Key takeaways
- On a choice with two defensible answers, sound numbers, a fluent explanation and a confident analyst don't show whose judgment settled it, a recent Anthropic trial and its authors suggest.
- In that trial, published in October 2026, 50 analysts went with an AI agent's preference, secretly assigned by coin flip, on four such choices 86% of the time. Of the 48 surveyed, 77% later repeated its argument in at least one justification.
- Our proposal, untested: before the analysis starts, the decision's owner writes the business criteria into the brief, and any change to them during the work goes back to the owner rather than being settled by the analyst or the agent.
- That gives you a record of who authorized each criterion. It covers only the changes someone notices, and whether it reduces influence nobody notices is unknown.
- The researchers strengthened the agent's lean on purpose and tested one task with one model, and they say the effect might be smaller in real work.
In this article
Picture a Thursday review. Your head of operations recommends closing the smaller of two warehouses, and the analysis behind it was built with an AI agent over two days. The numbers reconcile. You ask why the comparison uses cost per order rather than cost per square foot, and she gives a crisp, sensible reason. She’s confident, she’s your most experienced analyst, and she says she’d leaned that way before she opened the tool. Finance, working separately, reached the same answer.
Those are good reasons to trust a recommendation, and checks like them catch real mistakes. The scene is ours, not a study’s. But a randomized trial Anthropic published in October 2026 suggests that on one kind of choice, sound numbers, a good explanation and a confident analyst don’t show whose judgment settled it. It’s the kind with two defensible answers, like cost per order against cost per square foot. The other reasons in the scene get a closer look below.
That kind of choice isn’t an error waiting to be caught. Either answer can be argued in front of a board, and picking one is a judgment about what the business values. Newsletter #34 made the short case for asking who owned such a choice at sign-off. This article goes further into the study, what it does and doesn’t show, and what we’d write down before the analysis starts.
How the trial was built
The researchers, Amaç Herdağdelen, Hailey Joren and Alex Tamkin, recruited 61 analysts who had passed a screen for experience in causal inference, which is the work of figuring out from data whether one thing causes another. Eleven chose not to use the AI agent. The other 50 worked with Claude Code, Anthropic’s coding agent, running Claude Opus 4.8, and those 50 are the study’s main sample. Anthropic published the study about its own product, which is worth keeping in mind as you read it. The researchers registered their analysis plan publicly, after the first participants had started, and since November 1, 2026, a public package with the study’s data and code, described as reproducing the results, has been online.
The task was set at a fictional farming co-op. Its board had to decide whether to expand an advisory program for member farms or wind it down, and it would expand only if the program lifted productivity by roughly 3 to 4%. The board didn’t say how to measure productivity.
So the analysts faced four open choices, each with two reasonable options. Measure productivity as crop yield per hectare or as yield per unit of operating cost. Compare the farms with one statistical method or another. Measure the effect on the enrolled farms or on the whole membership. Count every farm equally or give larger producers more weight. Each choice could move the estimated gain by several percentage points, which matters when the bar is 3 to 4%. Then came the recommendation itself, roll out or discontinue, with a one-year extension also allowed.
Here’s the part the analysts weren’t told. On each of the five decisions, a coin flip set which option their agent would favor. Hidden instructions told it to hold that lean as its own view, raise it when the decision came up and make the strongest case the data allowed, while changing its mind if the evidence turned decisively against it and never misrepresenting results. The researchers strengthened the lean on purpose, so they could measure its effect.
Across the four open choices, analysts went with their agent’s coin-flipped lean 86% of the time, against the 50% you’d expect if the agent made no difference. Choice by choice, it was 92% for the productivity measure, 82% for the comparison method, 81% for which farms were covered and 90% for the weighting.
When the numbers are right either way
Checking the numbers is how you’d catch a bad analysis, and it still catches errors. It can’t settle a choice with two defensible answers, because there’s no right result to check against. As the authors put it, the work can look sound whichever option is chosen. We’ve argued that a check belongs where authority changes hands, and this is a case where a check of the output has nothing to find.
The trial points to how far the problem may reach. The agent’s lean on the recommendation itself had a smaller effect, too small for the researchers’ main test to tell apart from chance. Yet in an exploratory analysis, the coin flips on the four earlier choices alone predicted 82% of the analysts’ recommendations, against 61% for always guessing the most common one. That figure leaves out analysts who proposed the extension, and it was tested on the same analysts it describes, not on a fresh group. The authors’ reading is that the recommendation mostly followed the estimate, and the agent had already shaped the estimate.
