Customer Discovery Questions
The questions that surface the real problem, the constraints, and the definition of success — before you write a line of code.
Discovery is how you learn which problem is worth solving before you build anything. These questions are designed to surface specific past behavior — what people actually did, what it cost, who was involved — because opinions about the future are unreliable and polite people will praise almost any idea.
Run each call as an investigation, not a pitch: watch the real workflow, put a number on the pain, and leave with a metric, a baseline, the constraints, and the names of the people who can say yes or no.
The problem (not the feature)
- Walk me through it
- “Show me how you do this today, step by step.” Watch, don’t ask them to summarize.
- Cost of the status quo
- “What does this cost you now — time, money, errors, missed revenue?”
- Last time it hurt
- “Tell me about the last time this went wrong.” Specifics beat generalities.
- Who feels the pain
- “Who does this today, and how many of them?”
- Why now
- “Why is this worth solving this quarter and not last year?”
- What you tried
- “What have you tried already, and why didn’t it stick?” — shows real effort and failed fixes
- How often
- “Daily, weekly, at quarter-end?” Frequency times cost is the size of the prize
- The workaround
- “Can I see the spreadsheet or script you use today?” A workaround proves the pain is real
- If it vanished
- “If this problem disappeared tomorrow, what would you do with that time?”
Workflow & data
- Inputs and outputs
- “What comes in, from where, and what has to come out — in what format?”
- Volume
- “How many per day, and what does a busy day look like?” — this sizes the system
- Edge cases
- “Which weird cases eat half your time?” The long tail decides feasibility
- Real examples
- “Can you share 20 real examples, good and bad?” — the seed of your eval set
- Hand-offs
- “Who touches it before you, and after you?” Every hand-off is an integration
- Approvals
- “When does a person have to sign off, and who is that?”
Success & scope
- Definition of done
- “If this works perfectly, what’s different on Monday morning?”
- The one metric
- “What number would move that tells us it worked?”
- Good enough
- “What accuracy / latency / coverage is good enough to be useful?”
- The demo bar
- “What would you need to see to believe this is real?”
- Baseline
- “What is that number today?” Without a baseline nobody can prove an improvement
- Cost of a wrong answer
- “When it gets one wrong, is that annoying, rework, or a compliance incident?”
- Human review
- “Where must a person check the output before anything goes out?”
- Out of scope
- “What should the first version explicitly not do?”
Constraints (find the landmines)
- Where can it run
- “Cloud, your VPC, on-prem, air-gapped? Who approves that?”
- Data rules
- “Can data leave your network? Any PII / regulated data (HIPAA, PCI)?”
- Systems to touch
- “What systems hold the data — and who owns access to them?”
- Security review
- “What does it take to get a new tool approved here?”
- Deadline drivers
- “Is there a date this must work by — a launch, an audit, a renewal?”
- Existing tools
- “What do you already pay for that does part of this?”
- AI vendor rules
- “Are any AI vendors or models approved here, or banned?”
People (map the room)
- Champion
- “Who wants this to succeed and will fight for it internally?”
- Blocker
- “Whose approval could stop this — security, legal, a skeptical lead?”
- Decision maker
- “Who signs off / holds the budget?”
- The user
- “Who actually uses it daily — and do they want it?”
- How buying works
- “Walk me through the last time you brought in a tool like this.”
- Budget
- “Where would the money come from?” — tells you whether it is funded or a wish
- Who else to ask
- “Who else should I talk to about this?” — end every interview with it
Running the conversation
- Past, not future
- ask what they did last time, not what they would do — hypotheticals invite polite fiction
- Their life, not your idea
- pitching early turns discovery into a sales call; describe your idea last, if at all
- Compliments are noise
- “Sounds great!” is not data — steer back to a specific recent example
- Follow the emotion
- a sigh or an “ugh” is a lead: ask why that part is so painful
- Talk less
- let them speak most of the time; a pause after a question is fine
- No leading questions
- not “Would you use a tool that…?” but “How do you handle that today?”
- Ask for a commitment
- end with a next step that costs them something — time, data access, an introduction
After the call
- Write it up within a day
- problem, metric, constraints, people, open questions — while it is fresh
- Send a recap
- play back what you heard in their words; corrections are the most valuable reply
- Facts vs. guesses
- mark what they said, what you saw, and what you only inferred
- Red flags
- no metric, no owner, no data access, or “just add AI” — resolve these before building
Worked example: from “AI on the claims inbox” to a scoped pilot
An insurer’s operations lead opens with: “We want AI on the claims inbox.” That names a solution, so the call works through the question families in order.
