Simon Willison resurfaces a calculation from James Shore that few vendors put in their sales decks: if you double productivity with an LLM but don't halve maintenance costs, you have quadrupled technical debt. Three lines, and it changes how a CFO should read the TCO of an agent in production.
The operational point is this. The API is the most visible line item and almost never the most expensive one over twelve months. The refactor when the provider changes model version weighs more, so does the person maintaining the prompts, output monitoring, the evaluation harness that prevents silent regressions. Line items that don't make it into the initial quote and that push the break-even out by months.
Anyone evaluating an AI workflow in a company should demand three numbers from the vendor: steady-state API cost, expected maintenance hours per month, cost of a forced model swap. Without those three, there is no business case, there is a slide.
Why this matters for anyone building enterprise AI: the real cost of an agent in production is not the API, it is maintenance when the model changes version.