Prove demand before proving the demo
An AI demo proves the mechanism can work. It says nothing about whether anyone will change their behaviour or open their wallet for it. Before writing product code, find ten people with the problem and ask what they do about it today. If the honest answer is 'nothing, mostly', you are selling a vitamin in a market that tells itself it takes no vitamins.
The strongest signal is an ugly existing workaround: spreadsheets, duct-taped scripts, hourly workers doing it manually. People who already pay in time or money to solve the problem badly are the ones who will pay to solve it well.
The concierge test
Before automating a workflow, do it manually for a few users, yourself, with the AI helping you behind the curtain. You learn the real edge cases, the quality bar users accept, and what the workflow actually is (which is rarely what you assumed). Several of our best product builds started as a founder running the service by hand for a month.
The concierge test also prices the thing: what people pay a human (you) to do is evidence about what the software can charge to replace.
The honest MVP
An MVP is not a small version of everything; it is the full loop for one audience: signup, the core job, payment, return visit. If the core value is AI output, the MVP must include the unglamorous guardrails: what happens when the model is wrong, when input is garbage, when usage spikes.
Instrument everything from day one: activation, retention at week one and four, cost per user, quality complaints. These four numbers decide the next six months, and you cannot retroactively collect them.
- One audience, one job, full loop
- Guardrails are part of the product, not a later phase
- Instrument activation, retention, cost and quality on day one
The decision gates
Set the bar before you launch: what retention and cost numbers mean iterate, what they mean pivot, and what they mean stop. Writing this down protects you from the founder's version of sunk cost, which is a roadmap justified by hope. Killing a validated dead end in week six is a victory; most AI product post-mortems are week-six decisions made in month fourteen.