← All founder workflows

How to validate an AI wrapper idea when the model does the hard part

Validate an AI wrapper by testing the two things the model does not give you: distribution and defensibility. Deliver the output manually for ten users before building anything, and ask what they would pay to have it every week. A wrapper is a business when it owns workflow, data or distribution the model cannot supply — otherwise it is a feature with a subscription page, and the model provider or a weekend competitor removes it. Test the moat before the interface.

10 days $40 to run6 stepsUpdated 2026

The workflow

0/6 done

  1. There are only a few honest answers: it holds proprietary or hard-to-collect data, it fits into a workflow the user already lives in, it takes on liability or review, or it owns a distribution channel. "Better prompts" is not one — prompts are copied in an afternoon. If you cannot name which of these you have, the validation question is already answered and you saved yourself a month.

  2. Take requests by email or in a Slack channel, run the model yourself, review the output, send it back. This is the single most informative week available to an AI founder, because it shows you what fraction of outputs need human correction. If seven of ten need editing before they are usable, your product is a service with a model inside, and that changes the price, the margin and the roadmap.

  3. Two numbers decide the business: how often the output is good enough unedited, and what a run costs in tokens. A tool with a 40% correction rate cannot be sold as automation, and a workflow costing $0.80 per run cannot be sold at $9/month to heavy users. Both are measurable in the manual week, and both are routinely discovered after launch instead.

    Runway CalculatorFull tool
    Input
    $0.62 average cost per run, users average 40 runs/mo, price $29/mo, 120 hypothetical users
    What comes back
    $24.80 of inference against $29 revenue per user — a 14% gross margin before hosting and support. This does not survive a heavy-usage cohort. Either cap runs at 20, raise to $59, or cut per-run cost by two-thirds before anything else.

    Run it yourself — free, no signup:

    8%
    Runway
    5+ years
    Cash never reaches zero inside five years: you break even at month 12 and it climbs from there. That holds exactly as long as the growth rate does.
    Break-even month
    Month 12
    The month the borrowing stops. Worth more attention than the runway number.
    What cash ÷ burn would tell you
    12 months
    It says 12 months, because it assumes your MRR never moves. That assumption is the entire difference between running out next year and not running out.

    Under six months: cut, because you need a certain effect fast. Above twelve: grow, because growth compounds and cutting caps your ceiling. The founders who get this wrong grow when they should cut — growing feels like progress.

  4. "If someone released the same thing for free next month, what would keep you here?" The answers are your real moat, and they are usually unglamorous: the data you accumulated for them, the integration into their existing tool, the fact that you review the output. If nobody can answer, you have a demo — and demos churn hard the moment novelty fades.

  5. AI tools attract enormous free usage and very little payment. Charging in week one, before any interface exists, filters that instantly and gives you the only number that matters. Ten users paying $40 a month for a manually produced output is far stronger evidence than a thousand free trials of a polished app.

    Idea ValidatorFull tool
    Input
    AI tool drafting clinical letters for physiotherapists. 10 manual users, 7 paying $40/mo, 22% of outputs need edits, users cite "you check it before it goes out" as the reason to stay.
    What comes back
    GO, and the moat is the review, not the model. 70% conversion to paid on a manual service is exceptional, and the stated reason to stay is human verification in a liability-bearing context — which a free copycat cannot offer. Price on assurance and keep a human in the loop; do not automate the review away.

    Run it yourself — free, no signup:

    Who it's for, what it does, what they pay. The more specific the sentence, the sharper the read.

  6. After the manual phase you will know exactly which step people would not do for themselves — usually not the generation. Build that, keep the human review if it is your moat, and stay boring about the model underneath so you can swap it when prices halve again, which they will.

Questions founders ask about this

Are AI wrapper startups actually viable?
Some are. Viability comes from owning workflow, data, distribution or liability — not from the prompt, which anyone can reproduce. Validate which of those you have before building an interface.
How do I know if my AI product has a moat?
Ask ten users what would keep them if an identical free version appeared. Concrete answers — accumulated data, an integration they depend on, a human who checks the output — are a moat; "it's nicer" is not.
What margin should an AI tool have?
Measure cost per run against price with a realistic heavy-user profile before launching. Founders regularly price on average usage and discover the top decile of users consumes more inference than they pay for.
Should I offer a free tier for an AI product?
Be careful — free tiers on inference-backed products carry real marginal cost and attract the least committed users. A short trial or a small number of free runs achieves the same trust without an unbounded bill.
What if the model provider builds my feature?
Assume they might, and choose a wedge they will not: a specific industry's workflow, verified output where accuracy is a liability, or an integration with a system they will not touch. Generic productivity wrappers are the most exposed.

Next, founders usually do this

The tools used above have their own pages — MRR & Runway Calculator and Idea Validator — and the SOP SOP: Get your runway and burn under control runs the same ground in more depth. Also worth reading: the Nine Lives Doctrine, and real verdicts from ideas kitty has run this workflow on.

kitty.build runs this entire workflow for you

Every step above — the research, the competitor read, the numbers, the honest verdict — is what nine specialist AI boards do automatically when you feed her an idea. She will tell you to kill it if it deserves killing. First idea is free.

Feed her an idea — free

No card · Failed tasks are free · Your repo, your domain, your revenue