You validate an AI idea in two weeks by testing demand before building anything real.

Week one, fake the product. Run a concierge test or put up a simple landing page, and see if anyone cares.

Then build the smallest working slice in week two. Measure whether strangers actually use it.

If the numbers are weak, kill the idea. That is not failure. That is the system working exactly as designed.

Most AI ideas die in one of two places. The pitch deck, where they are never tested. Or six months into development, where testing them is ruinously expensive. The two-week sprint moves the death earlier, when it is cheap. Or it earns you the right to keep going, with evidence.

Why do AI ideas need a different kind of testing?

Because AI demos lie. A polished demo of an AI feature feels like a finished product. It is not. Demos show the happy path: clean inputs and expected questions, guided by a presenter who knows where not to click. Real users bring messy inputs and zero patience.

This fools smart people constantly. A team watches a demo, imagines it working at scale, and greenlights months of work. Then real users arrive and the idea collapses, because nobody checked whether anyone wanted it badly enough to change their habits.

The fix is to test the idea itself, before the technology. Forget “can we build this?” for now. The question that matters is “will anyone use this, and will they keep using it?” Those are different questions, and only the second one matters at the start.

Week 1: fake it before you build it

For week one, the rule is no real product. You are testing demand with the cheapest possible stand-in. Pick one of these three fakes based on your idea.

The concierge test. Deliver the service manually, behind the scenes, while the customer experiences it as a product. Example: your idea is an AI assistant that writes weekly social posts for restaurants. For one week, you write the posts yourself each morning and deliver them through a simple chat thread. The customer gets the experience. You learn what they actually ask for and whether they would pay for it. Five customers, one week, zero code.

The Wizard of Oz test. Like the concierge test, but the customer believes software is doing the work. They type into a simple interface. A human on the other end produces the output within minutes. This tests whether the experience itself is valuable, separate from whether the AI can do it yet. Be honest with testers afterward about what happened. People are remarkably fine with it when you explain you were testing the concept.

The landing page test. Put up one page that says what the product does and what it costs. Then drive a small amount of real traffic to it, from your network or a relevant online community, and measure signups. A page nobody signs up for is an idea nobody wants, and you learned it in days for almost nothing. The page must ask for something real, like an email address or a small deposit. Opinions are cheap. Emails cost a little trust, which is the point.

By Friday of week one, you want one of two outcomes. Real humans showing real interest, or silence. Both are valuable. Silence is the market telling you the truth early.

Hands placing colorful blank sticky notes on a white wall during a product brainstorming session
Week one is about ideas on the wall, not code in a repo.

Week 2: build the smallest real slice

If week one showed a pulse, week two builds the tiniest thing that could work: not a beta, not version one, just a slice.

Define the slice. Narrow it to one user and one task. Say your idea is “AI that handles customer support.” The slice is “AI that answers the five most common questions for one client, with a human reviewing every reply.” Everything else is cut. If you cannot describe the slice in one sentence, it is still too big.

Build it ugly. Stick to existing tools. No custom infrastructure this week. The goal is a working loop. Architecture can wait. You are buying information. The asset comes later. Anything you build this week should be disposable without regret.

Put it in front of strangers. Skip friends and skip your team. Strangers from week one, plus a few new ones. Give them the real task and watch without helping. Where they hesitate, your design is wrong. Where they smile, you found something. Record everything. Memory lies. Notes do not.

Decide by numbers. Set your targets before the week starts (more on that below). At the end of the week, the numbers make the call.

Founder showing a laptop prototype to a colleague in a startup office
Build the smallest slice, then watch a stranger use it.

What should you measure?

Track three numbers, no more. Activation first, then retention signal, then willingness to pay. Any more and you will drown in data and decide by gut anyway.

Activation. Count how many people completed the core task. If ten people try your support-answerer and two get a useful answer, you have a 20 percent activation rate. Below half is a warning sign. People who cannot succeed on the first try rarely come back.

