AI personas are not customer research

Core message

AI personas can generate hypotheses, test variants, and accelerate the preparation of real research. However, they should not be treated as a substitute for customer interviews, surveys, sales feedback, usage data, or market observation.

A study published in October 2026 in Nature Human Behaviour makes this limitation particularly clear. The researchers had language models simulate the economic preferences of 67,201 real respondents from 74 countries. Among other things, risk tolerance, patience, trust, and altruism were examined. The models recognized some demographic patterns but reflected cultural differences much less effectively. The authors conclude that simulated responses are not yet reliable proxies for representative cross-country surveys or policy impact analyses.^[1]

For companies, this does not mean a blanket rejection of AI personas. It follows a clear separation between a source of ideas and market evidence.

Why convincing answers are easily overestimated

Plausibility is not representativeness

A language model generates a linguistically appropriate continuation based on its training data and the input. When it is assigned the role of a purchasing manager, a design engineer, or a practice owner, it can formulate a plausible response.

This response initially shows what the model considers probable. It does not show how frequently a stance occurs in the target market, which customer groups decide differently, or what the actual willingness to pay is.

The difference is business-relevant. A convincing statement like "The simple integration is decisive for me" can trigger a good interview question. However, it must not become the basis for positioning, product development, or budget allocation without further evidence.

Averages can be correct even if the decision logic is wrong

Already in a 2024 in Political Analysis A published study compared synthetic ChatGPT responses with data from the American National Election Study. Some average values were close to the human comparison data. At the same time, the simulated responses showed too little variance, different correlations between characteristics and responses, and a sensitivity to small prompt changes as well as model changes over time.^[2]

This is an important warning. A result can appear plausible on the surface and yet miss the differences that are crucial for segmentation and decision-making.

The model does not know any unknown market truth

Companies often use research precisely because the answer is not yet known. They want to understand why prospects drop out, which alternative customers are actually comparing, or which problem is described differently internally than in the market.

An AI can condense existing public patterns. However, it cannot create a new customer experience that has never been collected. If the model merely varies known assumptions, a feedback loop easily emerges: the company inputs its perspective and receives it back in professional language.

What AI personas can be useful for

The boundary against pseudo-research does not exclude productive applications. AI personas can help structure the thought space before a survey.

Developing hypotheses and counter-hypotheses

A team can ask a model to collect possible reasons for a decision and subsequently formulate counterarguments or alternative situations. The result is a list of hypotheses, not a finding.

It is useful to label every statement as an assumption to be verified. This directly includes the question of which real-world source could confirm or refute the assumption.

Prepare interview guides

Open-ended questions for customer interviews can be derived from simulated objections. For example, 'The implementation effort seems too high' becomes not the identification of a market barrier, but the question: 'Which efforts or dependencies particularly influence your decision?'

The AI supports the preparation. The insight only emerges from the answers of real conversation partners and their classification.

Check concepts for obvious gaps

A model can take on different roles and point out missing information, unclear terms, or contradictory promises. This is suitable as an early editorial or concept review before real users invest their time.

Here, too, the statement remains limited. A found weakness is a hint for review. A weakness that is not found is not proof of clarity or market suitability.

Making surveys and tests more efficient

Question variants, answer options, and sequences can be played through in advance. This allows ambiguous wording to be noticed earlier. The final survey still requires real participants from the relevant target group.

What simulated customer responses are not sufficient for

AI personas should not be used alone to:

• quantify market size or demand;
• estimate willingness to pay;
• derive segment shares or conversion probabilities;
• evaluate cultural differences between countries;
• to label a positioning as validated;
• to completely replace customer interviews or user tests;
• to justify high investments or irreversible product decisions.

The greater the economic consequence, the more important real observations, verifiable data, and a documented methodology become.

A framework for action for companies

1. Define the decision question

Before use, it must be clear which decision is being prepared. Is it about an initial collection of ideas, the formulation of an interview guide, or a reliable market decision? Only in the first and second case can a simulation alone be a sensible state of work.

2. Label the evidence level

Separate at least three levels:

This separation prevents a linguistically elaborated draft from unnoticedly becoming supposed market knowledge.

3. Specify target group and context

A role like "B2B decision-maker" is too broad for research. Relevant differences can concern function, industry, company size, usage situation, existing system, risk, country, or purchasing phase.

The specification improves the simulation, but does not turn it into a representative sample. It primarily helps in selecting real conversation partners and comparison groups sensibly.

4. Validate against real data

Check central assumptions against available sources. This can include structured sales feedback, search queries, website paths, support cases, lost deals, interviews, and tests.

The National Institute of Standards and Technology recommends documenting test data, metrics, and limits for AI evaluations, representing relevant populations, and checking results against new ground truths as soon as they become available. For business practice, this means: A simulation is only closer to decision-making once its statements have been checked against real observations.^[3][4]

5. Document prompt, model, and time

When simulated answers are used in a work process, at least the model, version, date, prompt, context used, and evaluation should be recorded. Results can shift due to wording or provider changes.

Documentation does not make a simulation true. However, it makes it comprehensible how an assumption was formed and whether later repetitions are comparable.

6. Define a stop line

Define in advance which decisions will not be made without real customer evidence. This boundary can apply to things like pricing, positioning, market entry, product scope, or major campaign budgets.

This way, AI remains a tool for preparation and does not unnoticedly become a substitute for the market.

What companies should not do

A large number of generated responses is not a large sample size. A thousand outputs from the same model are not a thousand independent customers. Even different persona prompts do not automatically create the diversity of real experiences.

Companies should also avoid creating false precision. Percentages, rankings, or segment sizes from unvalidated simulations appear measurable, even though their population consists only of model responses.

Finally, the model should not be the hypothesis, data source, and validation instance all at once. An independent cross-check is then completely missing.

Consequences for companies

AI personas are valuable when they generate better questions. They become risky when their answers are treated as already collected customer reality.

The economically sensible use therefore lies primarily in preparation: making assumptions explicit, collecting counter-hypotheses, improving interview questions, and identifying obvious gaps. For decisions regarding market, offer, price, and investment, real customers, real usage, and reliable data remain indispensable.

The decisive question is not: "Does this simulated answer sound plausible?" It is: "What real evidence do we need before we act on it?"

Subject-matter connection

From assumption to verifiable decision basis

An SDC Discovery can structure decision questions, existing evidence, open assumptions, and an appropriate verification path. It determines where AI serves for preparation and which statements must be secured through real data or conversations.

Clarifying the initial situation in an SDC Discovery

Further insights

1. Review AI results: Who checks what and when?
2. Measure marketing: First the decision, then the dashboard
3. AI in SMEs: Business Case instead of Tool Hype