The first AI product you ship is mostly not the product. Auth for agents, context management, usage metering, guardrails, evals, monitoring — the visible feature sits on top of all that, and the feature is maybe a fifth of the work. The temptation is to bury the plumbing inside the product and call it done. Then the second AI product arrives, and you pay for everything twice.
We built the plumbing as a dedicated platform and put our merchant-facing copilot on it first. When usage-based pricing tiers came, metering wasn't a project — it was configuration. The real proof is the next two products, in build on the same base right now: those teams spend their time on what their product actually does, not on auth and guardrails. The platform didn't make the first product cheaper. It made every following one cheaper, and that was the point.
So the metric I'd offer: judge your first AI product by how much it lowered the price of your second. If every new agent starts from scratch, you didn't build a product — you built a very expensive demo. Platform work is boring until the roadmap gets long. Ours got long fast.