Retention signal. Look for anyone who came back unprompted or asked when they can use it again. One week is too short for real retention data, but unsolicited return interest is the strongest early signal there is. Count it carefully. It is rare and precious.

Willingness to pay. Ask directly, and make it concrete: “It will cost roughly this much per month. Want in?” Vague interest means nothing. A credit card or a signed letter of intent means everything. If nobody will pay, it is a hobby, however exciting.

Write these down before week two starts. Deciding the targets after seeing the results is how bad ideas survive.

When do you kill the idea? The kill criteria

Set these in writing before the sprint. Our default thresholds: kill it if activation is under a third or if the AI itself fails at the core task.

Kill it if activation is under a third. If most people cannot complete the core task even once, the concept is wrong. More building will not fix a task people do not understand or want.

Kill it if nobody asks to keep using it. Polite feedback (“interesting!”, “cool demo”) tells you nothing. Demand sounds like “when can I have this” and “how much.” If you heard none of that after two weeks of real exposure, believe the silence.

Kill it if the only users are being polite. Friends and colleagues will praise your prototype to your face. Strangers will ignore it. Only strangers count. If your numbers depend on politeness, you have no numbers.

Kill it if the AI is the hard part and it is not working. Some ideas need the AI to be genuinely good at something it is currently bad at. If your slice shows the technology failing at the core task, no amount of product polish helps. Park the idea and revisit it when the models catch up. That is a timing problem. It says nothing about you as a builder.

Killing an idea after two weeks costs almost nothing and teaches you a lot. Killing it after six months costs a fortune and teaches you the same lesson. Every experienced builder has a graveyard of two-week corpses. That graveyard is a sign of a working process.

What does a “go” look like?

The opposite of the kill criteria, and it is unmistakable when it happens. Strangers complete the task. A few ask for it again without prompting. At least one person commits money or a pilot agreement. The AI handles the core task well enough that the failures feel fixable.

A “go” means you earned the next test: a slightly bigger slice, a few more users, a little more realism. Validation is a staircase. Each step costs a bit more and teaches a bit more, and you stop climbing the moment the evidence goes quiet.

This is exactly how our Innovation Lab works with founders. Short, cheap experiments in sequence, each one earning the next. No six-month bets on untested ideas.

Frequently asked questions

We already built half the product. Is it too late to prototype?

No, but be honest about sunk cost. Test what exists with strangers this week using the week-two playbook. If the numbers are bad, the money already spent does not make them good. The cheapest time to kill a bad idea was two weeks in. The second cheapest time is now.

What if our idea needs custom AI training to work at all?

Then your first prototype should be a question. Ask whether a human, given the same inputs, could produce outputs customers value. If yes, the idea has demand and the training is an engineering problem worth solving. If no human could make it valuable, no model will either. Test the demand first.

How much should a two-week sprint cost?

Very little: you need a landing page and a small traffic budget. The main cost is one person’s focused time for two weeks. If someone quotes you a large sum for a “prototype,” they are quoting you a product. The whole point is that the test is cheap enough to run on ideas you are unsure about.

Can we prototype an idea for internal use, not a product?

Absolutely, and the method barely changes. In week one, do the task manually for one team and measure whether they actually use the output. In week two, build the smallest automated slice for that team. Internal tools die from neglect more than competition, so your key metric is whether anyone’s behavior changes.

What if the test succeeds but we cannot build it yet?

That is a great problem. A validated idea with committed early users is worth waiting for, and it is fundable. Document the evidence and keep the early users warm with honest updates. Build when the timing is right. Validation does not expire as fast as people fear.

Test it before you build it

Two weeks: fake it, then slice it, then believe the numbers. Most ideas will die, and each death makes the survivors stronger.

If you have an AI idea, you want a team that tests it honestly. Book a 30-minute call. We will help you design the two-week test, and we will tell you straight if the idea is not ready. That is what the Innovation Lab is for: testing ideas with evidence instead of betting on them blind.