# How to Validate a Business Idea with AI (Prompts + Framework) > Source: > Published: 2026-07-26 05:13:43+00:00 # How to Validate a Business Idea with AI (Prompts + Framework) **Short answer:** AI cannot validate your business idea — only paying customers can. What AI *can*do is design the test: turn a vague idea into falsifiable assumptions, write interview questions that don’t lead the witness, and audit your evidence for wishful thinking. The five-step sequence below uses AI at every stage **except** the one that matters most: talking to real humans. Ask a model “is this a good idea?” and you’ll get agreement. Ask it to design the experiment that could prove you wrong, and you get something useful. **TL;DR — Key Takeaways** **AI validates nothing.** It has no customers. Use it to design tests and interpret evidence — never as the source of the verdict.**43% of failed startups died of poor product-market fit**— the single largest cause in CB Insights’ study of 431 shutdowns ([CB Insights, 2024](#aiv-sources)).** Ask about the past, never the future.**“Would you use this?” produces false validation. “What did you do last time?” produces evidence.** Set kill criteria before you collect data.**Deciding the threshold afterwards lets you reinterpret a failed test as encouraging.** 10–20 conversations, not three.**Below ten you can’t tell a pattern from a few polite chats. **✔ Best for** Founders and operators with an idea they haven’t built yet, or a side project they’re deciding whether to commit to. Also useful for validating a new product line inside an existing business. **✕ Skip if** You already have paying customers and are optimising (that’s product work, not validation), or you’re looking for a business plan template or investor deck rather than an evidence-gathering method. **On this page** ## Can ChatGPT tell me if my business idea is good? **No — and the way it fails is dangerous, because it fails agreeably.** A language model has never met your customer, has no access to your market, and is optimised to be helpful. Ask whether your idea is good and it will find reasons to say yes, complete with a plausible-sounding market size and a list of competitive advantages it inferred from nothing. That’s not validation. That’s a very articulate friend telling you what you want to hear. The most expensive mistake in this entire category is mistaking fluent encouragement for evidence. **The core reframe:** validation is the process of trying to *disprove*your idea cheaply, before the market disproves it expensively. Every prompt in this guide is built to attack the idea, not support it. If a prompt makes you feel good, it probably isn’t working. The stakes are well documented. CB Insights analysed 431 VC-backed startups that shut down since 2023 and found **43% failed due to poor product-market fit** — building something not enough people cared to pay for. Those companies had raised a combined $17.5 billion, with a median of $11 million each. They didn’t lack money. They lacked evidence that anyone wanted what they were building. ## The 5-step validation sequence Each step feeds the next. Run them in order — skipping step one is why most validation efforts produce ambiguous results. **Step 1** **Problem** Turn the idea into falsifiable assumptions **Step 2** **Demand** Design the evidence hunt **Step 3** **MVP** Smallest disproving test **Step 4** **Runway** Can you survive the test **Step 5** **Outreach** Reach enough people Expect two to four weeks end to end. The output isn’t certainty — it’s a decision you can defend: proceed, change the idea, or stop. ## Step 1: How do I turn a vague idea into something testable? **Convert the idea into a list of assumptions that could be proven false.** “An app for freelancers to manage invoices” isn’t testable. “Freelancers earning $50–150k lose more than two hours a week chasing late payments, and would pay $20/month to stop” is — every clause can be checked against reality. This step is where most validation fails silently. If your assumptions stay vague, no evidence will ever feel conclusive, and you’ll drift toward building. ``` MY IDEA: [DESCRIBE IT IN 2-4 SENTENCES, PLAINLY] WHO I THINK IT'S FOR: [SPECIFIC CUSTOMER — role, company size, or life situation. "Small businesses" is not specific enough.] WHAT I THINK IT REPLACES: [WHAT THEY DO TODAY INSTEAD] Do not evaluate whether this is a good idea. Instead: 1. Rewrite my idea as a single testable problem statement in this form: "[SPECIFIC PEOPLE] experience [SPECIFIC PAIN] when [SPECIFIC SITUATION], and currently cope by [CURRENT WORKAROUND]." If I haven't given you enough to fill a blank, ask me rather than inventing it. 2. List every assumption that must be TRUE for this business to work. Split into: - PROBLEM assumptions (the pain exists, and is felt often) - CUSTOMER assumptions (these specific people have it) - SOLUTION assumptions (my approach fixes it) - MONEY assumptions (they'll pay, and enough of them exist) - CHANNEL assumptions (I can reach them affordably) 3. For each assumption, mark: - Is it already TESTED or UNTESTED? - If FALSE, is this FATAL or SURVIVABLE? 