There is no universal number of customer interviews that validates a Product decision. Plan enough relevant coverage to learn from the contexts that could materially change the answer, then review interviews in batches. Stop when another relevant interview is unlikely to change the decision or the explanation you need — not merely when people start repeating the same words.

If the remaining uncertainty is about prevalence, behavior at scale, usability, causality, or precise pricing response, another interview may be the wrong next method.

The practical question is:

What would another interview need to teach us to be worth running?

This guide owns that stopping decision. For the complete design-and-run workflow, use Customer Interviews for Product Discovery.

Two different questions are hiding inside “How many interviews?”

Before research: how many interviews should we plan?

This is a planning question.

You need an initial batch and review plan based on:

the Product decision;
the research question;
who has first-hand relevant experience;
which contexts could materially change the answer;
how heterogeneous those contexts are;
how much depth you need;
the quality of the interview dialogue you expect;
what evidence you already have;
how you plan to analyze the material.

The output is not a final required N. It is a defensible starting plan with a point where the team will stop and synthesize.

During research: how do we know when we are done?

This is an evidence and decision question.

After each batch, ask:

What changed in our model?
Are important participant contexts still missing?
Are contradictions explained or merely ignored?
Do we know the mechanism, or only the topic?
Is the evidence deep enough for the Product decision?
Would another interview plausibly alter what we do?
Does the remaining uncertainty now belong to another method?

The output is one of:

STOP AND DECIDE
RECRUIT MORE — SAME CONTEXT
RECRUIT — MISSING CONTEXT
SPLIT / RE-SCOPE THE SAMPLE
REPAIR THE RESEARCH DESIGN
SWITCH OR ADD A METHOD

Treat planning and stopping as separate jobs. A number chosen before the first interview should not become a quota the team finishes mechanically after the research question has changed.

Why a fixed interview count is misleading

A count such as “five interviews are enough” hides the variables that determine whether the sample is actually useful.

Five conversations with people doing nearly identical work on the same decision may carry substantial information. Five conversations spread across different roles, workflows, company contexts, lifecycle stages, and current solutions may be five different research contexts rather than one coherent sample.

A narrow question such as:

Where do first-time workspace admins get stuck while connecting their first data source?

is very different from:

What problems do finance teams have?

The second question can span multiple jobs, decision rights, workflows, tools, and constraints. Adding more interviews to a question that broad may simply produce more unrelated observations.

Interview count is therefore a planning input, not a certificate of truth.

Customer Interview Sample Planning Canvas

Use this before recruiting. Its purpose is to make the rationale inspectable, not to calculate a magic total.

# Customer Interview Sample Planning Canvas

Product decision:
[What choice will this evidence inform?]

Research question:
[What do we need to understand?]

Why interviews:
[What can conversation, history, or context reveal that another method cannot?]

Participant definition:
[Who has first-hand relevant experience?]

Contexts that could materially change the answer:
- [...]
- [...]
- [...]

Expected heterogeneity:
[Where might workflows, roles, constraints, triggers, or decision rights differ?]

Depth needed:
[Topic discovery / mechanism understanding / switching story / buying process / etc.]

Known prior evidence:
[What do we already know from analytics, support, sales, prior research, or observation?]

Initial batch:
[Operational plan — not a final required N]

Batch review point:
[When will we stop to synthesize?]

Before adding more interviews, ask:
[What evidence would another interview need to add to be worth running?]

The canvas deliberately does not output a required sample size.

Use information power as a planning lens, not a formula

Malterud, Siersma, and Guassora proposed information power as a way to reason about adequate qualitative interview samples. Their model considers the study aim, sample specificity, established theory, dialogue quality, and analysis strategy.

The useful Product implication is:

A narrow question asked of highly relevant participants through rich interviews can carry more decision-relevant information per session than a broad question asked across mixed contexts.

That does not mean Product Discovery is identical to academic qualitative research, and it does not produce a universal interview count. Use information power as a discipline for explaining why a sample may need broader or narrower coverage.

A Product-friendly planning translation

Ask five questions:

Aim: Is the research question narrow enough to answer?
Specificity: Are participants tightly relevant to that question?
Prior knowledge: What credible evidence or theory already constrains the space?
Dialogue quality: Are interviews likely to produce concrete histories and mechanisms rather than shallow opinions?
Analysis: Are you looking for a focused mechanism or trying to map broad variation?

The weaker and broader these become, the less defensible a tiny sample becomes. But the answer is still not a formula.

Heterogeneity matters only when it can change the decision

Do not create segments merely because demographic or firmographic labels are available.

A difference matters when it could plausibly change:

workflow;
constraint;
motivation;
trigger;
decision rights;
current solution;
consequence;
adoption behavior.

