A founder can leave a conversation with ten encouraging comments and almost no useful evidence. The customer was polite. The idea was easy to praise. Nobody examined the last time the task actually happened.
An interview becomes more informative when it reconstructs a recent episode: what triggered the work, what the person tried, where progress stopped, what followed, and who made the relevant decisions. The interviewer needs curiosity strong enough to survive an answer that weakens the product idea.
This chapter of The AI Software Factory, within the SalarsNet AI section, describes a proposed customer-discovery method. It is concerned with learning about behavior and constraints. Payment validation addresses the separate question of accepting a defined commercial offer.
No customer interview, transcript, or Salars discovery result is reported here. Illustrative questions and episodes are examples of how an investigation could proceed.
Write the uncertainty before writing the questions
An interview should answer something that can change a decision. “Learn about customers” is too broad to guide a small discovery effort.
A useful uncertainty might be whether independent sellers repeatedly lose time matching supplier item codes to internal records. Another might be whether the owner, rather than the employee doing the work, considers that burden worth paying to reduce.
Write what you already know and how you know it. A public post may support a reported problem. A founder’s operating experience may support one local episode. Neither establishes how broadly the pattern occurs.
Then identify competing explanations. Perhaps mismatches arise from unclear intake labels. Perhaps an existing tool solves them. Perhaps the work is infrequent and inexpensive. Interview questions should make those explanations visible instead of steering every answer toward new software.
The pain-qualification chapter explains the problem brief. The interview turns its unknown fields into questions about actual work.
Recruit for the task and role
A convenient participant may not be a relevant participant. Friends who like technology can provide thoughtful reactions while having no experience with the task or purchasing process.
Define recruitment conditions around the work. A hypothetical supplier-data study might seek people who recently processed supplier files and people responsible for approving tools used in that process. Their roles could overlap in a small business or differ in a larger one.
Avoid recruiting only people who already agree the problem is serious. Include people whose present method appears to work. They can reveal an adequate alternative or conditions that prevent the difficulty.
Record how participants entered the study. A trusted introduction, an existing business relationship, and an open invitation can produce different groups. The recruitment path limits how broadly the findings should be interpreted.
The UK Government Digital Service recommends planning interviews around research questions, recruiting current or likely users, and using open, neutral prompts focused on actual examples. Its guidance concerns service research; this article applies that discipline to commercial discovery without claiming that it guarantees demand. Using in-depth interviews.
Explain the conversation and obtain appropriate consent
The participant should understand who is asking, why, what will be collected, and how the information will be used. Recording requires a clear agreement; a person accepting a conversation has not necessarily accepted a recording or external transcription service.
GDS’s consent guidance identifies purpose, collection, use, sharing, recording, retention, and voluntary participation as information participants should understand. It also recommends appropriate expert review of the research documentation. This is a useful practice reference, not a substitute for requirements applicable to a particular jurisdiction or organization. Getting informed consent.
Keep the collection proportionate. A workflow can often be discussed with redacted examples. The researcher may not need customer names, account credentials, or a complete transaction history.
Make it easy for the participant to skip a question or stop. If they appear uncomfortable showing a record, return to a description or a harmless sample. Discovery should not create pressure to disclose information beyond the agreed purpose.
The consent and data-handling record belongs with the research evidence, outside the finished article. Published examples should not expose a participant’s identity or private information without appropriate authorization.
Reconstruct the most recent episode
Begin with a concrete occurrence. Ask when the task last happened and what initiated it. Follow the sequence from the person’s perspective.
For the hypothetical supplier-file study, the interviewer might ask: what did you receive, what did you need to do with it, and what happened next? If the participant says they cleaned the file, ask what cleaning involved. If they say it took too long, ask what work occupied that time.
Do not supply the answer. “Was it frustrating because the columns were inconsistent?” introduces both an emotion and a cause. “Which parts needed changing?” leaves the cause open.
If a permitted artifact is available, use it to clarify the sequence. The record may show that the task was easier or harder than remembered. Treat the artifact as evidence with its own limits; one file does not establish every occurrence.
Ask what happened after completion. Did another person review the result? Was a correction needed? Did the output enter a live system? The downstream handoff may contain the real problem even when the initial task looks straightforward.