4. Rank the FATAL + UNTESTED assumptions by how cheap they are to test. That ranking is my validation to-do list. 5. Name the one assumption I appear most emotionally attached to — the one I'd resist giving up. Say it directly. ``` **Why it works:** point 5 is the one competitors’ prompts skip. Founders don’t fail from lack of information; they fail by protecting a favourite assumption from testing. Naming it early makes it harder to defend later. ## Step 2: How do I test demand without building anything? **Talk to people about their past, not your idea.** The most reliable demand signal isn’t enthusiasm — it’s discovering what someone already does, pays for, or hacks together to cope with the problem. If they’ve built a spreadsheet workaround, the pain is real. If they’ve never thought about it before you asked, it isn’t. This is the core of Rob Fitzpatrick’s *The Mom Test*: ask about specific past events instead of hypothetical futures, because hypotheticals produce false validation. ``` CONTEXT: [PASTE YOUR PROBLEM STATEMENT FROM PROMPT 1] RISKIEST ASSUMPTION I'M TESTING: [FROM PROMPT 1, ITEM 4] WHO I'M INTERVIEWING: [ROLE / SITUATION] Write me a 20-minute customer discovery interview script following Mom Test rules. HARD RULES — the script must not: - Mention my idea or my solution at any point - Ask "would you", "could you see yourself", or any hypothetical - Ask anyone to rate, score, or predict their own behaviour - Contain a leading question that signals the answer I want The script MUST include: - An opening that doesn't reveal I'm building anything - Questions about the LAST TIME they faced this problem - What they actually DID about it, step by step - What it cost them — time, money, or something else - What they're using to cope today, and what they pay for it - Who else was involved and who holds the budget - One question designed to make it EASY for them to tell me the problem isn't a big deal For each question, add a one-line note on what a strong answer versus a weak answer sounds like. Finish with 3 things I should listen for that would suggest I'm wrong, and 3 that would suggest I'm onto something. ``` **The single strongest question:** “What are you using to deal with this today?” What people already build, buy, or hack together to cope is the clearest evidence of real demand — far stronger than anything they say about your idea. ### Get the Idea Validation Workbook All five prompts pre-loaded, plus an interview logging sheet, an evidence scoring rubric, and a kill-criteria worksheet you fill in before you start. Free. [Get the workbook →](#aiv-lead-magnet) ## Step 3: What’s the smallest test that could prove me wrong? **Design the minimum test that could disprove the idea — not the minimum product you could build.** Those are different things, and confusing them is how founders spend four months building an MVP that was never a test of anything. A good validation test has a result you’d act on either way, and it finishes in under two weeks. ``` ASSUMPTION I'M TESTING: [ONE FATAL, UNTESTED ASSUMPTION] MY CONSTRAINTS: [TIME AVAILABLE, BUDGET, SKILLS I HAVE] Design 3 tests that could DISPROVE this assumption. For each: - What exactly I do (concrete steps, not "run an experiment") - How long it takes and what it costs - The specific result that would DISPROVE the assumption - The specific result that would SUPPORT it - What it does NOT tell me (the limits of the test) Rank them by: fastest to a clear answer, not cheapest. Rules: - No test may involve building a working product. - Every test must be finishable in under 2 weeks. - If a test's result could be interpreted either way, discard it and say why. Then: name the cheapest test that a founder would AVOID because they're afraid of the answer. Recommend that one and explain the avoidance. ``` **Why it works:** the last instruction targets the real failure mode. Founders unconsciously choose tests that can’t fail — a landing page with no traffic source, a survey of friends. Asking the model to surface the test you’re avoiding cuts through that. ## Step 4: Can I afford to run this test? **Validation has a cost, and running out of room mid-test is its own failure.** The point of this step isn’t a financial model — it’s making sure you’ll still be in a position to *act* on the answer when it arrives. Plenty of founders validate successfully and then discover they have no runway left to build. ``` MY SITUATION: Cash available for this: [AMOUNT] Hours per week I can commit: [HOURS] Current income situation: [EMPLOYED / SAVINGS / OTHER] Months before I need this to earn: [NUMBER] THE TEST I'M PLANNING: [FROM PROMPT 3] 1. What does this test actually cost me in cash, hours, and opportunity cost? Include the costs founders forget. 2. If the test says GO, do I have enough left to act on it? Show the arithmetic and state your assumptions. 3. If the test says STOP, what have I lost, and is that an acceptable price for the information? 4. What is the cheapest version of this test that still gives me a clear answer? 