Why the estimate is always wrong
The prototype is the visible artefact, so it anchors expectations. It demonstrably works, so remaining work reads as finishing. The estimate that follows is a fraction of reality, and the overrun is attributed to poor execution rather than to the estimate having measured the wrong thing.
The prototype proved feasibility. It proved nothing about reliability, isolation, observability, cost, recovery or maintainability — and those, not feasibility, are what production means.
What is actually left
Error paths for every external dependency. Durable state and resumability. Run traces with cost and timing. Evals with a CI gate. Access control if the product is multi-tenant. Rate limiting and abuse handling if it faces users. Cost attribution. A rollback story for a model change. Documentation sufficient for someone else to operate it.
None of it is glamorous, all of it is load-bearing, and it is almost never in the original estimate.
The 'it already works' trap
The hardest conversation is with a stakeholder who has seen the demo. From the outside the remaining work is invisible, so it reads as gold-plating, and there is pressure to ship what exists.
The useful reframing is in terms of the failure the work prevents rather than the work itself. Not 'we need durable state' but 'without this, a run that dies halfway leaves records half-written and someone reconciles them by hand.' Not 'we need evals' but 'without these, the provider updating a model changes our product and we find out from a customer.'
Stakeholders are rarely against reliability. They are against paying for work whose purpose has not been explained in terms they can evaluate.
Estimate the gap, not the prototype
The practical fix is to scope production work as its own body of work with its own list, rather than as a percentage uplift on the prototype. It makes the number defensible, and it makes it possible to descope deliberately — a team may reasonably decide to launch to fifty pilot users without full durability, provided that is a decision rather than an oversight.