Your team asks an AI assistant whether a new sales channel is working. It produces a polished report and recommends investing more. A manager checks the calculations and signs off. But who chose what working meant: revenue, margin, repeat customers or the time spent serving those orders?

By the time someone reads the recommendation, the choice that shaped it may be buried in the analysis.

Research published by Anthropic in October 2026 tested this with 50 professional analysts. Each used an AI agent to assess a fictional farming cooperative. The researchers secretly gave the agents preferences on four reasonable choices, including how to measure productivity. Analysts picked the agent's preferred option 86% of the time across those choices.

The researchers deliberately strengthened those preferences, and this was one task with one model. The figure does not tell us how often this happens in a business. It shows why a human sign-off alone may not reveal who made the earlier judgment calls.

Find the decisions hiding inside the task

In a small business, those choices can be ordinary:

  • A stock report counts items sold rather than contribution margin, making a high-volume, low-margin line look like the priority.
  • A supplier comparison gives price most of the weight, while the team would pay more for reliable delivery during peak weeks.
  • A customer-service summary labels every fast reply a success, although repeat contact shows that some answers did not solve the problem.

There may be no single correct measure. Someone who understands the business needs to choose what fits the decision and explain why.

Before the assistant starts, write five lines:

  • Decision and owner: The managing director decides whether to expand the channel next quarter.
  • Success measure: Contribution margin after fulfilment costs, alongside repeat purchases.
  • Comparison: A similar period, with campaign and seasonal changes noted.
  • Constraint: Do not increase orders beyond fulfilment capacity.
  • Open choice: Should first orders and returning customers be counted separately?

The manager can still change a measure. Now the change is visible and has a reason.

Review the choices before the recommendation

Ask the assistant to show the two or three choices that most affect its conclusion. For each, ask what would change under the reasonable alternative. Would we make the same recommendation if we measured success differently? If the answer flips, put that beside the recommendation, not in a footnote.

For an important decision, have its owner write down their initial view before reading the AI report. Afterwards, note what changed and why. New evidence is a good reason to change your mind. An attractive argument built on a measure nobody chose deserves another look.

Use this for decisions that spend money, change service levels or are hard to reverse. A routine summary needs less ceremony. A pricing change or supplier switch deserves an owner and a visible alternative.

Try it on one real decision

Try it on the next recurring decision your team already makes. Write the five lines, use the AI report, then mark the assumptions that changed. If the exercise exposes conflicting figures, settle those definitions before building a bigger dashboard. If it exposes repeated manual work, a useful tool or integration may be worth building.

We can help put the decision, data and review step into a workflow your team can use and challenge.

Test one workflow before scaling it.

Estimate the time it could release, then check the effort, risks and evidence needed for a pilot.