A thousand answers. Not one customer asked.
Imagine a meeting about a new service. The screen shows a thousand customer responses: reactions to the price, preferred descriptions, reasons to choose. Nobody has spoken to a customer. The team asked AI to play the customers instead. This is a hypothetical scene, not a reported case.
The exercise is quick and convenient. The problem begins when we label it “customer research.” Answers we generated can start passing for answers we heard.
Analyzing an answer is not generating one.
AI can help organize interviews with real people. It can also act as a virtual respondent and produce new answers. These are different activities. Synthetic respondents may even be personalized using information about real individuals. That does not make the individual the author of the new response.
| Where the evidence starts | What it can establish | What it cannot establish on its own |
|---|---|---|
| Human surveys and interviews | What participating people said under the study conditions | Whole-market opinion or actual behavior |
| AI analysis of human responses | Categories and summaries checked against the originals | Perfect classification accuracy |
| AI-generated virtual responses | What a model produced from specified inputs | Human preference shares, market demand or sales forecasts |
This is my proposed distinction, not a quality certificate for human surveys. Sampling and question design still matter. But a limitation in observed responses is different from a limitation in generated ones.
A customer is allowed not to know.
In an experiment released on September 30, 2026, Pew Research Center compared U.S. panelists with synthetic versions of them. Across opinion questions offering “not sure,” human responses selected it 16% of the time; synthetic responses, 4%. The default reported model was Claude Opus 4.6. This was not a study of Korean customers or purchasing behavior. [1]
My interpretation for brands is narrower: uncertainty can be useful information. “I do not see the difference.” “I have not thought about it.” “What I already use is enough.” Perhaps the explanation needs work. Perhaps the problem matters less to the customer than it does to us.
An articulate virtual customer can make a market seem unusually attentive. Fluent reasons do not establish real interest. We may be thinking deeply about a question the customer has never asked.
Useful does not mean interchangeable.
A 2026 article in NIM Marketing Intelligence Review presents a brand-research experiment with both promise and limitations. Digital twins reproduced broad patterns and predicted individual answers better than chance. They also overstated positive brand attitudes and understated differences between people, particularly for familiar brands. [2]
An average can look plausible while hiding who cares and who does not. A brand needs to understand those differences, not just meet an imaginary average customer.
AI can help explore questions and possible objections before fieldwork. A separate NIM methods article stresses sensitivity to question design and answer order, alongside validation against human data. [3] My proposal is to use generated responses to prepare an investigation, rather than let them close it.
More answers are not different evidence.
Suppose virtual respondents like a proposed subscription service. Generate more responses and the presentation grows. We still have not observed what real customers pay for today, why they might switch, or where the price makes them hesitate.
A larger generated dataset does not by itself establish a representative human sample. A model responding to our description is not the same as a market responding to our offer. Even a real person saying they like an idea is different from choosing it.
These are my suggested checks, not a validated operating standard:
- Label generated responses clearly. Do not count them together with human survey responses.
- Retain the model, version, date, inputs and questions so the conditions can be compared.
- Decide what to ask actual target customers, including people who are uninterested or do not use the category.
- Name the real observation or response that would change the current hypothesis.
The answer we want is not always the answer we need.
A favorable reaction may become easier to produce. The uncomfortable questions remain: Does this need matter to anyone else? Can someone understand the offer and still not want it? What can we learn from a refusal?
AI can be a rehearsal partner for understanding customers. Agreement in rehearsal is not agreement in the market.
Generating an answer is not the same as hearing from a customer.
Sources and limits.
[1] Pew replicated U.S. public-opinion surveys from the first half of 2026. The percentages describe eligible opinion responses, not purchase rates or accuracy. They do not establish the behavior of every model or industry.
[2] Kaiser et al. (2026), NIM Marketing Intelligence Review, 18(1), 48–53: a discussion of a brand-research experiment, not evidence of Korean demand or sales. [3] Sarstedt et al. (2026), the same journal, 18(1), 24–29: methodological guidance. The scenarios, table and checks here are the author’s proposals, not a ban on synthetic research or a guarantee of results.