Planning allowance: 3–5 working days. Pin the model, endpoint, prompt, retrieval corpus, tool permissions and output constraints. Record date, service settings, region and actual token accounting.
Test at least: the current system; a reasonably improved prompt/schema version; and a cheaper plausible unadapted alternative. Add retrieval where factual access is the diagnosed problem. Include a ready-made specialist API when the workload has a credible one: for example, Extend describes a document extraction endpoint, but its accuracy must be tested locally.
Run the same task definition and record errors rather than silently discarding timeouts, invalid outputs or rejected documents. Report latency percentiles under representative concurrency, not only a single warm request. Keep optimisation effort bounded and comparable across candidates.
Gate: if this baseline already meets the business goal at acceptable cost, stop or redirect the work. Save the evidence so the next supplier cannot reset the comparison to an inferior starting point.