THE SHORT ANSWER

Use AI to help code, compare and summarise material collected from real people with appropriate permissions. Keep each interpretation connected to its source. Generated personas and simulated interviews can suggest hypotheses, but they do not constitute consumer research.

Keep three kinds of material separate

Observed material includes interview transcripts, reviews, support questions and sales notes. Interpretation is the analyst's explanation of what that material may mean. Synthetic output is generated text. Mixing all three into one document can turn a plausible suggestion into an apparently established customer need.

For a hypothetical subscription service, customers may repeatedly ask how cancellation works. That observation supports investigating uncertainty about cancellation. It does not by itself establish that price is too high, that most customers intend to cancel, or that a generated ‘budget-conscious persona’ exists.

Generated systems can fill gaps confidently; the existing guide to AI hallucinations explains why fluent synthesis needs checking.

Evidence & context: NIST

Use AI as a second reader with a source trail

  1. Define the research question and document how the material was collected.
  2. Remove unnecessary identifying details before using an approved tool. Confirm permissions, retention and access arrangements.
  3. Give each extract a reference ID; ask for themes with supporting IDs and contradictory examples.
  4. Inspect the cited extracts yourself. Keep ambiguous statements ambiguous.
  5. Compare a sample with human coding and investigate disagreements.
  6. Turn the strongest unresolved explanations into questions for further research.

Do not report the number of generated theme mentions as the number of customers who hold a view. Duplicated reviews, repeated contacts and a single vocal respondent can distort a summary. Keep the unit of analysis explicit.

Evidence & context: UK Information Commissioner's Office

Social listening does not reveal the whole market

A set of public posts can reveal language, questions and situations worth investigating. It describes the material you collected, under the platform and search conditions you used. People who post may differ from people who remain silent; a ranking system also influences what you see.

Ask for counterexamples and missing groups rather than a single average persona. A segmentation idea is useful when it changes a decision and can be tested against actual behaviour or further research. Giving a segment a memorable name does not validate it.

An insight should change a decision

In the cancellation example, the next action might be a clearer explanation before purchase, followed by checking comprehension with real customers. It might also be a change to a genuinely confusing policy. More targeted persuasion is not the only possible response to research.

Before briefing content or personalisation, state what is observed and what remains a hypothesis. That distinction lets the next team act without inheriting false certainty.

Sources & further reading

  1. Generative Artificial Intelligence Profile (NIST AI 600-1)

    NIST. Risk-management guidance, including confabulation. It does not establish a universal error rate.

  2. How should we assess security and data minimisation in AI?

    UK Information Commissioner's Office. UK regulatory guidance, checked 11 September 2026. Jurisdiction-specific context, not individual legal advice or permission for a particular use.

Examples and exercises are illustrative unless attributed to a source. No independent expert review is claimed.

A correction, a counterexample or an experience worth sharing?

Join the conversation ↗