5. Name the single financial assumption in my plan most likely to be wrong. Show your arithmetic so I can check it. Do not round in a way that hides a problem. ``` **Verify the numbers yourself.** Language models produce arithmetic that looks right and sometimes isn’t. Use this prompt to surface the variables and the questions — then check the maths in a spreadsheet, and involve an accountant before any decision with tax or legal consequences. ## Step 5: How do I find 20 people to talk to? **A reliable read takes 10–20 conversations, not three.** Below about ten you can’t distinguish a genuine pattern from a few polite chats. Most founders stop at four, hear encouraging things, and start building — which is how false validation happens. ``` WHO I NEED TO TALK TO: [SPECIFIC ROLE / SITUATION] MY ACCESS: [EXISTING NETWORK, COMMUNITIES I'M IN, COLD OUTREACH ONLY, ETC.] 1. List 10 specific places these people already gather — named communities, forums, events, subreddits, associations, Slack groups. Not "LinkedIn" or "social media." Mark which ones I can enter without being a spammer. 2. Write 3 outreach messages under 60 words each: - One for a warm intro - One for cold outreach - One for posting in a community Every message must: - Ask for 20 minutes about THEIR experience - Never mention that I'm building something - Give a concrete reason I'm asking THEM specifically - Make it easy to say no 3. Tell me what response rate to expect for each channel and how many people I need to contact to get 20 conversations. 4. Name the recruiting shortcut that would bias my results, and why I'll be tempted to take it. ``` Point 4 usually surfaces the same answer: interviewing friends, or people who already like you. Friendly samples are the most common source of false positives in early validation. ## What does a good validation prompt look like vs. a bad one? Nearly every “validate your idea with ChatGPT” prompt circulating online is on the left side of this table. The pattern is the same each time: asking the model for a verdict instead of a method. | Move | ❌ What most people do | ✅ What actually works | |---|---|---| The ask | “Is this a good business idea?” | “List the assumptions that must be true, and mark which are fatal if false.” | Market sizing | “What’s the market size for this?” — gets a confident, unsourced number | “What evidence would tell me this market is big enough, and where would I find it?” | Customer input | “Write a survey asking if people would use this” | “Write interview questions about what they did the last time they had this problem” | Competitors | “List my competitors and my advantages” | “What would a customer have to believe to choose me over what they use today?” | The verdict | Model gives a go/no-go — founder feels validated | Model designs the test; customers give the verdict | Emotional posture | Seeking permission to proceed | Actively trying to kill the idea cheaply | The through-line: **use AI for method, never for verdict.** A model can design a rigorous test in seconds. It cannot tell you whether anyone will pay, and it will not admit that unless you structure the prompt so it can’t. ## Can you show a real prompt with its actual output? Here’s Prompt 1 run on a real-shaped idea — a founder considering a scheduling tool for independent physiotherapists. ``` MY IDEA: A scheduling and reminder tool built specifically for solo physiotherapists. They currently use generic calendar apps that don't handle recurring treatment blocks or insurance paperwork, and they lose money on no-shows. WHO I THINK IT'S FOR: Independent physios running their own practice, 1-3 person clinics, in the UK. WHAT I THINK IT REPLACES: Google Calendar plus manual WhatsApp reminders plus a paper diary. Do not evaluate whether this is a good idea. Extract the assumptions, mark fatal vs. survivable, and name the one I'm most attached to. ``` **Testable problem statement:** “Solo physiotherapists running 1–3 person UK clinics lose revenue to no-shows and spend unpaid admin time on scheduling, and currently cope with a mix of Google Calendar, manual WhatsApp reminders, and a paper diary.” **Assumptions — fatal & untested (your to-do list, cheapest first):** **MONEY:** No-shows cost enough that a physio would pay monthly to reduce them.*Untested — fatal.*Cheapest test: ask 10 physios what last month’s no-shows cost them. If they don’t know the number, the pain isn’t salient enough to sell against.**PROBLEM:** Manual reminders are experienced as a burden, not as a two-minute routine.*Untested — fatal.*Many solo practitioners describe WhatsApp reminders as a client-relationship touchpoint they value.**CHANNEL:** You can reach independent physios affordably.*Untested — fatal.*They are geographically scattered, not concentrated in online communities, and rarely respond to cold outreach.**SOLUTION:** Scheduling is the binding constraint, not insurance paperwork.*Untested — survivable.