As teams across Singapore and Asia hand more analysis to Claude, ChatGPT, Microsoft Copilot and Google Gemini, this is the AI risk that checking for errors will not catch.
Anthropic's researchers asked 50 experienced analysts to use an AI agent (Claude Code) to advise a fictional farming cooperative on whether to expand an advisory programme. Along the way, each analyst had to make four judgment calls, such as which outcome to measure and how to weight large versus small farms. Each call had two reasonable answers, with no single right one.
Without telling the analysts, the researchers used a coin flip to decide which option each person's agent would lean towards. The agent was told to argue its lean honestly and never misrepresent the evidence. Then they watched what the humans chose.
| Finding | Result |
|---|---|
| Analysts who chose the option the AI leaned towards | 86% of choices (about 50% would be expected if the AI had no influence) |
| Analysts who already had a view and changed it when the AI disagreed | 60%, versus 6% when the AI agreed with them |
| Analysts who repeated the AI's arguments in their own explanations | 77%, sometimes word for word |
| Analysts who credited the AI for those arguments | 15% |
| Analysts who suspected the AI's preferences had been set for the study | None |
Experience, education and how much people said they trusted AI did not predict who deferred. Senior and junior analysts behaved much the same. About half still felt the final recommendation was more their own work than the agent's.
The researchers are careful about the limits: it was one short task, in one field, with one model, and the AI's leanings were deliberately made stronger than usual. Real-world effects may be smaller. But the direction is clear, and it matches decades of research on what psychologists call automation bias.
Automation bias is the tendency to accept a suggestion from an automated system more readily than we would accept the same suggestion from a person, and to look less hard for reasons it might be wrong. It has been documented with aircraft autopilots, medical decision tools and GPS navigation. Generative AI makes it stronger for three reasons:
An AI hallucination is a factual error: a figure, quote or rule that is simply wrong. You can catch it by checking sources. Judgment calls are harder. Choosing which supplier to shortlist, which metric to report to the board or which candidate to interview has no single correct answer, so there is nothing to fact-check. The output can be accurate, well argued and still reflect the AI's preferences rather than your team's priorities.
That also means the usual safeguard, "a human reviews it before it goes out", is weaker than it looks. If the reviewer simply adopts the AI's framing, the sign-off is real but the judgment is not fully theirs.
Most business uses of AI include small decisions that shape the result:
| Team | AI-assisted task | The judgment call inside it |
|---|---|---|
| HR | Screening CVs against a job description | Which experience counts as relevant, which gaps matter |
| Sales | Prioritising leads or accounts | What "high potential" means for your business this quarter |
| Finance | Building a forecast or variance analysis | Which assumptions, time period and comparison base to use |
| Marketing | Analysing campaign results | Which metric defines success: clicks, leads or revenue |
| Operations and procurement | Comparing suppliers or quotes | How to weigh price against reliability, location and risk |
| Management | Summarising options for a decision paper | Which options are presented, in what order, with what emphasis |
None of these are reasons to stop using AI. They are places to slow down and make the choice deliberately.
Our business prompt library includes ready-made prompts for balanced comparisons and assumption checks that you can adapt.
The same study found that analysts working with the AI finished faster (a median of 65 minutes versus 90) and made far fewer serious methodological errors (6% of analyses versus 55% without AI). The researchers note the no-AI group was small and less experienced, so the comparison is only indicative, but the message is clear: used well, AI makes teams faster and more thorough. The goal is to get that benefit while keeping the decisions that matter in human hands.
Check how ready your team is with our free AI readiness quiz. For hands-on practice, our AI literacy workshop covers automation bias and safe habits for every employee, and our responsible AI and governance training helps managers set the rules for human oversight.
It is the tendency to accept suggestions from an automated system, including AI tools, more readily than we would from a person, and to look less hard for reasons they might be wrong.
Analysts using an AI agent chose the option the agent leaned towards 86% of the time, even though the lean was set by a coin flip, and none of them suspected it.
No. The same study found AI-assisted analysts were faster and made fewer serious errors. The fix is to make judgment calls deliberately, not to avoid AI.
Write down your own view before asking the AI, then compare. If you often switch to the AI's position without a new fact to justify it, slow down and ask for the best case for each option.
This article is general information for business readers, not legal, financial or technical advice. AI tools and plans change quickly; check current vendor documentation before making decisions.
Tell us about your team and goals. We'll reply with a tailored training plan and quote.
Get a Quote