Bottom line
When AI recommendations disagree, first inspect the assumptions and objectives behind them. The useful question is not which answer sounds most certain, but which reasoning fits the decision, survives an evidence check, and acknowledges the trade-offs the responsible human must accept.
Restate the choice and the objective.
A recommendation to cut costs and a recommendation to improve customer service may both be internally coherent. They are not answering the same question if one optimizes this month's spending and the other optimizes retention over a year. Establish the time horizon, constraints, and success criteria before comparing the conclusions.
Create a disagreement table.
Use one row for each option and columns for expected benefit, supporting evidence, key assumption, downside, and missing information. Copy the reasoning accurately rather than rewriting it to favor your preferred answer. Label speculative claims. The table should make the source of disagreement visible, not manufacture a numerical score that implies precision the evidence cannot support.
Separate factual disputes from value choices.
A disagreement about current customer churn can be investigated using business records. A disagreement about how much short-term margin to sacrifice for a better customer experience requires a human priority decision. Ask the agents to identify which claims could be checked and which trade-offs remain even if all participants accept the same facts.
More on Decision Science — every week.
Join 2,400+ leaders getting the decision intelligence briefing.
Test the assumption that could change the outcome.
In a hypothetical pricing discussion, one recommendation might assume that customers are highly price-sensitive while another assumes reliability drives renewal. Review interviews, renewal reasons, or a carefully scoped experiment before treating either assumption as established. Do not ask the same model repeatedly until it produces the answer you wanted; repetition is not independent validation.
Keep a credible minority view.
If one option wins, preserve the strongest objection and the conditions under which the rejected alternative would become preferable. A minority argument can become important when new evidence arrives. Agreement among agents is useful as a summary of a discussion, but it is not proof of correctness, independence, or freedom from shared model biases.
Decide, defer, or run a bounded test.
The owner can choose an option with a documented trade-off, postpone while collecting a decisive missing fact, or authorize a reversible experiment. Record the reason and a review trigger. For decisions affecting safety, legal obligations, or significant financial exposure, obtain qualified review rather than using AI consensus as a substitute for expertise.
Research context
In a McKinsey survey, 72% of senior executives said bad strategic decisions were as frequent as good ones or were the prevailing norm.
McKinsey & Company: Untangling your organization's decision making (2017)This independently published statistic provides context for the topic; it does not substantiate every claim in this essay.
More on Decision Science, weekly.
One essay. Every week. No noise.