goagentdata

Your agent has an opinion. You may think it's yours

TL;DR. An AI agent can steer a data decision without anyone noticing. In a study Anthropic published in October 2026, 50 analysts matched an agent's secretly assigned lean 86 percent of the time, and few credited it. To keep the decision yours, write your measure of success and your expected result before you open the agent.

Checking the number rarely catches this, because a judgment call has no single right answer. So ask the agent to list its choices, and rerun the ones that could change the decision.

Agent first: the agent's lean becomes your decisionA question goes to an agent. Its answer carries a small lean marker toward Keep. Your view forms after it, toward Roll back, then slides silently until it sits on the agent's marker. The final recommendation carries the agent's marker while its label reads my decision. Illustration of the method, not measured data.Agent firstDid the new onboarding work?Agent's answerRoll backKeepleanYour viewRoll backKeepSlides to match, unnoticedRecommendationmy decisionCarries the agent's lean
You first: your written note keeps the gap visibleYou write a sealed note first, with your measure of success and expected result. The agent answers. The two are laid side by side on one line, and a warning-colored gap shows where they differ. The gap leads to a step, rerun with the other option, and then a named person decides. Illustration of the method, not measured data.You firstYour note, sealedMeasure, expected resultAgent's answerleanSide by sideRoll backKeepyour noteleanGapRerun with the other optionA named person decides
Illustration of the method, not measured data.

Do AI agents sway data decisions?

Picture a founder asking an agent whether the new onboarding worked. The agent must choose what counts as success, which customers to compare, and whether one huge account stays in. Together, those choices can decide the answer.

Anthropic, the company that makes the model tested, ran an experiment on this and published it on 27 October 2026. Analysts advised a fictional farming cooperative with an AI coding agent. For each of four analysis decisions, a coin flip set which option the agent's hidden instructions told it to favor.

  • They followed it. Choices matched the agent's lean 86 percent of the time, where chance would be about 50.
  • It changed minds. In 38 cases an analyst took a side first. They switched 60 percent of the time when the lean went against them and 6 percent when it agreed, an imprecise estimate (the gap ranges from 24 to 82 points).
  • The final call moved less clearly. 54 percent recommended rollout when the agent leaned that way, and 36 percent otherwise. On the paper's recommendation score, the adjusted range spans minus 6 to 40 points, so it could be zero.
  • They did not see it. 77 percent repeated the agent's arguments in at least one answer, and only 15 percent credited it anywhere. None suspected the agent's preferences had been manipulated.

It is one task and one model, and the researchers deliberately strengthened the lean, so everyday effects may be smaller. They also write that this is no argument for using AI less. With no lean assigned, the same model made the same five choices in 112 of 123 runs, so a default can act like a lean.

Haven't we seen this in meetings?

Partly, I think. The first-speaker link is my reading. The authors tie their result to automation bias, over-relying on automated tools.

Before agents, an analyst made these choices and a colleague pushed back. In a 2011 experiment, 144 students estimated facts, some seeing what others had answered. Their estimates grew more alike without becoming more accurate, and their confidence rose anyway. Anchoring is the other half: a first number pulls later guesses toward it, as Tversky and Kahneman showed in 1974. I think that is why writing answers down first helps.

An agent is that first voice, with three differences:

  • It often goes first. Told to raise its lean, the agent voiced a side first on 77 percent of 245 decisions, the analyst on 17.
  • Nobody else is in the room. A meeting has a second person who might object.
  • It argues its case. Even analysts who spoke first often switched, and its reasons can feel like your own.

Why doesn't checking the answer catch it?

The evals in our earlier post catch a wrong number. They cannot tell you whether seven-day activation beats 30-day retention, because that depends on your business.

So "a person reviews it" is a weak control here, because the reviewer reads the answer first. Ask reviewers for one line of what they expect before they open it.

How do you stop an agent steering your analysis?

Write first. Before you open the agent, write one short note: the decision, the measure of success, and the result you expect. You need this only for decisions about product, price, spend or hiring. If your metric is not defined in writing yet, that is the real work.

In a 2024 study, people judged mock defendants with an AI that was wrong once in three. In the first experiment, with 150 laypeople, 66 percent of those who judged before seeing the AI were right on the case where it erred, against 37 percent who had seen it first. But once everyone had seen the wrong advice, accuracy was not significantly different.

Writing first protects your view, not your reaction to the reply, so you need a second step.

Compare, then rerun. Where your note and the answer differ, do not reread the agent's argument. Rerun with your option and see whether the conclusion moves.

Ask for the choices. You need no repository for this. Paste the prompt below with your question, and leave your guess out of it, because the assistant may simply agree with it.