Distinguish fact, estimate, and interpretation
Participants can know some details precisely and estimate others. A recent file timestamp may establish when the task occurred. A remembered duration may be approximate. A claim about why the process fails may be an interpretation.
Preserve those distinctions in notes. “Participant estimates about an hour” is more accurate than “task takes one hour.” “Owner believes supplier inconsistency causes errors” differs from a verified explanation of the error mechanism.
Ask how an estimate was made when it matters. Did the person time the task, infer it from the workday, or include interruptions? The answer can change the meaning of a cost calculation.
Do not interrogate every ordinary statement. Focus clarification on consequential claims that determine the product decision. A conversation should remain intelligible and respectful, not become a cross-examination.
Afterward, separate observed artifacts, participant reports, and researcher inferences in the ledger. That structure allows later evidence to revise one field without rewriting the whole account.
Ask about alternatives and past attempts
A problem’s current workaround is part of the evidence. Ask what the person uses now, why they chose it, and what they have tried previously.
A spreadsheet may persist because it is flexible and inspectable. A paid tool may have failed because setup was difficult. A consultant may provide judgment that software cannot easily replace. These explanations shape the opportunity.
Ask what worked as well as what failed. If a feature solved the problem for one group, inspect whether the same conditions apply to the proposed segment. A discovery process that ignores adequate alternatives can turn a training gap into a new application unnecessarily.
Past purchases can reveal category spending and decision conditions. They do not establish willingness to buy your offer, but they clarify the buyer’s expectations and approval process.
The reuse-before-build chapter examines solution search. Interviews help identify which alternatives deserve inspection and why customers have accepted or rejected them.
Learn who owns the consequence
The person doing the task may experience inconvenience while someone else bears the cost of a mistake. An employee might care about speed; an owner might care about inaccurate listings; a support worker might need the history to resolve a dispute.
Ask who uses the output, who reviews it, and who is accountable when it is wrong. These relationships affect the product’s acceptance criteria and action boundary.
For a hypothetical intake-record tool, a fast generated description may be less valuable than a reliable link between the item, photographs, and condition evidence. The buyer’s priority can differ from the feature the interviewer expected to sell.
Also ask who can approve a tool and what a normal purchase requires. Keep that inquiry separate from asking the participant to buy during the discovery conversation. Mixing the two can make participants defend their earlier praise rather than explain their ordinary decision process.
A clear role map can be brief. Its purpose is to prevent the product from optimizing one person’s work while creating hidden obligations for another.
Leave room for silence and disagreement
An interviewer who fills every pause can answer their own question. A brief silence gives the participant time to remember details and choose their words.
When an answer conflicts with the hypothesis, ask for the episode behind it. If the participant says the task is easy, explore what makes it easy. Their process may reveal a useful alternative or a condition that narrows the segment.
Do not debate the participant into acknowledging pain. They are describing their work, not grading the founder’s idea. A defensive conversation produces weaker evidence and can damage trust.
Clarify ambiguous language without converting it into your terminology too quickly. “We reconcile it” might mean matching rows, checking totals, obtaining approval, or resolving missing information. Ask what the person actually does.
An unexpected answer can expose the missing condition. An interview should have a direction while retaining enough flexibility to discover a different bottleneck.
Analyze episodes before counting themes
After the conversation, reconstruct the task sequence while the context is fresh. Identify the trigger, inputs, steps, interruption, workaround, consequence, roles, and unresolved questions.
Then compare episodes across participants. Similar vocabulary does not guarantee the same mechanism. One person may complain about “inventory errors” caused by intake records; another may mean delayed supplier updates. Keep the distinction until evidence supports merging them.
A theme count should name its unit. Three comments by one participant do not equal three independent customers. Several employees from one organization can supply multiple perspectives without representing several purchase decisions.
Include counterexamples. A participant whose current method works can help define the boundary of the problem. A participant who experiences the problem but would not change their process can reveal adoption costs or low consequence.
Report the strongest evidence and limits together. A small discovery round can support a local problem pattern and a better next question. It cannot establish population prevalence without a suitable sampling method.
