The Monthly ReadingReading noteNovember 2026
When the agent leans, the reviewer follows
Chad Neptune, CEO and Chief AI Architect
4 min read
Reading note on: How AI agents shape experts’ judgment calls, by Amaç Herdağdelen, Hailey Joren and Alex Tamkin, Anthropic,
The approval step most companies trust may be weaker than they think. A person reads the agent’s work, agrees and signs, and the log says a human decided. Anthropic researchers have now measured how far expert analysts follow an agent’s lean.
On four method choices, the analysts sided with the agent’s lean in 86% of cases
Anthropic researchers recruited 61 professional analysts to advise a fictional farming co-op. Eleven opted out of using the agent, so 50 worked with one. Each agent had been told in secret, by coin flip, which way to lean, and to argue its case honestly. On four method choices the analysts sided with the agent’s lean in 86% of cases, where about 50% would be expected if the agent made no difference. On the final recommendation the pull was weaker: 54% recommended rolling the program out when the agent leaned that way, against 36% when the lean pointed the other way.
A signature counts only if the reviewer’s own call is on record
Resolon’s standing position is that a named person approves the decisions that carry risk. This study suggests that sentence is incomplete. A reviewer who reads the agent’s case before forming a view isn’t checking the agent. They are finishing its work. So the question to ask is whose judgment the signature contains. The authors say it themselves: a human reviewer is only an independent check if their judgment is their own
(Anthropic).
Writing the reviewer’s call down first is one fix, and the study didn’t test it. Its closest evidence is discouraging. Among 38 cases, an analyst had already named a choice before the agent showed its lean; they switched away 60% of the time when it disagreed, against 6% when it agreed. That is a 54-point gap, which the authors call imprecise: the plausible range runs from 24 to 82 points, and analysts who stated a preference may have differed from those who didn’t. The first call belongs on paper. A view held only in someone’s head can’t be counted or audited.
For a CEO or a PE owner, the audit trail can overstate the judgment applied
In the survey, 52% of analysts said the final recommendation was theirs more than the agent’s, yet their choices had moved about as far toward its lean as those who judged it more the agent’s: about 65 percentage points against 79, within the margin of error. Anthropic’s 2026 usage policy update, in effect from 12 November, still requires two things when Claude’s output can affect a person’s health, money, legal position, job or access to essential services: a qualified person with the authority to review its recommendations and change them if necessary, and notice to the affected person that AI was used. That reviewer may be one who thinks the call was more theirs than the agent’s. A credit exception can pass through an agent before a person and still carry a human signature while the agent sets the direction.
Where I could be wrong
The analysts knew causal inference well and farming little. The authors note that in their own field they might know more and hold firmer views, which could shrink the agent’s pull. The authors also reject reading the study as a case against using AI: three of the 50 who used the agent made serious methodological errors, against six of the 11 who worked without it. The no-agent group was small and had less experience, so I read the gap as a hint that the agent helped and don’t lean on it.
I wouldn’t take comfort from an agent that has no planted lean. With no lean assigned, the model picked the same five options in 112 of 123 runs. The authors add that agents with no assigned lean hold weaker preferences and sometimes push back. A consistent default is a lean nobody chose and nobody audits. That is my inference; the study didn’t test it. The study covers one task, one field and one model, with a lean strengthened on purpose, so the real effect may be smaller, and the fix may matter less than it seems.
What to do on Monday
Pick one decision type that matters and record the reviewer’s own call before they see the agent’s, so there is something to count. Each month, count how often reviewer and agent disagree, and how often the reviewer changes their call. A disagreement rate near zero can mean the agent is right or that nobody is checking.
Where Resolon comes in
We help you design the approval step so the reviewer’s judgment stays their own, starting with our AI governance and human control work.