Running and Synthesizing Customer Interviews

1. Blog
2. Strategy & Positioning

By Saroj Jha · October 5, 2026 · 9 min read

Good customer interviews ask about real, recent events. Good synthesis counts people, not quotes, and weights each problem by what it costs them. The five steps below show how to run both, and when to stop, with one example company.

Interviews are the cheapest research a startup can run, and the easiest to misread. One vivid story feels like a trend; a polite "I would use that" feels like demand.

Synthesis turns pages of notes into a count you can act on: which problems came up across people, what each costs them, and whether you have heard enough. To choose which segments to interview, use the Customer Segmentation Playbook. To measure a pattern across a larger group, use the Survey Playbook.

Running case: Example Co. Fictional

The running case: Example Co. Example Co. is a fictional seed-stage B2B SaaS startup whose software automates the month-end close for finance teams at companies with 50 to 500 employees. At the point this post picks up its story, it has 22 paying customers: 14 single-entity companies and 8 multi-entity groups that consolidate several legal entities each month. First, the founders want to hear what the product changes, and for whom. Round 1 is 14 interviews with controllers and finance staff: six single-entity customers, five multi-entity customers and three lost prospects (one single-entity, two multi-entity). Round 2 adds four single-entity customers. Every count and quote below is fictional.

Step 1: Decide what you need to learn, and from whom

Write the decision first. "Pick which problem our first message leads with" is a decision; "learn about our users" is not. Then tag every interview with its segment. Without the tag, one kind of customer's problem looks like everyone's.

Plan enough interviews per segment. In a study of 60 in-depth interviews, "saturation occurred within the first twelve interviews, although basic elements for metathemes were present as early as six interviews"^[2]. (Saturation: new interviews stop adding new themes.) A 2022 systematic review found that studies using empirical data "reached saturation within a narrow range of interviews (9–17)", "particularly those with relatively homogenous study populations and narrowly defined objectives"^[1]. Those numbers hold for a group of similar people, so two segments need two samples. In my view, plan about ten interviews per segment. Read patterns from the fifth, and stop early only when three in a row add nothing new. Go to 20 or more only when you need the complete list of needs, for example before designing a survey.

Example Co., step 1 Fictional

The decision: which segment the message and first marketing hire focus on, and which problem the message leads with. Each interview is tagged single-entity or multi-entity, and customer or lost prospect: seven per segment. Enough to compare the two; probably not enough to call saturation. All fictional.

Do this

Write the decision at the top of your notes, and add a "segment" column before the first call.

Step 2: Ask about the last time, not the future

People describe what they did last week far better than what they might do next year. So anchor every question in a real, recent event: "Walk me through your last month-end close." Then follow the thread: what they did, what it cost in time or money, what they use today to work around it, and who paid.

Keep your product out of it until the end; once you pitch, people get kind. Treat "I would pay for that" as weak. What they already pay for, and the workaround they put up with, tell you more.

Selling an AI product? Ask about trust the same way, through past events. In my view, three questions earn their place: "Tell me about the last time a tool got a number wrong. How did you find out?" "Who checks its output before it goes to your boss?" "What did IT ask about where your data goes?" The answers show how much accuracy proof you will need, and what IT will ask before anyone signs.

What Intuit's filing shows

Some companies make watching customers a formal habit. In its 2017 letter to stockholders, Intuit wrote: "The process begins with an understanding of our customers and the unique environments in which they work. Every Intuit employee conducts 'follow-me-home' visits with our customers, immersing themselves in the challenges our customers face every day, and infusing our organization with fresh excitement and new ideas for solving them"^[3]. This is what the filing says, not an account of how Intuit makes its choices internally or how well the visits work. My narrower lesson: watch the problem where it happens, and you see workarounds people forget to mention.

Example Co., step 2 Fictional

The guide: Walk me through your last close. Where did it slow down? What did that cost: hours, weekends, a late board pack? What still lives in spreadsheets? Who else had a say when you chose a tool, inside or outside the company? Lost prospects also get: what did you do instead? No pitch until the last five minutes.

Do this

Rewrite every "would you" question in your guide as a "tell me about the last time" question.

Step 3: Code each interview the same day

Coding means tagging each thing a person said with a short theme name. Do it the same day, while you remember the context. For each theme, record:

• Severity: 1 = annoying, 2 = costs time, 3 = costs money or customers.
• Quote: one line in their own words.
• Workaround: what they do about it today.
• Would pay: yes, no or unclear (weak evidence, as above).

Use the same theme names across interviews. "Our close drags on" and "reconciliations eat the second week" are one theme; merge them before you count.

Example Co., step 3 Fictional

Single-entity customer 1's controller. Theme: "close ran long on spreadsheets", severity 3, because the board pack went out late and the team worked weekends. Quote: "We used to close in about two weeks. We close in days now, and I didn't have to hire." Workaround before: spreadsheet reconciliations and chasing other teams by email. Would pay: already does. The call was recorded and AI-transcribed with the controller's permission, and a founder checked the quote against the recording. All fictional.

Do this

Book 20 minutes after every interview for coding.

Free template

Run Steps 2 to 5 from one workbook

The free Customer Interview Kit has the recruiting message, a screener, the interview guide and a workbook that codes and counts your notes.

If you use AI to transcribe or summarize

In my view, AI notes come with two rules. Ask permission at the start, both to record and to run the recording through an AI tool. Then check every quote against the recording before it goes into your counts, a persona or your website.

