Often, yes. In Anthropic's October 27, 2026 study, 50 professional analysts working with Claude Code went with a secretly coin-flipped agent lean 86% of the time, and 77% later repeated its arguments while only 15% credited it. John Whitmore's Coaching for Performance names the fix: responsibility needs a choice you made yourself, so write your lean first and ask for both cases before any verdict.

The comfortable story about working with AI agents is that the agent does the legwork and you make the calls. Anthropic has just tested that story on people who should be hard to sway, and the result deserves a slow read before your next board update.

The study, "How AI agents shape experts' judgment calls" by Amaç Herdağdelen, Hailey Joren and Alex Tamkin, was published by Anthropic on October 27, 2026. Fifty professional analysts, all screened for causal-inference skill, did a realistic analysis in Claude Code for a fictional farming cooperative deciding whether to expand an advisory programme. They faced four analytical choices with two reasonable options each, such as measuring productivity per hectare or per unit of operating cost, and then a final recommendation. Each analyst's agent was secretly told to favour one option per decision, set by coin flip, to argue its case honestly, and never to mislead.

Across the four analytical choices, analysts went with the agent's lean 86% of the time, against roughly 50% if the agent made no difference. In 38 cases an analyst had stated a preference before the agent spoke; when the agent disagreed they switched 60% of the time, against 6% when it agreed. Afterwards 77% explained their choices using the agent's arguments, sometimes word for word, and only 15% credited the agent. Experience, education, quiz scores and trust in AI did not predict who deferred. From the coin flips alone, the authors could predict 82% of the final recommendations. Those who followed the agent on every decision rated their confidence 71 out of 100, against 73 for the rest. They felt as sure, and as much the authors of the result.

The study is not an argument for working alone. The 11 analysts who chose not to use the agent took a median 90 minutes against 65, and six of them (55%) made serious methodological errors, against three of the 50 who used it, though that group was smaller and less experienced. The authors also strengthened the agents' preferences on purpose; with no assigned lean, the earlier Claude Opus model used in the study made the same five choices in 112 of 123 runs, so everyday leans may be weaker. The finding that matters is narrower and harder: on calls with no single right answer, the work looks sound whichever way it goes, so nobody notices who decided.

John Whitmore's Coaching for Performance explains why this is a performance problem and not only an ethics problem. His two pillars are awareness and responsibility, and responsibility in his sense means a choice you genuinely made, not one handed to you. Telling, he argues, breeds dependence and weakens self-belief. An agent that argues one side well is telling, however politely, and Charter's AI Pulse research this month found executives twice as likely as employees to trust AI output over their own judgment. The people with the most decisions to make are the most exposed.

The lean-first protocol (four steps)

  1. Write your lean before you open the chat. One line and one reason, on paper. The study's switch rate only exists because some analysts had a stated view to switch from.
  2. Ask for both cases, equally strong, and no verdict. The authors note their agents rarely made a strong case for the other option; make that the instruction.
  3. Ask what each case assumes and what evidence would flip it. That turns a persuasive paragraph into something you can check.
  4. Write your choice in your own words, then ask for the model's view. If you change your mind, note what moved you. That one line is the credit the study's analysts did not give.

I ran steps two to four as a single prompt on October 28, 2026 with Claude Opus 5.5, on an invented pricing decision with my lean written in.

Tested October 28, 2026. I had my research agent run this exact prompt against Claude Opus 5.5 (default settings, one run, no earlier conversation). The prompt:

I'm about to make a judgment call and I want it to stay mine. Follow these rules before you give any opinion:

1. Do not tell me which option you prefer, or hint at it, until I explicitly ask at the end.
2. For each option, write the strongest honest case, as someone who genuinely believes it would argue it. Keep each case under 120 words and make them equally strong.
3. Under each case, list the two or three assumptions it depends on and the one piece of evidence that would change the answer.
4. Then stop and ask me to write my choice and a one-sentence reason in my own words before you say anything else.

The decision: we ran a four-week pricing test on our B2B SaaS product, raising the annual plan from $900 to $1,200. Trial-to-paid conversion fell from 11% to 9%, but revenue per trial rose about 9%. I have to decide which metric defines success in the board update.
Option A: […]

The start of what came back, unedited:

## Option A: Trial-to-paid conversion rate

**The case:** Conversion measures whether the market accepts the product at this price. A drop from 11% to 9% is an 18% relative decline in customers won from the same pipeline. Each lost customer also takes away future expansion, referrals, and case studies, none of which a four-week revenue snapshot can capture. Revenue per trial reflects only first-contract value and says nothing about lifetime value. A shrinking customer base can quietly compound into a weaker funnel. The board needs the metric that shows whether we're winning customers, not one that a price increase can flatter by construction.

**Assumptions it depends on:**
- Lost customers would have had meaningful lifetime value through retention, expansion, or referrals.
- Customer count matters strategically, for example because of network effects, land-and-expand motion, or the fundraising narrative.
- The 2-point drop is real, not noise from four weeks of data.

**Evidence that would change the answer:** Cohort data showing the $1,200 customers retain and expand at least as […]

One run is one run: the same prompt gives a different answer next time, and a different model answers differently.

What it did well: the two cases were close to even in length and force, it named no preference, and it stopped and asked for my choice as told. Each case listed its weak assumptions, including the four-week sample, and named the evidence that would change it: cohort retention for one side, early churn for the other. The 18% relative fall in conversion it calculated is correct.

What it did badly: it accepted my two-option frame, when the obvious board answer is to report both metrics and say which you are optimising. My stated lean sat in the prompt, and one run cannot show whether it shaped the wording. And nothing here shows the protocol reduces deferral; the study did not test it, and the authors list "ask experts for their own view before offering one" as an idea for future work, with the honest worry that it could overwhelm people. Use it on the two or three calls a month that carry your name, not on every choice an agent makes.

For the related evidence on boards following AI over their own judgment, see whether you should follow an AI recommendation over your own judgment.

Sources: How AI agents shape experts' judgment calls, Anthropic, October 27, 2026; AI Pulse, Charter, October 2026; Executives Trust AI More Than Their Own Judgment, Study Finds, Allwork.Space, October 22, 2026; John Whitmore, Coaching for Performance (5th ed.).

Related: How to Find Your Passion · Best Self-Improvement Books · How to Make Better Decisions · What University Will Not Teach You