ARTICLE · Research

In-depth interviews: what the method proves and what it does not

Boris Kaptelov · 02.08.2026 · 5 min

An in-depth interview is not a single method with a fixed script but a family of qualitative interviews that differ in their degree of structure, the interviewer's role, the way they elicit a narrative, and the unit of analysis.

What it delivers: the substance of jobs and barriers, the language the audience uses to describe the problem, differences between cases, presumed behavioral mechanisms, and hypotheses to test.

What it does not deliver: market share, effect size, precise willingness to pay, or a prediction of actual purchase. Mention frequency within a corpus describes the corpus studied, not the population.

Below are short answers to the ten questions asked most often before and after a study, each with a link to a full discussion.

Summary: the question and the limits of inference

What to verify before investing in a product

A needs study must close four questions: what job the customer is doing, what they compare you against including the option of doing nothing, what stands in the way of buying and who takes part in the decision, and what they are willing to pay for. Not "how much" but "what for" — those are different questions.

Full discussion: customer needs research — what to verify before putting money in.

How many interviews you need

There is no single number. For a narrow question in a homogeneous group, nine to seventeen interviews are usually enough. When comparing segments, interviews are counted per group: thirty interviews across five segments is five studies of six each.

The "twelve" benchmark comes from the 2006 study by Guest, Bunce, and Johnson, where the basic elements of the themes surfaced after six interviews and reported saturation arrived around the twelfth. The authors cautioned that the result applied to their corpus and their coding approach. The caveat got lost in the retellings.

Full discussion, with a table of empirical tests: how much a study costs — six questions for your vendor.

Can you trust the customer's explanation

Four distinct objects are present in the conversation: the event, the memory of it, the participant's explanation, and the researcher's explanation. An explanation of a cause confirms the version of the cause the person has access to, not the cause itself.

Full discussion: the customer is not lying — they simply do not know why they chose.

What we call a need

The word covers at least nine different things: a goal, a task, a problem, a barrier, a constraint, a risk, a workaround, a preference, a requested solution. Each calls for a different response, and mention frequency does not equal importance.

Full discussion: the customer asks for a button — nine different things people call a need.

How to ask

The form of a question is not a channel for transmitting knowledge but an instrument that produces a particular type of data. Free narrative, episode reconstruction, critical incidents, comparison, artifacts, observation, and diaries yield different material about the same situation.

Full discussion: questions for an in-depth interview — seven techniques and the limits of each.

Who makes the decision in a corporate purchase

In a corporate purchase there is no single customer. There is a set of participation roles that shift as the decision unfolds, and needs that exist on several levels at once. Twenty interviews with happy users will not explain why the deal is stalling.

Full discussion: an organization is not a respondent.

What "testing the solution on customers" means

Four different procedures: testing the idea, testing the promise, testing the interface, and testing performance in a real context. Plus the pilot, which confirms that a specific configuration works and does not prove scalability.

Full discussion: "let's test the solution on customers" is four different procedures.

What a "customer job" is and the varieties of the approach

Under the name Jobs to Be Done — the job a product is hired to perform, JTBD for short — three traditions coexist: progress in circumstances after Christensen, desired outcomes after Ulwick, and switch interviews. They produce different types of results and are not interchangeable.

Full discussion: the "customer job" — three different schools that go by one name.

Can you trust analysis done by AI

"AI-assisted analysis" is not one operation but five: transcription, corpus organization, summarization, coding, and drawing conclusions. The first three save time at little risk. The last one does not save work — it trims accountability.

Full discussion: "the analysis was done with AI" is not a description of a procedure.

What to do with the report

Between a finding and a decision there is a separate step: priority, cost, timeline, owner, and a verification criterion. One finding unfolds into three different documents for three audiences. And every finding must trace back to a specific interview fragment.

Full discussion: the research report — how to carry findings through to a decision.

How we work with this

One narrow question in one segment is a five-day diagnostic: we frame the decision, run about ten interviews, check the findings against your data, and hand over a position on Friday.

Comparing segments, mapping a pricing contour, or a corporate purchase with multiple decision participants is a two-month project.

Reviewing someone else's study or a vendor's proposal takes a few days.

In every format, we put on record before the work begins what the study will show and what it will not. That determines which decision you will be able to make on the results.