- Walk me through it. On a screen-share, a handler opens an email, looks the claim up in the core system, then either replies with a status template or forwards it to an adjuster.
- How often, and what it costs. About 600 emails a day across 5 handlers, roughly 90 seconds each: 600 × 1.5 min = 900 minutes, or 15 handler-hours a day.
- The workaround. A shared spreadsheet of 14 reply templates. A week of the team’s own labels shows 62% of emails are “where is my claim?”
- Baseline and good enough. The median first reply takes 6 hours today. They want status questions answered within 15 minutes.
- Cost of a wrong answer. A wrong status means rework. A wrong statement about coverage means a regulatory complaint, so version one never discusses coverage.
- Constraints and people. The emails contain health data and must stay in the insurer’s Microsoft 365 tenant. Security reviews new vendors once a month. The champion is the ops lead. The likely blocker is the compliance officer, who has not been told yet.
- Commitment. She agrees to export 200 anonymized emails by Friday and book 30 minutes with compliance.
problem: ~370 status emails a day (62% of 600) take about 9 handler-hours
metric: median first reply on status emails (baseline 6 h, target 15 min)
scope v1: status questions only; anything mentioning coverage goes to a person
risk: compliance not yet involved → meeting booked before any build
Seven questions turned “AI on the inbox” into something a two-week pilot can prove or disprove. They also found the one person who could have stopped it in month two.
Recipes
A 45-minute discovery call
0-5 min intros, purpose, permission to take notes: "I'm here to learn, not to pitch"
5-20 min "Walk me through how you did this last time" (screen-share the real workflow)
20-30 min pain and cost: how often, time per case, errors, what breaks at quarter-end
30-38 min success: the one metric, today's baseline, what "good enough" means
38-43 min constraints and people: data rules, where it can run, who approves
43-45 min a next step with a date, and "Who else should I talk to?"Spend most of the time on the walkthrough: watching one real case beats an hour of description. If you run out of time, cut the agenda, not the next step.
Turn a feature request into a problem statement
request: "We need an AI chatbot for our support team."
ask: "What would it handle on day one?" -> password-reset tickets
ask: "How many, and what does each one cost?" -> 1,200 a month, about 6 minutes each
ask: "What does done look like?" -> half of them resolved with no agent
problem: 1,200 password-reset tickets a month take about 120 agent-hours
metric: share resolved without an agent (baseline 0%, target 50%)
scope v1: password resets only, in English; hand off to a human if the identity check failsThe request named a solution; the questions found the job, its size and a measurable target. Scope v1 is narrow on purpose — it can ship and be proven in weeks.
Spot and fix a leading question
leading: "Would you use a tool that summarizes your contracts automatically?"
-> "Sure, sounds useful." (polite, hypothetical, worthless)
better: "Tell me about the last contract you had to review."
-> "It took two days, and I missed an auto-renewal clause. We paid for a year."
follow-up: "What did that cost? What do you do differently now?"The better question produces a story with a cost attached, which is evidence. The first only tells you the person is polite.
A discovery recap email
Subject: Recap: invoice matching, and next steps
What we heard
- 3 AP clerks match about 2,000 invoices a month to POs by hand (~4 min each)
- mismatches over $500 need controller approval: about 8% of invoices
- success = 70% matched automatically, with zero false matches over $500
Constraints: data stays in your Azure tenant; security review takes ~3 weeks
Open questions: who owns the ERP export? can we get 200 sample invoices?
Next step: data walkthrough with Priya, Thursday 2:00 pmWriting it down exposes gaps before they become rework, and the customer’s corrections are often the most useful thing you learn all week.
When to use it, and when not
Use these questions whenever someone gives you a solution before a problem: a new customer engagement, an internal tool request, a feature ask passed on by sales, or a pilot that has run for weeks with no agreed metric. They matter most when the build is expensive and the customer’s environment is unfamiliar.
- Watching beats asking. If you can sit beside a user for a shift, do that first, then use the questions to fill the gaps.
- Not for a scope that is already decided. Once the scope is signed and funded, switch to requirements and acceptance criteria. Reopening “why” in every meeting stalls delivery.
- Not for usability tests. To learn whether people can use a prototype, give them a task and watch in silence. Discovery asks about their life without your product.
- Data beats memory for volume. When tickets or logs exist, count from the export instead of trusting an estimate.