For a buying-process question, end user versus procurement owner may be a real split. Two job titles that perform the same workflow under the same constraints may not be.

The practical rule is:

Heterogeneity matters when it creates different mechanisms relevant to the Product decision.

If the research population contains materially different mechanisms, “more interviews” may be the wrong instruction. You may need to split the sample or narrow the research question.

Higher stakes require stronger evidence design, not automatically more interviews

The higher the cost of being wrong, the stronger the evidence burden should usually be.

That can mean:

broader relevant context coverage;
deeper interviews;
explicit investigation of contradictory cases;
analytics;
observation;
experiments;
financial modeling;
legal or expert input;
sales or support evidence;
research triangulation.

For a large, hard-to-reverse platform investment, another block of interviews may add less value than the right non-interview evidence.

Higher stakes therefore mean:

stronger evidence design

not:

larger interview N by default.

Stop-or-Recruit-More Decision Board

Use this after each batch. Move through the gates in order.

Gate 1 — Is the Product decision and research unknown still clear?

Ask:

What decision are we trying to inform?
Which unknown can interviews actually reduce?
Did the research question drift after what we learned?

If no → SPLIT / RE-SCOPE or REPAIR.

More interviews against a vague or moving question are unlikely to fix the research design. Rewrite the decision and unknown first. If the full study design is the problem, return to the Customer Interviews for Product Discovery workflow.

Gate 2 — Are the relevant participant contexts represented?

Look for decision-relevant gaps such as:

only end users when the buying process matters;
only successful adopters when non-adoption or churn matters;
only SMB workflows when enterprise workflow is in scope;
only one current solution when alternatives may change the mechanism;
only one role when decision rights differ.

If no → RECRUIT — MISSING CONTEXT.

Do not multiply the sample by every available demographic label. Add a context only when it could materially change the answer.

Gate 3 — Are new interviews still changing the model?

A useful new interview may reveal a:

workflow;
trigger;
constraint;
consequence;
role;
alternative;
contradiction;
mechanism.

If yes → RECRUIT MORE — SAME CONTEXT when the current context is still generating decision-relevant novelty.

Do not recruit “more” generically. State what kind of evidence you expect the next sessions to add.

Gate 4 — Are topics repeating while the mechanism remains shallow?

Suppose everyone mentions “setup permissions,” but the team still cannot explain:

when the problem occurs;
what causes it;
why some people recover and others fail;
which role controls the permission;
what consequence changes behavior.

That is repetition without sufficient understanding.

If yes → REPAIR / DEEPEN.

Possible causes include:

surface-level questions;
weak probes;
participants without first-hand experience;
interviews that are too broad or too short;
pitching instead of investigating;
coding labels without reconstructing the story.

Use Customer Interview Questions That Reveal Real Behavior when wording and probing are the problem.

Gate 5 — Do contradictions map to materially different contexts?

When one interview conflicts with the emerging explanation, ask whether the difference maps to:

role;
workflow;
lifecycle stage;
current solution;
constraint;
trigger;
decision rights.

If yes → SPLIT / RE-SCOPE.

You may have combined two research populations that should not be averaged together.

If the contradiction does not map cleanly to context, run targeted recruitment or deeper interviews to investigate it. Do not use majority vote across interviews.

Gate 6 — What uncertainty remains?

If the main remaining question is:

prevalence or rate → use quantitative measurement or an appropriately designed survey;
behavior at scale → use analytics when available;
task usability → use usability testing or observation;
causal impact → use an appropriate experiment or quasi-experimental design;
exact pricing response → use pricing research, behavioral tests, sales evidence, or other fit-for-purpose methods;
cross-source consistency → use research triangulation.

If another method fits the uncertainty better → SWITCH OR ADD A METHOD.

Gate 7 — Would another relevant interview plausibly change the Product decision?

Write down the answer before scheduling anyone else.

If you can state:

“The next interview could change our decision if it shows X in context Y.”

then another interview may be justified.

If you cannot state what another relevant interview needs to teach you, and the remaining uncertainty is acceptable:

→ STOP AND DECIDE.

Stopping does not mean “truth proven.” It means the expected value of another interview is now lower than acting, switching method, or carrying the remaining uncertainty explicitly.

Evolve the evidence tracker from “new topic” to “new understanding”

Use a tracker that preserves both novelty and depth.

# Customer Interview Evidence Tracker

Decision:
[What Product decision are we informing?]

Key unknown:
[What remains uncertain?]