The measure that decides this is: <your measure>.
Before you recommend anything:
1. List the choices you made: what you measured, who you counted,
   the dates, and what you left out.
2. For each, say what the other sensible option was.
3. Run it again with the other option and show both results side by side.
4. Tell me which choices change the answer.
Then give your recommendation in plain words.

In a test on a made-up data set where the new onboarding raised activation but not paid conversion, the assistant recommended keeping it when no measure was named, even with this prompt. With the measure named, it said the data showed no win. One run each with one model, so this illustrates, it does not measure.

Can't we just think harder?

Only partly. In the study, experience, expertise, and measures of open-minded thinking and enjoying hard thinking did not predict who followed the agent, though 50 people is a small test.

Training helps modestly. In a 2026 study of 226 students, a 2-hour session combining a questioning AI tutor with role play reduced how much groups copied from AI, without raising self-rated critical thinking. A short reading on how ChatGPT works did not significantly reduce following wrong suggestions among 36 high school students in a 2025 preprint.

Written prompts show mixed results. In a 2026 preprint with 214 people on AI-assisted writing, people who named the assumption behind an AI's plan accepted 63 percent of flawed outputs, against 71 with no prompt, a gap that was not significant. A what-if prompt scored 81, and combining the two did not help. I still think of the note and the prompt as critical thinking on paper, but the evidence is thin.

What does this look like at each stage?

I think stakes matter more than stage: a pricing call before launch deserves more care than a button test later. Stage decides who does the work.

Before launch. Data is thin, so this mostly applies to pricing and market questions. Write the note in a chat message and use the prompt above.

First paying customers. Name one owner per decision, and keep dated notes in a shared doc. Whoever asked the agent runs the prompt.

Growing with a first data person. The data person sets which decisions need a note, reviews the choices as well as the numbers, and moves the notes and the rule into the repository.

What goes wrong, and how do you catch it?

The agent chose the measure. You asked whether onboarding worked, and it filled the gap with seven-day activation.

GuardrailThe decision owner names the measure in the note before the agent runs.

CheckThe decision owner compares the note's date with the analysis date. The note must be older.

The conclusion rests on one quiet choice. Dropping one large account flips the result, and nobody mentions it.

GuardrailThe person who ran the analysis shows both results in one table.

CheckThe decision owner asks the agent to rerun with the account kept in. If the result crosses the line in the note, the choice is the decision.

For your data team

For your data team

This part gets technical, and it assumes a repository and a coding agent. Without those, the note and the prompt above are enough. The standard techniques are a pre-analysis plan (preregistration), a sensitivity check that changes one choice at a time, and, with many defensible choices, specification curve analysis.

The note, committed before the session:

# decisions/2026-11-onboarding.yaml
decision: keep or roll back the new onboarding
owner: <named person>
written_on: 2026-11-04
success_measure: paid conversion within 30 days of sign-up
who_is_counted: new accounts, excludes is_test and is_internal
window: sign-ups from <start> to <end>, only cohorts at least 30 days old
comparison: <new vs old onboarding, and how accounts were assigned>
changes_the_decision_if: <the lift at which you would keep it>

Your expected result stays out of this file, since the agent reads it.

A rule for your agent's instructions file, such as CLAUDE.md, or AGENTS.md for Cursor. It is part of the harness:

For any analysis that feeds a decision about product, price, spend or hiring
(skip plain lookups such as counts and status checks):
1. Read the matching note in decisions/. If none exists, ask for one first.
2. Open with the judgment calls (measure, who is counted, window, exclusions,
   method), the option you used for each, and one line on the case for the
   main alternative. Then give your recommendation and name the calls it
   depends on. If asked for one line, give it and name the call that would
   change it.
3. For each call that could reasonably go another way, rerun with the other
   option, one change at a time, and show the results in one table. Say which
   calls flip the conclusion.
4. Use read-only data access. Write only to analysis/, never to production.

Caveats:

  • Whether rule 2 lowers deference is untested.
  • The rule guides the agent. Read-only credentials enforce only the write limit.

Check it in 10 minutes

Take the last agent analysis that fed a decision.

  1. Ask the agent which filters, dates, thresholds and definitions could reasonably have gone another way.
  2. Mark each one: did a person decide this before seeing the result?
  3. Pick the unmarked choice that touches the headline number. Rerun a copy with the other option, same data.
  4. If the conclusion flips, the answer depends on a choice nobody made on purpose. Name who owns it. A flip does not say which option is right.

See where your setup stands with the two minute self-check. Then read the draft checks your coding agent can run.

Spotted a mistake or a newer source? Tell us and we check it against the source. How we work.

Check your setup in 2 minutesAll postsGet new posts by email