Trade-offs and limits
Discovery buys certainty with calendar time and the customer’s patience. Each way of gathering evidence covers a different blind spot:
| Method | Gives you | Costs you |
|---|---|---|
| One long call | the whole story from one person | one point of view, often missing the edge cases |
| Short calls by role | users, budget holder and security each heard | weeks of scheduling, and conflicting stories to reconcile |
| Watching the workflow | the real steps, including the ones nobody mentions | access and permission, which are harder to arrange |
| Data samples | real volumes and the seed of an eval set | data-access approval, sometimes weeks of it |
The usual failure is not too little discovery but discovery with no output: notes nobody wrote up, and no metric. Set a limit. If a week of calls has not produced a metric and a baseline, the problem may not be ready to solve.
Common mistakes
- Pitching in the first ten minutes. Once you describe your idea, every answer bends toward it, and you hear politeness instead of evidence.
- Accepting an average. “About ten minutes each” hides the long tail. Ask about the last five cases and the worst one, because the long tail decides feasibility.
- Talking only to the champion. The person who invited you wants the project to work. The daily user and the security reviewer may not, and they are the ones who change the scope.
- No baseline. “Make it faster” with no number for today means nobody can prove the pilot worked, and the renewal becomes an argument.
- Leaving without a commitment. “Let’s keep in touch” is a polite no. A data export, an introduction or a date on the calendar is a real yes.
- Writing conclusions instead of quotes. Keep what they said and what you saw apart from what you inferred. Otherwise your guess has become a “requirement” by next week.
Check yourself
A prospect says, “Our analysts would definitely use a tool that drafts these reports.” What do you ask next, and why?
Something like “Tell me about the last report you wrote. How long did it take, and what was hardest?” Their statement is a prediction, which the sheet treats as polite fiction (Past, not future; Compliments are noise). The last real report gives you a time, a frequency and a pain you can measure.
You hear “300 cases a week, about 20 minutes each.” How big is the prize, and what must you still ask before you can claim an improvement?
300 × 20 = 6,000 minutes, or 100 hours a week (How often: frequency times cost). You still need the Baseline for the quality side, such as today’s error rate, because hours saved only count if quality holds. Asking for the last five cases also tests whether “20 minutes” hides a long tail.
The call went well and the champion is excited. You close with “I’ll send something over.” What did you miss?
Ask for a commitment and Who else to ask. End with a next step that costs them something (data access, an introduction to the security reviewer, a date) plus the names of other people to talk to. Excitement with no commitment is not a signal.
Your recap said 2,000 invoices a month. The customer replies: “It’s more like 800.” Is that bad news?
It is the point of Send a recap: corrections are the most valuable reply. It also changes the sizing. At 800 a month, the same saving per invoice is 40% of what you estimated, so check that the pilot is still worth running before you build.
Questions about Customer Discovery Questions Cheat Sheet
What questions should I ask in a customer discovery interview?
Ask about specific recent events rather than opinions: walk me through the last time you did this, what did it cost, what have you tried, and why did that not work. Then quantify: how often it happens, how long it takes, how many people are involved. Ask what success would look like and how it is measured today, so you get a metric and a baseline. Uncover constraints such as data rules, security review and where software can run, and map the people: who uses it, who pays, who can block it. Close by asking who else you should talk to.
What is the Mom Test?
The Mom Test is a set of interviewing rules from Rob Fitzpatrick’s book of the same name, named for the idea that even your mother should not be able to lie to you if you ask the right questions. Talk about their life instead of your idea, ask about specifics in the past instead of generalities or the future, and talk less while listening more. Compliments, hypothetical promises and feature requests are treated as weak data, while past behavior, money already spent and commitments such as time, introductions or a pilot are strong signals.
How many customer discovery interviews do I need?
Enough to stop hearing new things within a segment. Patterns usually start to repeat after five to ten conversations with similar people, such as the same role in similar companies, and if every call sounds different your segment is too broad. For an enterprise or forward-deployed project with one customer, cover every role that touches the workflow: the daily users, their manager, the budget holder, IT or security, and whoever owns the data. Keep interviewing after you start building, because the first prototype will surface questions nobody could answer in the abstract.
How do I avoid leading questions in customer interviews?
Avoid questions that contain your solution or invite agreement, such as would you use, do you think it would be useful, or would you pay for. They produce polite yes answers that predict nothing. Ask about what already happened instead: how do you handle this today, tell me about the last time, what did you do next, what did that cost. Keep questions open, avoid suggesting answers, and stay quiet long enough for people to fill the silence. If you need a reaction to an idea, show it at the end and watch for a concrete commitment rather than enthusiasm.