Build, Buy, or Fine-Tune
The Hidden Iceberg: Total Cost of AI
The invoice isn't the cost
Teams budget the API bill and get ambushed by everything else. Under the waterline: evals (someone builds and maintains the test sets that prove the model works), guardrails (input filtering, output checks, fallback logic), observability (logging every call, tracing failures), iteration (prompt tuning is a recurring chore, not a task you finish), data plumbing (your data, clean enough to feed a model), and on-call (when the provider degrades at 2 a.m., your feature degrades).
A defensible rule of thumb: all-in, production AI runs 3–10x the model bill. The demo cost you 5% of the real total.
warning
The classic budget failure: pricing a feature on token costs alone, shipping it, then discovering the human review and eval maintenance dwarf the API spend.