How should product teams learn to use AI during discovery?

Product discovery asks teams to understand users, problems, constraints and risky assumptions before committing to a solution. The evidence is often incomplete and open to interpretation.

AI can organise notes, suggest themes and help teams explore alternatives. It can also create a convincing synthesis that hides contradictory evidence or turns a generated idea into an apparent user need.

Product teams need role-specific practice that protects the discipline of discovery. Learning should strengthen their contact with evidence, not place another layer between the team and users.

Discovery decisions are made with incomplete evidence

Discovery brings together user research, service data, business goals, technical constraints and professional perspectives. Teams decide what the problem is, which assumptions are risky and what needs to be learned next.

AI may support parts of this work, but it does not know which interview comment reflects a rare yet important accessibility need. It may group similar language while missing differences in context. It can generate a plausible opportunity statement without evidence that users experience the problem described.

Product capability therefore includes knowing what the output represents. A summary is an interpretation of supplied material. A generated hypothesis is a proposition to test. Neither becomes user evidence because it is clearly written.

Generic AI exercises miss the discipline of discovery

General demonstrations are useful for showing how a tool summarises, classifies or generates ideas. They can help learners understand iteration and common limitations.

Discovery adds requirements that generic material rarely contains. Learners need traceable research evidence, conflicting views, incomplete information, inclusion considerations and a decision about what to investigate next. The quality of an output depends on the research method and intended use, not only the instruction given to the tool.

A role-specific exercise should preserve these conditions. It should ask learners to make product decisions and justify them, rather than reward the largest list of generated ideas.

Build capability around evidence and hypotheses

Start with a user problem or discovery question. Give learners an approved, fictional or safely prepared evidence set, including material that does not fit the dominant pattern.

Practice should require the team to:

• Define the problem without assuming AI is the solution.
• Decide whether the material and approved tool are suitable for the activity.
• Preserve links between themes and source evidence.
• Look for omissions, contradictions and minority experiences.
• Separate direct findings, human interpretation and AI-generated hypotheses.
• Apply product, research and accessibility standards.
• Decide what needs further user research or another form of validation.

Comparison is useful. Teams can conduct an initial human review, examine AI-supported synthesis and discuss what each surfaced or lost. Feedback should focus on evidence and decision quality, not whether the AI output matched a prepared answer.

Learners also need to decide when AI adds little value. A small, sensitive or highly contextual evidence set may be clearer to examine directly.

Protect user contact and multidisciplinary judgement

AI-generated personas or simulated users may help teams explore questions during ideation. They are not evidence about real people. Product teams should not replace research with synthetic responses that reproduce assumptions already present in the prompt or model.

Using raw research material requires approval, consent and appropriate privacy and security controls. Removing names may not make sensitive qualitative data safe. User researchers and information owners should shape the method before material enters a tool.

Multidisciplinary learning reveals different risks. Researchers can challenge weak claims, designers can examine inclusion and journey implications, engineers can test feasibility assumptions, and product managers can connect evidence to priorities. Relevant legal or risk expertise should join where the scenario requires it.

End the activity with a decision record. State the evidence, interpretation, hypotheses, uncertainty and next research action. This provides a practical artefact that can transfer into discovery work.

AI can make exploration faster and broader, but speed is useful only when the team retains traceability, user contact and responsibility for what it chooses to believe and do.

Example

A product team uses fictional interview notes about an account-recovery journey. An approved AI tool proposes themes and opportunity statements.

Learners trace each theme to the notes. They find that the summary has lost a minority accessibility need and combined two different causes of failure. The team restores those distinctions, labels new opportunity statements as hypotheses and plans questions for further user research. An accessibility specialist challenges an assumption about the preferred recovery channel.

The team uses AI as a synthesis and exploration aid without allowing it to manufacture evidence or replace contact with users.

FAQs

• Can product teams use AI-generated personas in discovery?

  Generated personas may help explore assumptions or prepare research questions, but they are not evidence about real users. Keep them clearly labelled and validate relevant hypotheses through appropriate research with the people affected by the product or service.

• Should AI analyse raw user-research notes?

  Only where the research method, consent, privacy, security, tool approval and data controls permit it. The team also needs source traceability and a review method that preserves contradictory or minority evidence. Anonymisation may not remove every risk.

• Which discovery task should a product team practise first?

  Choose a low-consequence task connected to current work, with approved material and a checkable output. Comparing an AI-supported synthesis with source notes is often more useful than open-ended idea generation because the reasoning can be examined.