Skip to contentSkip to main content
← Back to all posts

From Customer Interview Notes to Product Decisions: An Evidence-First AI Workflow

Turn customer interview notes into traceable product decisions without inventing quotes or confusing feature requests with evidence. Includes a research template and five tests.

Share by email

You interview five prospective customers. Three mention difficult onboarding, one asks for an integration, and another says the product is complicated. An AI summary turns this into “Customers want simpler onboarding.” But which customers struggled, what were they trying to accomplish, and did anyone request the proposed solution?

AI-assisted interview analysis should preserve evidence, not turn a transcript into a roadmap. The useful output is a chain from an observed problem to a testable hypothesis, a reviewed decision, and a bounded next action.

This workflow works with ordinary Markdown. Phasoric can support local knowledge capture, reviewed decisions through Strata, change signals through Pulse, and approval-first follow-ups through Forge. It does not replace interview consent, research judgment, or decision authority.

1. Preserve observations before generating conclusions

Each interview is an evidence source, not a vote. Record the research question and relevant participant context: role, workflow, environment, and desired outcome. Use pseudonymous IDs where possible.

Make one source note per interview. Include its date, interviewer, consent restrictions, and whether the source is a transcript, live notes, or a later recollection. These carry different certainty. Do not present reconstructed wording as a direct quotation.

Define an observation as a specific behavior, incident, or stated difficulty linked back to the source. “P03 described retrying an expired invitation” is more defensible than “Onboarding is broken.”

Before asking a connected model to analyze material, confirm the data can be processed by that provider. Remove names, account identifiers, and unnecessary personal information. Local-first storage does not automatically make remote model processing local.

2. Make a portable evidence card

The template below is a suggested research practice, not an automatic Phasoric extraction format.

# OBS-017 — Invitation retry during setup
Participant: P03
Interview date: [actual date]
Research question: Where does initial setup stall?
Source: interviews/P03-notes.md#invitation
Capture type: Interviewer notes, not verbatim
Observation: Participant described retrying an expired invite.
Interpretation: Invitation expiry may interrupt onboarding.
Alternative: The invite may have been misconfigured.
Confidence: Provisional — one participant
Status: Needs corroboration
Owner: Research lead
Next check: Observe setup with a fresh participant

Separate the reported observation from your interpretation. A later usability test might reveal that the problem is reproducible, or that the participant used an old link. Both findings matter.

If a recording is deleted under the research retention policy, update the evidence status instead of implying the original recording remains inspectable. Derived summaries are not independent proof of their own claims.

3. Cluster by user job, not keyword frequency

AI clustering can group passages quickly, but shared words can hide different intentions. “Export” might mean regulatory recordkeeping to one customer, collaboration to another, and fear of vendor lock-in to a third.

For every proposed theme, ask:

  1. Who? Which roles and environments experienced the issue?
  2. What job? What did they need to accomplish?
  3. Which sources? What passages actually support the finding?
  4. What conflicts? Who completed the same job successfully?

Maintain a lightweight theme table with observation IDs, distinct participants, context, counterexamples, and unknowns. Five qualitative interviews are not a statistically representative market survey. A repeated concern may justify further discovery; it cannot establish market prevalence.

Distinguish a spontaneously reported difficulty from agreement with a leading question. An AI-generated synthesis that removes the counterexamples is less useful than the original notes.

4. Convert findings into testable opportunities

“Add Slack notifications” is a request for a solution, not a complete statement of need. The underlying job might be exception handling, ownership, or auditability. Shipping an integration immediately skips the most valuable question.

Use this opportunity matrix:

FieldQuestion
JobWhat outcome did the user need?
FrictionWhat observable step was difficult?
EvidenceWhich source observations support this?
AlternativeWhat else could explain it?
ConsequenceWhat impact did the participant describe?
UnknownWhat must be verified?
Next testWhat small experiment could reduce uncertainty?

Synthetic example: Several participants describe trouble resuming interrupted setup. The next step could be observing a restart with disposable accounts, not committing to a notification feature.

Prioritize by consequence, evidence strength, strategic fit, and reversibility. If conversion impact or frequency was not measured, label it unknown, rather than fabricating a number.

5. Separate the product decision from the synthesis

A research brief may recommend options. A decision record identifies which option an authorized owner accepted, why it was selected, and when it should be reviewed.

