Building something impressive with a language model takes an afternoon. Building something you would let a customer depend on takes considerably longer, and almost none of the additional time goes into prompting.
You cannot improve what you have not measured
The single most common gap we find in stalled AI projects is the absence of an evaluation set. Without one, every change is judged by whether the last few outputs felt better, which means the system drifts sideways for months while everyone remains busy.
An evaluation set does not need to be elaborate. A few hundred real inputs with agreed-upon expectations, scored automatically and run in continuous integration, will tell you more than any amount of manual review. It also changes the conversation with stakeholders: you stop debating impressions and start reading a number.
Failure is a design surface
A prototype has one path. A production system needs an answer for every way the path can break, and those answers are product decisions rather than technical ones.
- What does the system do when it is not confident? Silence, a hedge and a handover to a person are three very different products.
- What happens when the model provider is degraded or down? Failover changes cost and quality, so somebody has to choose the trade.
- What is the system explicitly forbidden from doing, and is that enforced in code rather than in a prompt?
- Who reviews outputs after launch, how often, and against what?
Cost is an architectural constraint
Per-request cost that is trivial in a demo becomes the dominant line item at scale, and it is very difficult to retrofit. Caching strategy, model routing, context size and how much retrieval happens per call all need to be decided while the architecture is still soft.
The interesting question is not whether a model can do the task. It is what the system does on the day the model gets it wrong.
None of this is glamorous, and none of it demos well. It is, however, most of the work — and it is the part that decides whether an AI feature survives contact with real users.