Experts went with AI’s coin-flip picks 86% of the time. Here’s how to keep the call yours

On Tuesday, Anthropic published a study with an uncomfortable finding for anyone who uses AI on their numbers. Professional analysts working with an AI assistant went along with its preferences 86% of the time, even though a coin flip had decided what those preferences would be.

The AI was told never to mislead them. It simply made a good case for one option, and most of the time, that settled it.

What the study found

Anthropic recruited 61 analysts experienced in working out cause and effect from data. Their task: tell a made-up farming cooperative whether its paid advisory program for member farms was working and should be expanded. Fifty of them used Claude Code, Anthropic’s AI coding tool.

Each analyst faced four choices between two reasonable options, such as whether to measure productivity as crop yield per hectare or as yield per unit of operating cost. Each choice could move the answer by several percentage points, against a board target of roughly 3 to 4%. Without telling the analysts, the researchers had the AI favor one option on each choice, picked by coin flip, and make its best honest case.

• They followed the AI’s lean 86% of the time. If the AI made no difference, you’d expect about 50%.
• They gave up their own picks. When an analyst had already chosen and the AI disagreed, they switched 60% of the time, compared with 6% when it agreed, though that rests on only 38 cases.
• They adopted its reasoning. Asked afterward to explain their choices, 77% repeated the AI’s arguments, sometimes word for word. Only 15% credited it.
• It moved the bottom line. When the AI leaned toward expanding the program, 54% recommended expanding it, compared with 36% when it leaned the other way.
• The AI still helped. Six of the 11 analysts who skipped it made serious errors that invalidated their results, compared with three of the 50 who used it, though the groups weren’t directly comparable.

Anthropic deliberately strengthened the AI’s preferences, and it studied one task, in one field, with one model. But its main point holds beyond the lab: “Keeping a human in the loop assumes that the human’s judgment provides an independent check.”

Why it matters when AI runs your numbers

As we wrote last week, more small business owners are using AI to run the numbers. Most of their questions aren’t math problems with one right answer. They’re judgment calls dressed up as math.

Say a restaurant asks an AI whether its two-for-one Tuesday paid off. By revenue per night, it might look like a win. By profit per table, or by whether new customers came back, it might not. The AI will pick a measure and explain it well. If the study is any guide, you’ll go along and later remember the reasoning as your own.

Closer to home

The made-up co-op’s question is one farm families across the Wiregrass ask every season: did that program or practice actually pay? Businesses in Dothan and Enterprise ask it too, about a new ad, reminder texts, or a price change. How you define “worked” can decide the answer.

Five habits that keep the call yours

1. Decide what “worked” means first. Write down the measure and the bar, such as profit per job up at least 5%, before you open the chat.
2. Ask for the other side. Ask for the strongest case for the option the AI didn’t pick. Anthropic found its AI rarely made one on its own.
3. Ask what changes. Have it show the result under each reasonable choice. If the answer flips, the call is yours.
4. Notice whose reasons you’re using. If your explanation sounds like the AI’s, make sure you agree with it.
5. Test predictions before you trust them. For a forecast you’ll use again and again, check it against past months it hasn’t seen.

None of this means using AI less. It means making the important calls yourself.

Where we can help

Our predictive analytics service starts from the decision you want to improve, tests models on periods they haven’t seen so you know how accurate they are, and puts plain-language reasons behind every score. Our AI strategy and training workshops teach staff to use AI for analysis on their real work. Not sure where to start? Book a free strategy call.

Sources

• Anthropic: How AI agents shape experts’ judgment calls, October 27, 2026
• Anthropic: Who’s Aligned to Whom? How AI Agents Shape Experts’ Research Judgment Calls (PDF), October 2026