*If paperwork is the real pain, this becomes a different product. **Already partially tested:** that physios use generic tools — observable, low risk. **The assumption you appear most attached to:** that no-shows are the primary pain. You led with it and framed the whole product around it, but you haven’t cited a single physio saying so. It’s also the most convenient assumption for you, because it’s the one your solution addresses. Test it first and test it hardest — specifically, ask what they’ve already tried to reduce no-shows. If the answer is “nothing,” they don’t consider it a problem worth solving. Note what the output didn’t do: it never said whether this was a good idea. It converted a plausible-sounding pitch into four falsifiable claims and identified which one the founder was protecting. That’s a two-week work plan — and it’s far more useful than a verdict. ## Level-up: the evidence audit prompt This is the part competitors’ guides don’t have. Everyone tells you to talk to customers. Almost nobody helps you interpret what you heard — which is where validation actually goes wrong. Founders come back from twenty interviews, remember the three enthusiastic ones, and call it validated. Run this **after** your interviews, with your raw notes pasted in. ``` MY ORIGINAL ASSUMPTION: [WHAT I SET OUT TO TEST] MY KILL CRITERIA, SET IN ADVANCE: [WHAT RESULT WOULD MAKE ME STOP — if you didn't set one, say so] RAW INTERVIEW NOTES: [PASTE EVERYTHING — messy is fine. Include who each person was and roughly what they said.] Audit this evidence. Be harsh. Respond in these six sections: A. SAID vs DID — separate every data point into what people SAID (opinions, praise, intentions) and what they have actually DONE (paid for something, built a workaround, changed behaviour, gave up time or money). Only the second column is evidence. B. LEADING QUESTIONS — quote back any question I asked that signalled the answer I wanted, or where I pitched instead of listened. Be specific about which responses those questions contaminated. C. SAMPLE BIAS — who did I talk to, and who is missing? Flag if these people are unusually friendly to me, unusually accessible, or unrepresentative of the buyer. D. THE PATTERN I'M IGNORING — what shows up repeatedly in these notes that doesn't support my idea? Quote it. E. VERDICT AGAINST MY OWN CRITERIA — did I hit the kill criteria I set? Answer yes or no plainly. If I didn't set criteria in advance, say that this audit is much weaker as a result and explain why. F. WHAT I'D NEED TO HEAR to genuinely validate this, and how far I am from hearing it. Rules: - Do not soften the finding. Do not open with encouragement. - If the honest read is that I have interest but no evidence of demand, say exactly that. - Distinguish "people confirmed the problem" from "people will pay to solve it." These are not the same finding. ``` **Why this is the unlock:** Section A alone reframes most founders’ results. Enthusiasm collapses into a short “DID” column and a very long “SAID” column — and seeing that gap laid out is more persuasive than any advice about confirmation bias. Section E is the accountability mechanism: it holds you to a standard you set before you knew the answer. **Setup tip:** paste your notes verbatim, including the parts that felt discouraging. Cleaning them up first defeats the purpose — the audit is only as honest as its input. ## How do I decide: go, pivot, or kill? **Decide the threshold before you gather data, then hold yourself to it.** Setting kill criteria afterwards lets you reinterpret weak results as encouraging — the most common way founders talk themselves past a failed test. **▲ Go** Multiple people have already spent money or meaningful time coping with this problem. Several offered something scarce — a referral, a deposit, a pilot slot. The pattern held across people who don’t know you. **◆ Pivot** The problem is real but you’re solving the wrong part of it, or for the wrong person. The repeated complaint wasn’t the one you expected. Keep the customer, change the product — or keep the product, change the customer. **▼ Kill** People confirmed the problem exists but have never tried to solve it. Nobody has paid for a workaround. Enthusiasm was high and commitment was zero. “That sounds useful” with no follow-through is a no. The distinction that matters most: **“people confirmed the problem” and “people will pay to solve it” are different findings.** Nearly every idea passes the first test. Most fail the second — and conflating them is what produces a confident launch into an empty market. ## Which model for which step? The prompts are model-agnostic. Practical notes as of **July 2026**: | Step | Best fit | Why | |---|---|---| Assumption extraction (1) | Claude or ChatGPT | Both hold an analytical stance without drifting into encouragement, which is the failure mode here. | Market & competitor research | A model with live web search | Training data alone invents market sizes and misses competitors launched this year. Never cite a figure you can’t click through to. | Evidence audit (level-up) | Claude | Large context windows hold twenty interviews of raw notes without you chunking them, which matters for spotting cross-interview patterns. | Interview script (2) | Any | A well-constrained prompt does the work here; model choice barely matters. | Runway