Every product roadmap has an AI feature on it this year — usually with a deadline, rarely with a plan for what it costs to run. The demo takes a weekend. Production is where it gets interesting.
We've shipped AI features in our own products — strategy briefs in Refractr AI, voice and text practice in OsceSetGo — and the lessons repeat. Here they are.
Meter it from day one
Model calls cost real money per use, which makes an AI feature the first part of your product with a marginal cost. Refractr bills in credits; OsceSetGo enforces usage limits per plan. Both were designed alongside the feature, not bolted on after the first surprising invoice. Decide how usage maps to revenue before you ship — or your most enthusiastic users become your most expensive.
Structured beats chatty
Users don't want a chat box; they want the job done. Refractr doesn't hand marketers a conversation — it generates a strategy brief in the shape the product already uses. Constraining the model to structured output makes results predictable, testable and cheap to render into your existing screens. Save open-ended chat for the rare product where conversation is the point.
Model choice is a product decision
We run OpenAI and Gemini across our products and pick per task — different jobs justify different cost and quality trade-offs. Build the thinnest possible seam between your product and the model, so you can swap when prices or quality move. They do, every few months.
Ship one seam, then widen it
The pattern that works is boring: one well-instrumented feature, real usage data, then expand. The pattern that fails is a platform-wide “AI layer” announced before anyone has checked what a month of tokens costs. If you're planning the first kind, we can usually tell you in one call what it will take.