How to validate an idea with AI before building too much

AI can help you design a validation process. It cannot validate demand by agreeing that your idea sounds promising. Validation means reducing a specific uncertainty with relevant evidence. Separate the existence of a problem, willingness to change behavior, ability to use a solution and the economics of delivering it.

Write the riskiest claim

Imagine a service that helps independent instructors fill last-minute class openings. The key claim may be: instructors lose enough revenue to spend time on a new distribution channel. A beautiful marketplace prototype does not answer that question. Ask AI to list the assumptions and rank them by consequence if false; then apply your own judgment to the ranking.

Use an assumption ledger with the claim, current evidence, uncertainty, cheapest informative test and result that would change your mind. Avoid a single composite “idea score,” which can conceal an unanswered existential question.

Investigate actual behavior

Talk with people who recently experienced the problem. Ask about a specific empty class, what they tried, what it cost and what happened. Do not lead with your solution. AI can help improve the interview guide and flag leading questions; it cannot provide substitute interview participants.

Research existing alternatives using primary sources and dated observations. Include spreadsheets, social posts, agencies and doing nothing. A model's competitor list is a starting hypothesis until you verify that the businesses and features actually exist.

Choose a test matched to the uncertainty

For problem importance, investigate recent events and workarounds. For demand, offer a clearly described pilot with a meaningful next step. For usability, test a prototype with relevant participants. For feasibility, attempt the difficult operational or technical step. These tests answer different questions and should not be collapsed into “users liked it.”

A landing page can measure response to a specific message and traffic source. State honestly whether the product is a pilot or waitlist. A click does not prove willingness to pay, and an email address does not prove retention. If you accept money, make delivery terms and refund handling explicit.

Decide in advance what the evidence can support

Choose a timebox, recruitment approach and decision rule before seeing results. Small early tests are directional. Report the denominator, source and context rather than presenting a tiny sample as population demand. If friends enroll, say so; that audience may behave differently from cold prospects.

Ask AI to challenge your interpretation: what else could explain the result, who is missing, and what would disconfirm the story? Keep the original records available to the reviewer.

Finish with a decision

Choose to continue, narrow the audience, revise the offer, run another test or stop. Document the next uncertainty. A failed assumption discovered cheaply is a useful result. The artifact you want is an evidence-backed next step, not a polished pitch claiming certainty.