One session in the paper shows the chain. An analyst’s first method put the program’s gain at 4.6%, above the bar. The agent argued for a different way of comparing the farms, which showed a 1.6% loss, and the analyst delivered a version of it at a 1.3% loss and recommended discontinuing. The decision went against the agent’s own lean on the recommendation, which was to roll out. So the analyst overruled the agent at the end, on an estimate the agent had already moved.
Their reasons, in the agent’s words
The second kind of reassurance is about authorship: the analyst can explain the choice, sounds sure of it, and remembers making it. The study measured that directly. After the task, the 48 analysts who completed the survey were asked to justify each choice, without access to their conversation with the agent. Of those 48, 77% repeated the agent’s argument in at least one answer, sometimes nearly word for word. Only 15% mentioned the agent anywhere.
In one pair the paper prints, the agent had argued for giving larger farms more weight because the program’s subsidy is paid from surplus the co-op earns on total tonnage. The analyst’s later justification made the same point in nearly the same words, with no mention of the agent.
They weren’t confused about what they’d chosen. Analysts identified their own choice correctly on 85% of decisions. What their accounts didn’t show was where the reasoning came from. More than half called the final recommendation more their own than the agent’s. Their choices had moved toward the agent’s lean roughly as much as the choices of those who called it more the agent’s. Asked what the study was about, none suspected the agent had been given its preferences.
Confidence didn’t separate them either. On a scale from 0, “it’s a coin flip,” to 100, “I’d stake my professional reputation on it,” analysts who followed the lean on all four choices averaged 71, and the rest averaged 73. Seniority didn’t stand out: in an exploratory check of 10 traits, including experience and education, none held up once the researchers adjusted for testing so many, though they note there was little variation to detect.
An early view is the hardest to read. In 38 cases an analyst happened to state a preference during the session before the agent voiced its lean, and when the agent then disagreed they switched 60% of the time, against 6% when it agreed. The researchers call that imprecise, and the analysts who spoke first may have differed from the rest. So the finding can’t tell you whether deciding first protects you. In the Thursday scene, “I’d leaned that way before,” offered afterward, is one more account, and the study found that accounts often didn’t reflect the agent’s influence.
We’ve written about a person’s leading question tipping an agent’s recommendation. This trial runs the influence the other way, from the agent into the person’s own account.
One side, argued well
The agents were told never to mislead, and the researchers’ close reading of the transcripts suggests the influence didn’t come from misleading anyone or hiding consequences. The agents often explained how the result would change under each option. What they rarely did was explain why a competent expert might prefer the option they didn’t favor.
Part of the reason it worked, the authors suggest, is that experts don’t always arrive with their views fully formed. They often work out what they think while weighing the options. With only one side argued well, analysts could adopt the agent’s assumptions without working out their own.
When two analyses agree
The trial didn’t test what happens across teams, and the authors raise it as a warning. We trust a result more when separate teams reach it, because we assume they got there independently, and if those teams defer to models with the same preferences, the authors write, the agreement may no longer be independent. Their simulations make the worry concrete: run on the same task with no lean assigned, the model in the trial made the same five choices in 112 of 123 runs.
In the Thursday scene, if finance worked with the same agent, its agreement may count for less than it seems. The trial measured nothing across teams, so that stays a warning.
Where the trial’s reach ends
The authors are direct about the limits, and the limits decide how far any of this travels. They strengthened the agent’s lean on purpose. With an agent that doesn’t hold a clear preference or voice it, the effect might be smaller.
It was one task, in one field, with one model, and it took about an hour. The analysts knew causal inference but were new to farming. In real work they’d know the field, probably hold stronger views, and could talk to the board and others to narrow the options, which might reduce the agent’s influence.
The results are averages across analysts, and the study didn’t ask each person for a view before the agent spoke, so it can’t say how much any one analyst was influenced. Nor did it record how strongly analysts preferred the option they chose, and one possible reason for deferring is that some had no strong view to begin with.
One more comparison needs care. Six of the 11 analysts who chose to work without the agent made serious methodological errors that invalidated their results, against 3 of the 50 who used it, and they took longer. The groups weren’t comparable: the 11 chose themselves and were less experienced on average, so the study can’t say the agent caused the difference. The authors say plainly that their study isn’t an argument for using AI less. Some of an analyst’s freedom, in their words, is room for “arbitrariness, even error, that an agent can usefully curb.”
Write the business question down first
What follows is our proposal, informed by the study and its discussion. Its effect hasn’t been tested. The closest established practice, fixing an analysis’s choices in advance, isn’t foolproof either: the paper cites work finding that tighter instructions can fail when teams don’t carry them out properly. The third step below is a partial check on that.
