How to Validate an AI Business Idea Before Building Anything

Most AI ideas fail for ordinary reasons — nobody wanted it enough to pay — and a few fail for AI-specific reasons: the data was not available or the accuracy was not good enough. Validation has to test both. Here is the framework I use.

Stage 1 — Validate the problem (one to two weeks)

Talk to ten people who fit your target customer. Do not pitch. Ask how they do the task today, how long it takes, what goes wrong and what they have tried. Listen for:

• Frequency: daily or weekly problems beat quarterly ones.
• Pain: words like "hate", "waste", "always late".
• Existing spend: time, staff or tools already dedicated to it.
• Workarounds: spreadsheets and chat groups are a strong signal.

If fewer than half describe the problem as painful without prompting, change the customer or the problem.

Stage 2 — The concierge test (two to four weeks)

Do the job manually for three to five of them using a model and your own effort. No product. Deliver the output the way they would receive it from a product — an email, a spreadsheet, a message.

You learn three things that no survey can tell you: what "good" output means to them, what inputs actually exist (and how messy they are), and whether they ask for it again. Ask for a token payment. Willingness to pay a small amount now predicts far more than a promise to pay later.

Stage 3 — AI feasibility (parallel with Stage 2)

Collect 30–50 real examples from the concierge work. Run the simplest technical approach — a well-designed prompt with structured output, plus retrieval if needed — and measure:

• Accuracy against what the customer accepted.
• Cost per item at realistic context sizes.
• Latency and whether it fits the workflow.
• Failure modes and whether a review step catches them.

If accuracy is far from the bar, ask whether a human-in-the-loop product is still valuable. Often it is: eighty percent automation with review beats zero.

Stage 4 — Pricing and channel test (one to two weeks)

Write a one-page offer with a price. Show it to twenty prospects through the channel you intend to use — an association, a partner, a community, direct outreach. Count conversations booked and pre-orders or pilots agreed. This tests distribution, which kills more products than technology does.

Go / no-go criteria

Proceed to an MVP when you have:

• At least three customers who used the concierge output more than once.
• At least one who paid something.
• Measured accuracy that meets the bar with review, at a cost that leaves room for margin.
• A channel that produced conversations without a warm introduction.

If any is missing, iterate on that stage rather than building.

AI-specific traps

• Demo bias. The first ten examples work; the messy hundredth does not. Test on real inputs.
• Hidden data dependency. The idea requires data the customer cannot legally or practically share.
• Platform overlap. A model provider or incumbent ships the feature natively. Ask what you would still have.
• Accuracy theatre. Measuring on examples you chose rather than a random sample.

What validation is not

It is not a survey, a landing page with a waitlist, or a pitch deck. Those measure interest, not behaviour. Behaviour — using the output again, paying, introducing a colleague — is the only reliable signal.

Related reading

• 25 AI Business Ideas for Entrepreneurs
• How to Turn an AI Idea Into a Software Product
• AI Consulting — if you would like a structured review of your validation results

Key takeaways

Interview for pain, do the job manually, measure accuracy and cost on real examples, test price and channel, and only then build. If you want a second set of eyes on an idea, book a consultation.