Relevant participant contexts:
[Roles / workflows / situations that could materially change the answer]

For each session capture:
- Relevant context
- New topic
- New dimension / mechanism
- Contradiction
- Existing pattern strengthened
- Decision impact
- Coverage gap
- Next action

Batch review:
- Which topics are repeating?
- Which mechanisms are now better understood?
- Which contradictions remain unexplained?
- Which decision-relevant contexts are missing?
- Which assumptions are still being treated as facts?
- What would another interview need to teach us?
- Is another interview still the best method?

A new topic means a new issue appeared.

A new dimension / mechanism means an existing issue became materially better understood.

That distinction matters because a team can stop hearing new topics while still learning important new reasons, conditions, consequences, and variation.

The Insight Clustering Helper can help compare observations across sessions without forcing them into a single saturation score.

Code saturation and meaning saturation are not the same thing

Hennink, Kaiser, and Marconi distinguished between two useful ideas:

Topic or code coverage

Have the main issues appeared?

Meaning or mechanism depth

Do we understand the dimensions, nuance, variation, and reasons behind those issues deeply enough?

Their empirical study reached these forms of saturation at different points in one specific research population. Those study-specific counts are not Product Management thresholds.

The practical lesson is the distinction:

You can have “heard the issue” without understanding why it happens, when it matters, who experiences it differently, or what consequence drives behavior.

So “the interviews are repetitive” is not enough. Ask whether the repetition reflects deep understanding or only a stable list of labels.

Saturation is not one universal research endpoint

“Saturation” is used in several ways across qualitative research, and its appropriateness depends on the methodology.

Braun and Clarke, for example, argue that simple “no new themes emerge” saturation logic is not conceptually appropriate for every form of thematic analysis. CraftUp does not need to turn a Product team into qualitative-methods academics, but it should avoid using the word as a scientific-looking certificate.

For Product Discovery, prefer precise statements such as:

no new decision-relevant contexts are appearing;
the current explanation has stopped materially changing;
major contradictions have been investigated;
the remaining uncertainty is acceptable;
the remaining uncertainty belongs to another method.

A useful CraftUp shorthand is decision-relevant redundancy: additional interviews in the relevant context are no longer changing the explanation or Product decision enough to justify their cost.

That is a practical Product concept, not a universal academic definition of saturation.

Interview quality is part of sample adequacy

A sample of weak interviews does not become adequate merely because the count grows.

Watch for:

participants without first-hand experience;
leading questions;
hypothetical answers replacing concrete behavior;
shallow probes;
pitching;
missing context;
one interview trying to cover too many jobs;
notes too weak to preserve evidence;
a guide that changed so much that early and late sessions no longer answer the same research job.

When interviews keep producing shallow repetition, ask:

Is the sample actually sufficient, or is our instrument only capable of producing shallow answers?

If question quality is the issue, repair it before increasing N.

Do not confuse repeated statements with independent evidence

Ten people repeating the same opinion is not automatically stronger than several detailed accounts of actual behavior.

Ask what kind of evidence is recurring:

remembered preference;
current workflow;
observed workaround;
purchase or switching event;
measurable outcome;
consequence described in context.

Also ask whether the observations are genuinely separate contexts. Ten participants repeating the same internal sales script, account policy, or shared organizational narrative may not be ten independent mechanisms.

Do not over-formalize “independence” statistically. The point is to preserve provenance and context.

Keep Evidence / Inference / Assumption / Unknown / Decision separate.

Repetition does not equal prevalence

If a large share of your purposively recruited interview sample mentions a problem, the defensible statement is:

it appeared frequently in this qualitative sample.

It is not automatically:

the same share of the market has the problem.

Likewise, hearing an issue in every interview does not prove universality, and not hearing it in a small sample does not estimate zero prevalence.

If prevalence matters, switch to appropriate quantitative evidence.

The “five users” usability rule does not answer this question

The famous small-N guidance from Nielsen Norman Group is about qualitative usability testing: observing people interact with an interface to uncover usability problems.

It is not a universal sample-size law for:

generative discovery;
market understanding;
buying-process interviews;
JTBD research;
pricing;
segmentation;
broad problem discovery.

The same N cannot be transferred across methods simply because both involve talking to users.

The rule to keep is:

sample adequacy follows the research question and method.

Hypothetical decision trace A — narrow question, coherent context

A PM wants to understand:

Why do newly invited workspace admins fail to connect their first data source?

The team recruits recent first-time admins who attempted the same core workflow.

Early batch

The first sessions reveal several failure modes: unclear credentials, missing permission, and uncertainty about which data source to choose.

Decision: RECRUIT MORE — SAME CONTEXT.

Reason: the explanation is still changing and new mechanisms are appearing.

Later review

New sessions repeat the same top-level issues. But one participant contradicts the apparent permissions story: their organization requires a separate security approval before an admin can obtain the needed access.

