Customer Discovery for Startups: From Hypotheses to Build, Pivot or Stop (2026) cover image
Photo by C D-X on Unsplash
Back to blog
Product

Customer Discovery for Startups: From Hypotheses to Build, Pivot or Stop (2026)

The whole customer discovery process in five steps: hypotheses with kill lines, problem interviews, a solution test, a commitment ask, and the build, pivot or stop decision most founders put off.

Koji

Koji Team

Product · · 9 min read

Customer discovery is the work of finding out, before you build, whether the customer, the problem and your answer to it are real. Most guides stop at the interviews. This one covers the whole process end to end: writing hypotheses you can prove wrong, the interviews that test them, the solution test, and the part founders skip, deciding on a date whether to build, pivot or stop. If you want the detail on running the interviews themselves, our reference guide Customer Discovery Interviews: The Complete Guide covers that.

The short version: write five or six hypotheses with a "proved wrong if" line for each, test the problem in 10 to 20 interviews, test the solution with a mockup, ask for a small commitment, and then make one of three calls. Build when the problem, the workaround spend and the commitments line up. Pivot when one hypothesis broke and the rest held. Stop when the problem itself did not show up.

Why the decision matters more than the interviews

Founders rarely fail because they skipped talking to customers altogether. They fail because the conversations never forced a decision. CB Insights' review of 431 VC-backed companies that shut down since 2023 found that 43% cited poor product-market fit, ahead of bad timing at 29% and unit economics at 19% (CB Insights, startup failure reasons). An older Startup Genome analysis of more than 3,200 high-growth tech startups found that 74% of high-growth internet startups fail because they scale prematurely, hiring and spending ahead of the evidence (Startup Genome, 2011).

There is also direct evidence that treating an idea as a set of hypotheses changes outcomes. In a randomised trial with 116 Italian startups, Camuffo, Cordova, Gambardella and Spina taught one group to state hypotheses and test them, and the other to rely on judgement. The hypothesis group performed better and was more likely to pivot (Management Science, 2020). A later replication across 759 firms in four trials found the opposite on pivots: founders trained this way pivoted less, and fewer pivots went with higher performance (CEPR Discussion Paper 15972). The two studies disagree on pivots, but agree on the part that matters here: founders who tested their ideas as hypotheses did better than those who went on judgement.

Step 1: Write hypotheses you can prove wrong

Steve Blank's customer development model starts here, and for good reason. Every startup idea rests on a handful of beliefs. Write them down as plain sentences, and next to each one write what you would have to hear to drop it. That second column is what most founders skip, and it is what turns interviews into a decision.

Table of six customer discovery hypotheses: customer, problem, current fix, value, buyer and solution, each with an example and what would prove it wrong.
Write the last column before the first interview, not after · Download or share
HypothesisExampleProved wrong if
CustomerOps leads at 20 to 100 person agencies feel this mostFewer than half of them had it in the last 3 months
ProblemOnboarding a new client takes them over a week of adminMost describe it as a day or less
Current fixThey use spreadsheets and email, and dislike itThey already use a tool they are happy with
ValueThey have spent money or staff time trying to fix itNobody has spent anything on it
BuyerThe ops lead can buy a tool without the founderEvery purchase needs the founder to sign off
SolutionA checklist that fills itself in would save daysNobody would swap their current fix for it

Two rules make this work. Put the riskiest hypothesis first: usually the problem, because nothing else matters if it is not real. And set the threshold now. If you decide what "proved wrong" means after the interviews, you will find a way to keep believing.

Step 2: Test the problem in interviews

This is the part most guides cover, so briefly. Talk to 10 to 20 people from your narrowest customer hypothesis, about their life and not your idea. Ask about the last time the problem came up, what they did, what it cost and what they have tried. Rob Fitzpatrick's The Mom Test is still the best short guide to doing this without being misled, and our reference on the Mom Test method covers its rules.

How many is enough? For a fairly uniform group, Guest, Bunce and Johnson found that 92% of their codes had appeared by the twelfth interview (Field Methods, 2006). That makes 10 to 15 a sensible first batch for one segment. Our guide to how many interviews are enough covers when you need more.

If this is your first round, our founder playbook on your first 10 customer discovery calls has who to book, a booking message and a timed script. If you have no one to call yet, start with finding 20 people to interview when you have no audience.

After each interview, score it against your hypothesis table, not against how the call felt. A friendly call where someone said "that sounds useful" moves nothing. A call where someone described losing two days last month and showed you the spreadsheet they built moves the problem and current-fix rows.