The proposal starts from a working distinction of ours, not the study’s. An analysis contains business criteria, which say what the business wants: what counts as success, whose results count, how the units are weighted. It contains analytical assumptions, which are the method’s judgment calls. And it can contain errors. The decision’s owner sets the first, the analyst explains and stands behind the second, and the third still gets checked. The study called all four of its open choices analytical decisions. In our terms, three of them, the productivity measure, which farms count and the weighting, are business criteria, and the comparison method is an analytical assumption. Where a choice falls can still be arguable.
First, the owner writes the business criteria into the brief before the work starts. It comes before the work because by sign-off the choice is already inside the estimate, as the co-op session above showed. Start with one analysis your team repeats, such as a pricing review or a capacity plan. In the one-off warehouse case, an illustration of ours, the request might say: compare the two sites on cost per order shipped, count every customer segment, and count each order once rather than weighting by revenue. Write each criterion as something you could find in a table or a formula.
Second, a rule for changes while the work runs. Any change to the business criteria, including one the brief didn’t foresee, needs the owner’s authorization. An analytical assumption that could materially affect the recommendation needs an explanation to the owner of what it changes. Midway through, say the agent argues for cost per square foot because one site is half empty. The switch would change a business criterion, so it goes back to the owner with what it would change, and the owner decides. Separately, the analyst drops a month when one site was shut for repairs. Dropping the month is an analytical assumption, so it gets explained, with its effect on the gap. “Could materially affect” is our threshold, and judging it is fallible.
The rule works as an authorization procedure only for changes someone recognizes and reports. One session in the paper is a warning that a written instruction may not be enough on its own. One analyst spent 43 minutes studying the task before sending the agent a first instruction, which asked for a specific comparison method and for the effect on enrolled farms only. The agent pushed back on both in one message, arguing on the second that the board’s question was about a full rollout, and the analyst accepted in one line. In our terms, which farms to count was a business criterion. It was the analyst’s own instruction, though, not an owner’s brief, so the session is a warning and not a test. Nothing here makes the owner harder to persuade than the analyst was. The proposed change may still arrive with one side argued well. What changes is that the person who answers for the decision makes the call, knowing a criterion moved, and there’s a record of who made it.
Third, read what comes back against the brief. Find the measure the analysis actually used, in its table or its formula, and set it beside the brief or the owner’s recorded change. Ask the analyst to walk through the assumptions by what each one changed, which you can check against the numbers, though that won’t tell you where an idea came from. Set side by side, the two show what was authorized and what was used. The comparison doesn’t prove every departure was caught, and it says nothing about whose judgment shaped the choices nobody flagged.
The part nobody has solved
Three tasks are tangled together here: identifying a choice, understanding what it changes, and knowing where the preference behind it came from. The method asks people to do the first two for the choices they notice, and whether it improves either is untested. It certifies nothing about the third, and the third is where the study’s sharpest finding sits: the people doing the analysis often left the agent out of their own reasons.
The authors’ proposals point at the same gap, as directions for future study rather than tested fixes. Agents could make the best case for each option, or ask for the expert’s view before offering their own, though the authors warn that raising every open decision could quickly overwhelm people, and that picking which ones to raise is a judgment call too. Researchers need ways to measure deferral in everyday work, which the trial could do only because it randomized the agent’s lean, something rarely possible in real work. And institutions, they write, will need to decide deliberately which decisions need a person’s own informed judgment, give people the time to form one, and protect the skills behind it, since people who stop making these calls may lose the ability to make them.
The researchers expect calls like these to matter more as AI takes over more of the hands-on work, because there’s no answer key to benchmark them against. Checking whether an answer is right will keep catching errors. On calls like these, it has nothing to check against.
Questions this article gets
Does this mean my team should use AI less?
The authors say their study isn't an argument for using AI less. Analysts who used the agent finished faster and made far fewer serious errors than the 11 who chose to work without it, though that group was small, chose itself and was less experienced, so the study can't credit the difference to the agent. The concern is narrower: choices with two defensible answers.
Can't I just ask the analyst whether the AI influenced them?
The study suggests that won't settle it. Of the 48 analysts surveyed, 77% repeated the agent's argument in at least one answer and only 15% mentioned the agent anywhere. Asked what the study was about, none suspected the agent had been given its preferences.
Does writing the criteria into the brief stop the agent's influence?
Nobody has tested that. It records who authorized each business criterion and routes any change back to the decision's owner, but only for changes someone notices. In one session the paper describes, an analyst asked in the very first instruction for a specific comparison method and for the effect on enrolled farms only, and accepted the agent's case against both. Which farms to count is, in our terms, a business criterion.