Shipping AI features that actually help users

One model is the wrong number of models

The strongest model is too slow and too expensive for simple tasks; the fast model falls over on complex reasoning. Production systems route: quizzes and flashcards go to quick, cost-effective models, while step-by-step exam answers go to specialised reasoning engines. The routing layer — not any single model — is the product.

Build that layer behind your own interface from day one. Models improve monthly, and you want swapping one out to be a configuration change, not a rewrite.

Track cost per generation, not per month

Token costs feel negligible until a power user finds the expensive path and runs it in a loop. We record the precise cost of every generation in real time, attributed to the user session that incurred it. That single decision makes pricing honest, abuse visible, and the monthly invoice boring — which is what an invoice should be.

Moderation is a feature, not a filter

Anything that generates content for classrooms or customers needs to check both the request and the response — against a list that the client can update without a deploy. Restricted topics change, and a moderation registry frozen into code is out of date by the first safeguarding review.

Design moderation as a first-class feature with its own admin surface, audit trail, and owner on the client side.

Design for the bad output

Models produce mediocre output sometimes. The product question is what happens next: can the user regenerate, edit, or salvage the result without starting over? The teams that win with AI are not the ones with the best prompts — they are the ones whose product stays useful when the model has an off day.

If a feature only works in the happy path, it is still a demo, whatever the launch post says.