Inspect your own incentives
The interviewer often wants the idea to survive. They may remember enthusiastic answers more clearly than quiet objections, or ask follow-up questions only when an answer supports development. A written analysis rule can reduce that selective attention.
Before reviewing a session, identify the evidence needed for each consequential claim. If the claim is recurrence, look for actual occurrences or an explicitly labeled estimate. If the claim is buyer value, look for the consequence and decision role. If the claim is an inadequate alternative, examine what the current method fails to do under the participant’s conditions.
Ask a reviewer to find the passage that most weakens the preferred interpretation. They should explain the consequence for scope or next action. The purpose is to expose a neglected condition, not require criticism for its own sake.
Retain uncertainty when two interpretations remain plausible. A seller’s difficulty might arise from confusing instructions or from a genuinely unsupported workflow. The next observation should distinguish those explanations instead of adding more interviews that ask the same broad question.
If the discussion guide changes, record why and preserve the earlier version. New questions can improve discovery, but the answers should not be pooled as if every participant received identical prompts. This matters especially when a later prompt mentions the product’s proposed benefit.
A modest record of recruitment, questions, changes, and interpretations makes the evidence more reusable. It also gives the founder a way to reconsider the idea without relying on a memory shaped by hope.
Use AI summaries as a draft of the evidence
An AI transcription or summary tool can help organize a permitted conversation. It can also omit uncertainty, merge roles, or turn an approximate statement into a precise claim.
Review important claims against the original authorized record. Check who said what, whether a statement was a question or answer, and whether the summary preserved conditions. Keep the source available within the agreed retention and access rules.
Do not ask a model to invent customer personas or fill missing interview answers. A gap is a research result. Filling it with plausible language makes the record harder to distinguish from fiction.
If multiple researchers use summaries, calibrate their classification on a small set and record disagreements. Model agreement can be useful operationally, but it is not independent evidence when the models share the same incomplete transcript or prompt.
The next decision should remain grounded in traceable episodes. The summary is an aid to reading, not a substitute for what the participant actually supplied.
Turn the interview into a next decision
An interview round should end with a decision-relevant statement. Perhaps the problem exists only when a supplier changes item identifiers. Perhaps the current tool already handles it after configuration. Perhaps the user values the improvement but the buyer does not.
For each candidate, write what changed in the problem brief, what evidence supports the change, and what remains unknown. Name the next action: inspect a documented feature, observe another episode, present a bounded offer, or close the candidate.
Avoid a vague conclusion that customers want efficiency. It cannot guide scope or distinguish your product from every other business tool.
If a product concept is shown during the conversation, record that separately from discovery about current work. The concept can influence later answers. A participant’s reaction to the prototype should not be mixed with their earlier description as if both were spontaneous evidence of the problem.
The interview has succeeded when it improves the decision, including when the result is to stop. Its value is the clearer account of work and constraints.
What Would We Do at Salars?
A proposed Salars discovery round would start with one uncertain workflow claim and recruit relevant operators plus the purchasing role where distinct. A supplier-data or intake-record candidate would require recent episodes rather than general reactions to an AI app.
The team would explain the purpose, collect only needed information, and use permitted redacted records. Questions would reconstruct the last task, current workaround, consequence, alternatives, and approval conditions. The interviewer would keep product presentation separate from the account of current work.
Analysis would preserve participant reports, artifacts, and inferences as different evidence types. Counterexamples and working alternatives would remain in the ledger. AI summaries, if authorized, would be checked against the original record for consequential claims.
The result would be a revised problem brief and a concrete next decision. It might justify a bounded paid offer, a setup service using existing software, another observation, or closure. No interview or customer finding is asserted here.
A discovery conversation becomes useful when the founder can describe the customer’s work more accurately afterward, even if that accuracy makes the original idea less attractive.
Sources
- Government Digital Service: Using in-depth interviews, official guidance on planning, neutral questions, and actual examples.
- Government Digital Service: Getting informed consent for user research, official guidance on explaining collection, recording, use, sharing, and participation.
Sources inspected October 7, 2026. Commercial application and analysis workflow are proposed. Illustrative questions are not quotations from actual customers, and no interview study was executed for this article.
Loading comments…