Decision: RECRUIT — MISSING CONTEXT.

Reason: the contradiction points to a materially different setup context that could change the Product solution.

Targeted follow-up

After covering that context, the team can explain the main mechanisms and when they differ. The remaining question becomes:

How common is each failure mode across all new admins?

Decision: SWITCH METHOD.

Reason: prevalence is now the uncertainty. Product analytics is a better next source than more generative interviews.

The stopping logic was never “we reached a particular N.”

Hypothetical decision trace B — broad question, falsely mixed sample

A team asks:

Why do finance teams struggle with monthly reporting?

Its sample mixes:

analysts;
finance managers;
CFOs;
SMBs;
enterprise organizations;
spreadsheet-heavy workflows;
ERP-heavy workflows.

After several interviews, the evidence looks contradictory.

One person cares about reconciliation work. Another cares about executive review. Another is blocked by data access. Another is focused on forecast governance.

The wrong response is:

schedule many more random finance interviews.

The better diagnosis is:

SPLIT / RE-SCOPE.

The original question and sample combine materially different workflows and decision rights. Narrowing the research job — for example to the month-end reconciliation workflow for analysts in a defined operating context — creates a sample with higher information value.

Contradictions are a branch, not noise

When a participant contradicts the current model, inspect the contradiction before deciding it is an outlier.

Ask:

Is this a different context?
Is this a different role looking at the same workflow?
Is this a different lifecycle stage?
Is this a different prior solution?
Is this a genuinely conflicting behavior inside the same context?

Then choose deliberately:

targeted recruit;
split the sample;
deepen the interview;
preserve unresolved contradiction;
use another method.

Never reduce qualitative interviews to majority vote.

What about pricing research?

Interviews can reveal:

how buyers describe value;
which alternatives they compare;
how budgets and approvals work;
previous purchasing behavior;
which trade-offs matter in a buying decision.

They are weaker evidence for estimating a precise market-wide willingness-to-pay distribution.

Depending on the decision, combine qualitative buying interviews with pricing experiments, sales evidence, historical behavior, or quantitative methods designed for the job.

Do not use an arbitrary interview minimum to create false precision around price.

What about a completely new product idea?

Do not ask:

How many interviews prove the idea?

No interview count converts a Product idea into a validated fact.

Break the uncertainty apart:

Do relevant people actually encounter the problem?
What do they do today?
What triggers the situation?
What consequences or constraints matter?
Which contexts create different mechanisms?
What alternatives already compete for the job?
Which assumptions still need another method to test?

Interviews can reduce specific uncertainties. The responsible next step may be another research round, a prototype, analytics, a smoke test, an experiment, a pricing test, or a decision not to proceed.

A final review before you schedule another batch

Ask the team:

What Product decision are we now able to make?
Which observations support it?
Which parts are inference?
Which important assumptions remain?
What evidence contradicts our preferred explanation?
Which participant context could still change the answer?
Are we repeating topics without understanding mechanisms?
What exactly would another interview need to teach us?
If another interview would teach us little, which method fits the remaining uncertainty better?
Is the decision reversible enough to act while carrying the remaining uncertainty?

If the answer is still “we just want more confidence,” define the uncertainty before adding interviews.

Continue the research workflow

If the Decision Board says REPAIR, use Customer Interview Questions That Reveal Real Behavior for question and probe repair, or the broader Customer Interviews for Product Discovery workflow for study design.

If it says RECRUIT, use the User Interview Script Generator to create the focused guide for the remaining unknown and participant context.

If it says SWITCH OR ADD A METHOD, use the research triangulation guide to choose evidence sources without pretending they answer the same question.

If the research job is specifically reconstructing switching and progress decisions, use the JTBD interview framework.

The goal is not to finish a ceremonial quota. It is to justify why the next unit of research effort is an interview — or why it should be something else.

Method notes and sources

The methodological concepts above are adapted carefully rather than treated as Product formulas:

Malterud, Siersma & Guassora, “Sample Size in Qualitative Interview Studies: Guided by Information Power” — Qualitative Health Research.
Hennink, Kaiser & Marconi, “Code Saturation Versus Meaning Saturation: How Many Interviews Are Enough?” — Qualitative Health Research.
Braun & Clarke, “To saturate or not to saturate? Questioning data saturation as a useful concept for thematic analysis and sample-size rationales” — Qualitative Research in Sport, Exercise and Health.
Nielsen Norman Group, “Why You Only Need to Test with 5 Users” — qualitative usability-testing guidance.

These sources do not supply one universal Product interview count. They help explain why sample adequacy depends on the research purpose, participant relevance, dialogue quality, analytical depth, and method.