Shipping an AI MVP in Six Weeks: Scope, Data and the Demo Trap

Pick one workflow, not one feature list

An MVP that touches five workflows shallowly teaches nothing. One workflow, end to end, with real inputs and a real user, answers the question you are paying to answer.

Write down what a success looks like in a single sentence before any code. If it cannot be measured in the first month, it is not the right first scope.

Data beats model choice early

Most early quality problems come from inputs, not from the model. Fifty well-labelled real examples change more than a model upgrade.

Collect those examples before building. They double as an evaluation set, which is what turns tuning from guesswork into work.

The demo trap

A curated demo hides the cases that break the product. Every prototype looks excellent on the three inputs its author chose.

Test on inputs nobody cleaned: messy files, half-filled forms, the wrong language. The failure list from that hour becomes the actual roadmap.

Cost shows up later than you expect

Per-request costs look negligible during development and become the main line item at volume. Measure cost per completed task early, not cost per call.

Caching, shorter prompts and a smaller model for the easy 80 percent usually cut the bill more than any clever optimisation.

Decide, then commit

At the end of the six weeks, choose one of three: continue, change the scope, stop. Keeping a prototype alive without a decision is the expensive outcome.

Whatever you choose, keep the evaluation set. It is the only artefact that stays valuable after the code is rewritten.