On October 27, 2026, Anthropic published a randomized controlled trial on how AI agents steer experts’ analytical choices. Fifty professional analysts with causal-inference experience used Claude Code running Claude Opus 4.8 to decide whether a fictional farming cooperative should expand or discontinue an advisory program. Four decisions (outcome measure, comparison method, farm coverage, weighting) each had two reasonable options, and a coin flip secretly set which option the participant’s agent favored.
Analysts matched the agent’s lean 86 percent of the time, against about 50 percent if the agent had no influence (92 percent on the outcome measure, 82 on comparison method, 81 on farm coverage, 90 on weighting). When the agent leaned toward rollout, 54 percent recommended rollout, versus 36 percent when it leaned the other way. The four leans alone predicted 82 percent of final recommendations. Analysts who had already picked an option switched 60 percent of the time when the agent disagreed, versus 6 percent when it agreed, though that rests on 38 cases.
Seventy-seven percent of analysts reused the agent’s arguments in their explanations, but only 15 percent credited the agent, and experience did not predict who deferred. Agent users also made fewer serious methodological errors (3 of 50) than a small no-AI group (6 of 11) and finished faster (65 versus 90 minutes median).
Why it matters: the study puts a number on a quiet risk of agentic tools - an agent’s defaults can decide contested analytical choices while the human still believes the conclusion is their own.
What it does not show: the researchers deliberately strengthened the agents’ leans, so real-world effects may be smaller; it covers one roughly one-hour task, one field and one model; participants were new to farming; and the AI and no-AI groups differed in experience, so the error and time gaps cannot be attributed to AI alone.