AI Idea Validation Checklist

25 checks to run before building your AI product or service. If you can't tick a box, that's your next task — not a reason to skip it.

From the Awesome AI Builder Series · print or save as PDF (Ctrl/Cmd+P) · CC BY 4.0

1. Problem

A good problem is frequent, costly, and currently handled badly.

• I can state the problem in one sentence"[Role] struggles to [task] because [root cause], which leads to [consequence]."
• The problem is frequentIt happens weekly or daily — not once a year.
• The problem is costlyI can estimate what it costs the user in time, money, or risk.
• I know how people handle it todaySpreadsheets, interns, agencies, ignoring it — the current workaround is my real competitor.
• I found 3+ independent public signalsForum threads, job posts, reviews, support tickets that mention this pain. Assumptions without evidence are just opinions.

2. Customer

Pain without a buyer is a hobby, not a business.

• I can name my primary segment preciselyRole, industry, company size, geography — not "businesses" or "everyone".
• I know who is NOT my customerExplicit exclusions keep the scope honest.
• I know who owns the budgetThe person who feels the pain is often not the person who signs.
• I've talked to 5+ potential customersReal conversations, not surveys of friends. Ask about the problem, don't pitch the solution.
• At least 3 of them described the pain unpromptedIf you have to explain why it's a problem, it isn't one.

3. Market

Small validated beats big imagined.

• People already pay to solve thisExisting spend (tools, services, headcount) is the best proof of willingness to pay.
• I can name 3 existing alternativesZero competition usually means zero demand.
• I know why nowWhat changed (tech, regulation, cost) that makes this solvable today?
• I can reach this segment repeatablyA named channel — community, LinkedIn, partnerships — not "word of mouth".
• The niche is big enough for my goalRough math: reachable buyers × plausible price vs. what I want to earn.

4. AI Fit

AI must be the best way to solve it — not just a way.

• AI is genuinely neededIf a form, a template, or a rule engine solves it, build that instead — it's cheaper and more reliable.
• The task tolerates imperfectionAI output is probabilistic. Drafts and suggestions: fine. Irreversible decisions: dangerous.
• A human review step exists where stakes are highLegal, medical, financial outputs need a human in the loop — plan for it in the workflow and the pricing.
• I have (or can get) the data the AI needsAccess to the user's documents, context, or history — with their permission.
• I've checked compliance basicsGDPR, sector rules, data residency. Regulated domains need compliance notes before code.

5. Commitment Test

Validation ends with a commitment, not a compliment.

• I can deliver value manually firstA concierge version (you + AI tools behind the scenes) tests demand before you build product.
• I have a testable offerOne sentence: what they get, in what timeframe, for what price.
• Someone has pre-committedA paid pilot, a signed LOI, a deposit — anything stronger than "sounds cool, keep me posted".
• I know my kill criteriaDecide in advance what result means "stop" — e.g. fewer than 2 committed pilots after 20 conversations.
• I have a 7-day next stepValidation momentum dies in weeks, not months. What will you test this week?