The model is the smallest part. Production readiness is an evaluation set, a cost per request, and a defined answer to the question of what happens when the output is wrong.
A demo proves a model can be right. Production requires knowing how often it is wrong, how you find out, and what the system does when it happens. Those three things are engineering work, and they take longer than the model selection everyone spends their time on.
An evaluation set you did not write last night
Without a held-out set that reflects real inputs, "it seems better" is the only quality signal available, and it is worthless six weeks in. The set needs to be built from production data, kept out of any prompt or training run, and re-scored on every change.
A cost per request you can multiply
Cost per request times expected volume is the number that decides whether the feature ships. It is also the number that most proofs of concept never produce, because at demo volume it rounds to zero.
A defined failure path
Every AI feature needs an answer to: what does the user see when the model is confidently wrong? Silent failure is the default, and it is the worst option. Fallbacks, confidence thresholds and human review paths are part of the design, not a later refinement.
Where the effort really goes
In the projects we run, model choice is rarely more than a tenth of the work. The rest is data access, evaluation, serving cost, and the plumbing that lets you change your mind about the model in six months without rewriting the product.