maths (4) | Any — then verify | Use it for structure and variables. Check every number in a spreadsheet regardless of model. | We re-check these notes whenever a major model ships. If you’re reading this more than two weeks after the date above, verify your model versions still match. ## Frequently asked questions ### Can ChatGPT validate my business idea? No. AI cannot validate a business idea because validation requires evidence from real potential customers, and a language model has none. What AI does well is design the test: sharpening your idea into falsifiable assumptions, writing unbiased interview questions, and auditing the evidence you collect for wishful thinking. Asking a model whether your idea is good produces agreement, not validation. ### How many customer interviews do I need to validate an idea? A reliable read generally takes 10–20 interviews, not three. Fewer than about ten and you can’t distinguish a genuine pattern from a few polite conversations. What matters more than the count is whether you asked about specific past behaviour rather than hypothetical future interest. ### What is the biggest reason startups fail? Poor product-market fit. CB Insights analysed 431 VC-backed startups that shut down since 2023 and found 43% failed because of poor product-market fit. Running out of cash appears in around 70% of failures, but CB Insights describes it as the final symptom rather than the root cause — teams ran out of money while searching for demand that was never there. ### What questions should I ask to validate a business idea? Ask about specific past events, not future opinions. Good questions: when did you last deal with this problem, what did you do about it, what did that cost you in time or money, and what are you using to cope with it today. Avoid asking whether someone *would* use your product — hypothetical answers produce false validation. ### How do I know if my idea is validated or if people are just being polite? Look for commitment rather than compliments. Praise, enthusiasm and “that sounds great” are not evidence. Evidence is someone giving up something scarce: money, time, a referral, access to their calendar, or a written commitment to try it. If nobody has given up anything, you have interest, not validation. ### How long should validating a business idea take? Two to four weeks for a first read in most cases. The goal isn’t certainty but a decision: proceed, change the idea, or stop. Validation that stretches past a month usually means the test wasn’t designed to be falsifiable, so no result ever feels conclusive enough to act on. ### Should I set kill criteria before I start validating? Yes, and in writing before you collect any data. Decide in advance what result would make you stop — for example, fewer than three of twenty people having actively tried to solve the problem in the last six months. Setting the threshold afterwards allows you to reinterpret weak results as encouraging, which is the most common way founders talk themselves past a failed test. ### Can AI do the customer interviews for me? It shouldn’t. The value of a discovery interview is in the unexpected tangent, the hesitation, and the follow-up question you didn’t plan. Use AI to write the script, to debrief you afterwards, and to audit your notes for leading questions — but conduct the conversations yourself. ### Download: The Idea Validation Workbook All five prompts plus the evidence audit, an interview logging sheet, a said-vs-did scoring rubric, and a kill-criteria worksheet you complete *before* you start. Everything in one file. [Send me the workbook →](#) Enter your email and we’ll send the workbook plus a short monthly prompt update. Unsubscribe anytime. **Written by the Narracomm team** Narracomm is a communications and content strategy team that helps business owners, operators, and founders use AI to produce clear, credible, high-performing work. We build and test these prompt systems inside real client engagements — including early-stage validation work — and revise them as models change. [Add specific credentials, ventures validated or advised, years of experience, and a named reviewer here to strengthen E-E-A-T.] ## Sources & further reading [CB Insights — Why Startups Fail: Top Reasons (431 shutdowns analysed)](https://www.cbinsights.com/research/report/startup-failure-reasons-top/)[User Intuition — Research on “no market need” as a failure cause](https://www.userintuition.ai/reference-guides/why-startups-fail-no-market-need-research/)[Perspective AI — 60 Mom Test-approved customer discovery questions (2026)](https://getperspective.ai/blog/60-customer-discovery-questions-for-2026-mom-test-approved)[The Mom Test for Customer Interviews (2026)](https://www.koji.so/blog/mom-test-customer-interviews-2026)[Harvard Innovation Labs — Validate Customer Demand](https://innovationlabs.harvard.edu/how-to/validate-demand)- Rob Fitzpatrick, *The Mom Test*(2013) — the origin of the past-behaviour interview method. Last reviewed and updated: **July 25, 2026** · Prompts tested against current model releases. Next review due within 14 days.