AI personas are not customer evidence: a practical research sorting exercise
An AI-generated customer can give you an articulate answer in seconds. That answer may even mention the exact frustration your product promises to solve. The tempting next step is to put it in a slide labelled “customer insight”.
Before doing that, ask where the information came from. A synthetic user is a simulated participant generated by software. It can help you explore possibilities, but its response is not a conversation with a person who has used your product.
The practical challenge is keeping real observations, your interpretation and generated suggestions distinguishable as they move through research notes and product decisions.
Give every claim an origin
Use three plain labels: observed, interpreted and generated. These describe the origin of a statement, not whether it is automatically true.
- Observed: a participant completed a task, said something or encountered a problem. Keep a reference to the original record.
- Interpreted: you proposed an explanation for what happened. State what supports it and what else could explain it.
- Generated: an AI produced a possibility or simulated response. Turn it into a question to investigate.
An observed comment can still be misremembered or drawn from an unrepresentative sample. A generated suggestion can still inspire a useful question. The labels help the team decide what each item can support.
Try the evidence sorting exercise
These three statements come from a fictional mentoring platform. Decide how you would label each before opening its explanation.
“One participant reopened the calendar twice during the booking task.”
Observed. If the session recording supports the note, it describes a behaviour. It does not explain why the calendar was reopened or establish how common the problem is.
“The participant may have been checking whether the time suited their working hours.”
Interpreted. This is a possible explanation. Ask what they were checking and examine alternatives, such as a confusing time-zone label.
“As a busy mentor, I would definitely pay for automatic scheduling.”
Generated if an AI produced it while acting as a mentor. Do not turn it into a customer quote or evidence of willingness to pay. Investigate how actual mentors schedule today and what they have paid for.
Use simulation to improve the next question
A useful role for AI is rehearsal. Give it a fictional scenario and ask where an interview question is leading, vague or difficult to answer. You are testing the wording, not validating demand.
For example, “Would our brilliant new scheduling tool save you time?” invites agreement. A more useful starting point is: “Tell me about the last time you arranged a mentoring session. What did you do first?” Follow the person’s account rather than forcing it into the feature you want to build.
AI can propose missing situations, such as shift work or a shared device. Treat these as recruitment and design questions. Do not assume a generated persona accurately represents women in tech, disabled users or any other group.
A research note you can reuse
- Question: What are we trying to learn?
- Origin: Where did this information come from?
- Record: Which session, event or document supports it?
- Interpretation: What might it mean, and what remains uncertain?
- Decision: What does this justify doing now?
- Next check: Who or what could challenge the conclusion?
Illustrative note: “A participant changed the session time after reopening the calendar. We suspect the displayed time zone was unclear. Check the recording, ask a follow-up question and observe the revised label in the next session. This does not yet justify building a new scheduling system.”
Keep AI-assisted synthesis traceable
Synthesis means bringing research together to look for patterns. If AI helps summarise real interviews, keep source references attached to claims, verify quotations against the original and look for contradictory accounts. An elegant summary can omit the participant who did something different.
Use only tools and data handling arrangements approved for the research. Removing a name does not necessarily remove identifying details. Plan consent, access and retention before transferring interview material.
Synthetic data can be useful for testing a system with hypothetical inputs. That use is different from claiming to have learned what customers want. Keep the purpose visible so a test case does not become a customer finding by accident.
Make one decision easier to challenge
Before your next product meeting, pick one slide containing a customer claim. Add its origin, source and limitation. If the claim is generated, replace “customers want” with a testable question.
You do not need to remove AI from your workflow to make the evidence clearer. You need to give the team enough information to question the conclusion and choose the next real-world check.
You May Also Like
Celebrating International Creativity Day: Unlocking Innovation in the Tech World
Unleashing Creative Innovation: “Design the Thought”
