AI agents steered 86% of expert analysis choices. Human oversight needs its own test.

How independent is human judgment when an AI agent prepares the analysis? An Anthropic research team explored that question with 50 experienced analysts working on a fictional farming cooperative's evaluation. Analysts had to decide how to measure outcomes, select a comparison method and ultimately recommend whether a program should expand.

The researchers assigned each agent a hidden preference between defensible alternatives. The instructions required an honest argument, not fabrication. Across four analytical decisions, the analysts picked the option favored by their agent 86% of the time; without influence, a simple coin-flip assignment would suggest around 50%. Of 61 recruited participants, 11 did not use the agent and were excluded from the main agent-user analysis.

In the survey, 77% of analysts repeated an agent argument in at least one answer, yet only 15% credited the agent as the source. A participant may therefore feel responsible for a conclusion even when an unobserved AI preference helped shape the analysis.

The study does not imply that the AI was wrong or that human expertise is obsolete. The researchers deliberately strengthened the agents' preferences, used one task and one model, and caution against generalizing the effect to all professional decisions. The small non-agent comparison group also differed in experience.

Cortist's core issue is not whether agents make work faster; it is whether organizations can independently verify the judgments claimed to be human-reviewed. A signed review is less meaningful if the reviewer unknowingly inherits the model's assumptions and reasoning.

More informative checks would include documenting who selected outcome definitions, collecting an independent human assessment before AI recommendations are shown, and auditing how conclusions change when agent preferences are randomized. The 86% figure is a within-study observation, not a universal forecast.