Analysts follow AI agents’ recommendations in four decisions

Anthropic reported that analysts followed their AI agents’ recommendations in 86% of four analytical decisions. The company said the rate would have been about 50% if the agents’ input made no difference, indicating that analysts often aligned their choices with an agent’s lean in the study.

The pattern was also evident when analysts had stated a preference before receiving the agent’s input. When the agent disagreed, analysts changed their initial preference 60% of the time. When the agent supported their preference, they changed course 6% of the time.

The figures describe two different conditions: an agent challenging an existing view and an agent agreeing with it. Anthropic said each of the four decisions could shift the estimated productivity gain by several percentage points. That mattered in relation to a board threshold of 3–4%, the company said.

The account links the decisions’ potential effect to that threshold but does not state whether the board ultimately accepted or rejected a productivity estimate. The study also asked analysts to rate how confident they were in their judgments, using a scale from 0 to 100.

Anthropic described 0 as equivalent to a coin flip and 100 as a level of certainty at which someone would stake their professional reputation. These endpoints indicate how the scale was framed; the account does not give an average confidence score or report how confidence changed after analysts received an agent’s input.

In an exploratory analysis, Anthropic examined whether 10 characteristics predicted how often analysts followed their agents’ recommendations. The characteristics included experience, self-rated expertise in causal inference, education, performance on a causal-inference quiz, trust in AI, frequency of AI use, enjoyment of effortful thinking, open-mindedness, gender and age.

The account describes the analysis as exploratory and does not identify any characteristic as a predictor. The reported results center on analysts’ choices and the possible effect of those choices on a productivity estimate. They do not state that an AI agent made the decisions independently: the measured outcome was whether analysts adopted the agent’s lean.

The study’s reported percentages therefore concern analysts’ decision patterns, while the board threshold provides the stated benchmark for the estimated productivity gain.