Step 3: Test the solution

Only when the problem rows hold do you show anything. Keep it rough: a sketch, a landing page, a clickable mockup, a one-paragraph description. The goal is to learn whether your answer fits the problem they described, not to collect praise.

Show it, then ask what they think it does before you explain it. Ask what they would stop using if they had it. Ask what is missing. If people cannot tell what it does in a few seconds, that is a finding too.

With Koji you can run this step as an AI interview: attach the mockup or landing page to a question, and the interviewer asks about what people see, then follows up on their answers. For a first-impression test, the image can be shown once, full screen, for 3, 5, 10 or 15 seconds and then hidden, after which Koji asks what they remember. That lets you show the same concept to 30 people in a day instead of booking 30 calls.

Step 4: Ask for a commitment

Interviews tell you what people say. A commitment tells you what they will do. Before you decide to build, ask the people who were most excited for something that costs them a little: a paid pilot, a deposit, a letter of intent, an intro to their boss, an hour to try a prototype with their real data.

A founder and a potential customer talking across a table, the moment an interest turns into a pilot or an intro
A founder and a potential customer talking across a table, the moment an interest turns into a pilot or an intro

Photo by Amy Hirschi on Unsplash

Not everyone will say yes, and that is the point. If ten people loved the idea and none would give you an hour, you have learned something important. Our guide to customer discovery vs customer validation covers how this step leads into validation.

Step 5: Build, pivot or stop

Set a date for this before step 2 starts, two to four weeks out, and keep it. On that date, go through the hypothesis table row by row and mark each one held, broke or unclear.

Flowchart for deciding after customer discovery: three questions on the problem, workaround spend and commitments, ending in building the smallest version.
Make the call on the date you set, with the table you wrote before the first interview · Download or share

Build when the problem held (at least six in ten described it unprompted, with a recent example), people are already spending time or money on a workaround, and some of them committed to something. Build the smallest version, for the people who committed, and keep interviewing them.

Pivot when one hypothesis broke and the rest held. The kind of pivot follows from which row broke:

  • The customer row broke, but someone else had the problem worse: a segment pivot. Rewrite the customer hypothesis and run ten more interviews with that group.
  • The problem was real but smaller than you thought, and a neighbouring problem kept coming up: a problem pivot. Write new hypotheses for that problem.
  • The problem held, but nobody would swap their current fix for your solution: a solution pivot. Keep the problem, change what you offer, and test again.

Our founder playbook on pivot decision research goes deeper on choosing between these.

Stop when the problem itself did not show up: few people had it recently, nobody had spent anything on it, and no one committed. That is a good result. You learned it in weeks instead of after a year of building. Pick the next problem from what people did talk about, and start again at step 1.

The mistakes that keep founders in discovery forever

No kill criteria. Without a "proved wrong if" column, every interview can be read as encouraging. Write it first.

Mixing segments. Ten interviews across five kinds of customer gives you five sets of two. Pick one segment per round.

Pitching in problem interviews. The moment you describe your idea, people start being kind. Save the solution for step 3.

Counting compliments. "I would definitely use that" is not evidence. Past behaviour, money spent and commitments are.

No decision date. Discovery that never ends is a way of avoiding a hard call. Put the date in the calendar.

Where Koji fits in the process

Koji helps most in steps 2 and 3, when you need to hear from more people than your calendar allows, or in more than one segment at once. Describe what you want to learn and the research assistant drafts a study you can edit. Add screening questions so only people in your customer hypothesis get in, then share one link. Each participant has a voice or text conversation with an AI interviewer that asks your questions and follows up on their answers, in their own time. Koji then groups what people said into themes, with the quotes behind each one, which makes the row-by-row review in step 5 faster.

Koji report summary with a side panel of supporting participant quotes behind one finding
Every theme in a Koji report links back to the quotes behind it, so you can check a hypothesis against what people actually said

It does not replace the first handful of calls you run yourself, and it cannot ask for a commitment on your behalf. Use it to widen the evidence once you know what to ask.

A free account comes with 10 credits, and a text interview uses one, so you can test your first hypothesis table without paying. Start a free discovery study, or read Koji for founders to see how other founders set theirs up.

Run your first AI-moderated study in 10 minutes

100 free credits on signup. No credit card required.

GDPR compliantEU or US data residencyNo AI training on your data
Koji

Koji Team

Product

Share this article

Keep reading