OpenAI's model card for Whisper, its speech-recognition model, shows why. It warns that "the predictions may include texts that are not actually spoken in the audio input (i.e. hallucination)", and says: "we caution against using Whisper models to transcribe recordings of individuals taken without their consent"^[4]. This is what the model card says about that one model, not a test of other tools, and not a recommendation.

Step 4: Count people, weight by cost, and separate patterns from anecdotes

Count interviews, not mentions: five mentions from one person are one vote. A theme's share is the interviews that raised it divided by all interviews. Its priority is share times average severity, so a problem half your people have that costs money outranks one everyone has that only annoys.

Then label each theme. These are the AAJ skill's rules of thumb, not research findings:

Example Co., step 4 Fictional

Two themes are patterns. The costliest, "consolidation is still manual", is in 7 of 14 (50%), severity 3, priority 1.5. All seven are multi-entity, including both multi-entity lost prospects, and the product cannot yet solve it without custom work: a segment boundary. "Close ran long on spreadsheets" is in 9 of 14 (64%), severity 2.2, priority 1.4, and 7 of its 9 mentions (78%) are single-entity, where customers say the problem is gone. "Audit prep takes days" is emerging: 4 of 14, all single-entity, with the outside accountant chasing support by email. The most quotable interview, one controller's forecasting wish list, is an anecdote. The segment has not saturated, so Example Co. books more. All fictional.

Do this

Lay out a grid of people by theme, grouped by segment, before anyone writes a persona.

Step 5: Stop when new interviews stop adding new problems

Guest, Namey and Chen (2020) propose a simple saturation check built on "Base Size, Run Length, and New Information Threshold". In plain words: count the themes in your first few interviews, then ask what share of themes the next two or three interviews add. They propose two thresholds, "≤5% new information and no (0%) new information"^[5]. The practical version I recommend: five or more interviews in the segment, and nothing new in the last three. Then read the result:

• Enough: saturated, with at least one pattern. Act on the top one.
• Saturated, no pattern: the segment is too broad, or the problem is not theirs.
• Keep going: new themes are still arriving. Check again every two or three interviews.
• Too few: under five. Probe the emerging themes next.

Example Co., step 5 Fictional

Four more single-entity customers, 11 in the segment, with added questions on audit prep and outside accountants. New themes by interview: 1, 0, 1, 0, 1, 0, 0, 1, 0, 0, 0. The last three add nothing: saturated. "Close ran long on spreadsheets" is raised by all 11 (severity 2.5, priority 2.5), and the customers keep saying "we close in days now". "Audit prep takes days": 7 of 11 (priority 1.3). So the message leads with the solved problem: "Close the books in days, not weeks, without adding headcount." Eight of the 11 rely on an outside accountant, and 4 of the 10 customers first heard of Example Co. from one, so accountant partners become the channel to test. Multi-entity groups wait for consolidation. Next, the "very disappointed" survey in my product-market fit post sizes the pattern. All fictional.

Do this

Log how many new themes each interview added. Three zeros in a row, after five or more, means stop and decide.

How do you turn customer interviews into homepage copy?

To turn customer interviews into homepage copy, capture the exact words customers use, sort each phrase by where it belongs on the page, and pick the phrases the most people said.

Run the interviews for messaging

Run them the usual way: ask about a real, recent event, keep your product out of it until the end, and record with permission. For messaging, add: "What was happening when you started looking?", "What did you use before, and what was wrong with it?", "What nearly stopped you buying?" and "How would you describe us to a colleague?" Write answers down word for word. The same day, tag each phrase with a theme and a slot: problem, trigger, old way, outcome or objection.

Build a message bank

Keep one sheet with the phrase, who said it, and a tag for where it could go: headline, subhead, proof or objection.

Pick by count, not charm

Group close variants and count how many different people said each. In my view, a headline phrase needs at least three people behind it. A vivid line from one person is an anecdote.

Map it onto the page

Worked example: Example Co. Fictional

Example Co. automates the month-end close. Most of its single-entity customers say some version of "we close in days now", and one adds "I didn't have to hire." The headline becomes: "Close the books in days, not weeks, without adding headcount." A fictional objection, "we can't risk the audit trail", becomes the line beside the button.

Test it

Show the first screen to people who match your buyer for five seconds and ask what you do and for whom. Or ask three interviewees: "Is this how you'd describe it to a colleague?" If they reword it, use their words.

By stage

At pre-seed, counts are small, so treat the copy as a draft. At Seed, run to saturation in the segment you plan to win before you lock the headline. At Series A–B, refresh the bank quarterly from sales calls and win-loss interviews.

The common mistake is writing the headline from your favourite quote, in your language rather than theirs.

What changes by stage?

• Pre-seed — You are looking for the problem. Plan about ten interviews per candidate segment and do not write a persona yet.
• Seed — You are choosing a first segment and message. Run to saturation in the segment you plan to win.
• Series A–B — Use interviews to explain what your usage data shows, such as why one cohort churns, and add win-loss interviews (Win-Loss Analysis playbook).

What mistakes should I avoid?

• Building a persona from the most quotable interview.
• Counting quotes instead of people.
• Skipping severity, so a frequent annoyance looks like a positioning platform.
• Stopping because you are tired, not because the last three added nothing.