The maintenance bill nobody quotes for AI features
The build is quoted. The upkeep usually is not. Model deprecations, price moves and prompt drift are recurring costs, and they belong in the estimate.
An AI feature is not a website that sits still once it is finished. The ground under it moves: providers deprecate models, prices change, behaviour shifts between versions, and the data feeding it drifts. None of that is a defect. All of it is a cost, and it should appear in the quote rather than as a surprise in month four.
What actually recurs
- Model deprecations. The version you shipped on will eventually be retired. Migrating is rarely a config change — output shifts subtly and your evaluation set has to be re-run.
- Price and rate-limit changes. Usually downward per token, but throughput limits and tiering change too, and a feature priced against last year assumptions can quietly stop making margin.
- Prompt drift. Not the prompt changing — the world around it. New product names, new policies, new edge cases staff started sending.
- Data drift. The documents behind a retrieval system age. Nobody notices until an agent cites a policy that was replaced in March.
- Ordinary software rot. Dependencies, OS updates, the September mobile releases, expiring certificates. AI features do not get an exemption.
Why it is left out of estimates
Partly optimism, partly because the first three months genuinely are quiet. The bill arrives later, and by then the project is “done” and nobody owns it. We would rather name it up front and be beaten on price by someone who does not, than win the work and have the conversation in month six.
What a support arrangement should cover
Monitoring that reaches a human rather than a dashboard nobody opens. Re-running the evaluation set when a model version changes. Dependency and OS updates. A defined response when accuracy drops below the threshold you agreed at launch. And a quarterly look at cost-per-request, because that number moves.
The honest framing
If machine-assisted development made the build shorter — and it did — that saving belongs in your quote, not silently in our margin. The flip side is that the running costs of an AI feature are real, ongoing and worth budgeting. A cheaper build with no support plan is not cheaper. It is the same money, paid later, usually in a hurry.