AI Can Summarize 10,000 Customers. It Can Also Erase the One Group That Matters.

Research presented at the Federal Trade Commission in March 2026 found that AI summaries of 96,630 hotel reviews were systematically more positive than the reviews beneath them and compressed differences between properties.

A Stanford Digital Economy Lab study of 18 large language models found a second weakness: minority positions were persistently underrepresented when AI aggregated many opinions. For marketers, the decision is no longer whether AI can summarize customer evidence; it is whether a summary is trustworthy enough to steer positioning, creative and product promises.

The risk is not a dramatic hallucination. It is a polished consensus that never truly existed. If a small but valuable customer group uses a different buying criterion, an average-sounding brief can erase the very signal that should change the campaign.

Start With the Decision, Not the Dataset

Do not begin by dropping every review, transcript and survey answer into a model. Begin with one decision: Which promise should lead the next campaign? Which objection deserves a landing-page answer? Which audience should receive a different offer?

Then assemble five inputs: the decision question, a defined time window, the original customer language, any known customer segment or outcome, and the business constraint the answer must respect. Remove personal information that is not needed for the decision. Keep each comment attached to a source label such as prospect, first-time buyer, repeat customer or churned customer; do not ask the model to infer identities that the evidence does not contain.

The output is not “a summary.” It is an evidence map: recurring themes, meaningful disagreements, source counts, verbatim examples and unanswered questions.

Run a Five-Step Evidence-Preserving Synthesis

First, ask AI to classify each contribution without combining it with the others. Give it a short, human-approved codebook: desired outcome, stated obstacle, proof requested, emotional language and buying stage. Require an “unclear” label when the evidence does not fit.

Second, aggregate the labels and preserve their denominators. “Seven of 42 renewal customers mentioned setup friction” is decision-grade evidence; “customers dislike setup” is not.

Third, force a dissent pass. Ask for comments that contradict each leading theme, appear in a small segment, or change meaning when tied to customer value or lifecycle stage. The Stanford work matters here because the researchers found both minority underrepresentation and sensitivity to input order. Shuffle the material and run the synthesis again. If the conclusion changes substantially, the result is unstable.

Fourth, have a person inspect the original comments behind every recommendation. AI may organize the evidence, but a human must decide whether a minority signal is commercially important, whether sarcasm or context was misread, and whether the proposed promise is one the business can actually keep.

Fifth, convert only approved findings into a one-page decision brief: audience, problem, evidence, promise, proof, exclusions and the next test. OrionPilot is designed to connect business context, strategy, creative work, data and learning rather than treat each task as an isolated prompt. That makes the useful OrionPilot pattern here explicit: keep the evidence map, the human-approved promise and the measurement decision connected, so the next campaign inherits the reasoning—not merely a fluent conclusion.

Put the Promise Through a Human Gate

Illustrative example: A subscription meal company reviews 240 cancellation comments before a reactivation campaign. AI identifies price as the largest theme. The dissent pass reveals a smaller group of former weekly customers who did not object to price; they left because delivery windows felt unpredictable.

The team should not let AI choose between a discount and a reliability message. A human checks whether delivery operations can support a narrower promise, whether the smaller group is economically meaningful and whether “more predictable” can be substantiated. The approved output might be two test cells: a value-led message for price-sensitive former customers and a delivery-control message only for people whose comments or behavior support that concern. This is segmentation grounded in evidence, not invented personalization.

That approval gate protects the company from a common failure: converting a qualitative pattern into a universal claim. The FTC presentation found the strongest positivity bias among lower-rated properties and those with fewer reviews—exactly where uncertainty was highest. Sparse evidence should produce a narrower claim, not greater confidence.

Measure Whether the Decision Improved

Use one primary measure tied to the decision. In the illustrative reactivation test, measure the rate of former customers who restart and remain active through a second scheduled order, not opens or clicks. Run the test long enough to observe that second order—typically four to six weeks for a weekly service—and compare each message with a holdout that receives the existing campaign.

Also audit the synthesis itself. Sample at least 20 source comments from the largest theme and every comment from a strategically important minority cluster. Record unsupported classifications, omitted contradictions and claims that overreach the evidence. If those errors would have changed the promise or audience, stop the campaign and rebuild the map.

AI makes customer listening dramatically more scalable. It does not make consensus more truthful. The advantage belongs to the team that can compress thousands of voices without compressing away the disagreement that should change the decision.