Phasoric's Strata decision model preserves reviewed choices with supporting evidence and distinguishes confirmed decisions from candidates. The weekly decision-review guide shows a source-linked decision register. A meeting note saying “we should ship this” is not automatically an approved roadmap decision.

Record decision owner, confirmation date, source links, alternatives, constraints, lifecycle status, and review trigger. Keep rejected alternatives and rationale. A new teammate should be able to tell whether an idea was abandoned or simply never discussed.

In a fresh assistant session, ask: “What supports this decision, what evidence contradicts it, and what would trigger reconsideration?” Demand actual source paths. Missing authorized evidence must be reported as missing, not invented.

6. Review changes without silently rewriting the roadmap

Suppose an early team postpones bulk import because pilots need only single-file uploads. Later, enterprise interviews reveal recurring multi-file workflows. That change deserves attention, but the earlier decision may have been appropriate for the original audience.

Compare segment, time, scope, and approval state before declaring a conflict. New evidence does not automatically supersede a decision.

Phasoric's Pulse surfaces source-linked change and review signals in supported contexts. You can also use a dated manual register: list new observations, the prior decision they challenge, and who must review that decision. Pulse does not establish market demand or independently approve a new roadmap.

For an evidence-review approach that preserves disagreement, see resolving conflicting AI knowledge.

7. Turn confirmed decisions into reviewable next actions

A confirmed choice should yield a bounded task: run a usability test, build a disposable prototype, revise help text, or interview a different segment. Name the owner and completion criterion.

Forge supports previewable local note changes, checklist extraction, and draft follow-ups under an approval-first workflow. A proposed action is not a completed task or sent email. External mutations need a supported executor and explicit permission.

Before approving, inspect the target workspace, exact change, source evidence, and privacy impact. A customer follow-up draft must not reveal another participant's confidential statement just because both influenced a product decision.

8. A 40-minute review for each interview batch

Use a batch of three to six interviews, with one research question in scope.

  • 0–8 minutes: Check note completeness, consent, identities, and source links.
  • 8–18 minutes: Extract observation cards; label interpretations as hypotheses.
  • 18–26 minutes: Group by user job and add counterexamples.
  • 26–33 minutes: Propose opportunities with supporting and opposing evidence.
  • 33–40 minutes: Choose the next test, assign an owner, and record only explicitly approved decisions.

Keep four artifacts distinct: original notes, observation cards, opportunities, and decision or test records. A summary can link them; it should not erase their boundaries.

9. Five tests before trusting the workflow

Try these with synthetic interviews and independently known answers.

  1. Ambiguity: Two people use “sync” for different jobs. The model must not merge them into one feature request.
  2. Counterevidence: One participant succeeds where others struggle. The synthesis must preserve the counterexample.
  3. Uncertainty: A recollection lacks exact wording. The model must not invent a verbatim quote.
  4. Authority: A research recommendation conflicts with a confirmed decision. The assistant must distinguish them.
  5. Access: A confidential interview is outside the agent's authorized scope. Its details must remain unavailable.

If a test fails, correct capture, classification, retrieval, or permissions before expanding automation. The 12-question AI memory checklist provides a broader evaluation framework.

Frequently asked questions

Can AI analyze interviews without transcripts?

Yes, but the result is limited by the notes. Label paraphrases and uncertainty; do not manufacture quotations, chronology, or missing details.

Should a repeated request automatically enter the roadmap?

No. Similar language can refer to different jobs. Validate the underlying need and consequences before committing to development.

Does Phasoric automatically approve product decisions from interviews?

No. It offers knowledge, decision, change, and approval-oriented capabilities. This research workflow is user-directed; consequential choices require appropriate human review.

Begin with one research question

For the next three interviews, create evidence cards for the most consequential observations. Ask a second reviewer to locate each source. Then choose one assumption worth testing. Good AI-assisted research makes product decisions inspectable, not merely persuasive.

Frequently asked questions

Can AI analyze customer interviews without full transcripts?

Yes, but conclusions are limited by the notes. Clearly label paraphrases and uncertainty, and never generate verbatim quotes or chronology absent from the source.

Should every repeated customer feature request become a roadmap item?

No. Similar language can describe different jobs, and small qualitative samples do not establish market prevalence. Investigate the underlying problem first.

Does Phasoric automatically turn interviews into approved product decisions?

No. Phasoric supports knowledge capture, reviewed decision records, change signals, and approval-first follow-ups. This article's research routine is user-directed and decisions require human confirmation.