AI MVP development: a useful first version in weeks.

01What an AI MVP is, and isn't

An MVP is the smallest version of a product that real people use for real work. The word that matters is viable: it has to work on real data, for real users, every day. An AI demo that answers well in a meeting is not an MVP. A narrow tool that a sales team uses every morning is.

AI makes MVPs both easier and riskier. Easier, because a capable model can replace months of hand-written logic. Riskier, because a product that works 80% of the time in a demo can fail in front of customers.

02What to cut

• Features. One job, done well. Everything else goes on a list for later.
• Platforms. Web or mobile first, not both, unless the job only makes sense on a phone.
• Admin polish. Internal screens can be plain for the first months.
• Rare cases. Handle them by hand at first and automate them once you know how often they happen.
• Extra languages, unless language is the point of the product.

03What never to cut

• Evaluation. A set of real examples with expected results, so you know when a change makes things worse.
• Logging. Every AI decision recorded, so mistakes can be traced and fixed.
• A human fallback. A clear way for users to reach a person when the AI can't help.
• Authentication and data protection, from the first user.
• Payments, if paying is part of what you are testing.
• Analytics on the one or two actions that prove the product is useful.

04A week-by-week plan

Before the build, a discovery sprint of 1 to 2 weeks tests the idea on real data and fixes the scope and price. Then:

1. Week 1: the core flow. Data model, authentication, the main screen and the path a user takes to get value.
2. Week 2: the AI and the integrations. Model calls, retrieval if needed, connections to the systems the product depends on, and the evaluation set.
3. Week 3: real users. A closed group uses it for real work. You watch, log and fix.
4. Week 4: harden and launch. Monitoring, cost limits, documentation and the public release.

Simple products ship in the first two weeks; products with heavier integrations use all four. Either way you see working software every week.

05A stack that keeps you fast

The stack we use for our own products is chosen to be boring and quick: Next.js for web, React Native with Expo for mobile, Go for the backend, Postgres with pgvector for data and retrieval, and managed models from several providers so we can choose per task. It scales well past the MVP stage, so nothing has to be thrown away when the product works.

06What to measure after launch

• Activation: how many new users complete the core job once.
• Task success: how often the AI gets the job right, measured on your evaluation set and on samples of real use.
• Retention: who comes back next week.
• Cost per active user: hosting plus model usage, divided by people who actually use it.

Review them weekly and ship one improvement a week. For what an MVP costs, see our cost breakdown.

07Questions people ask