"What an AI feature really costs: a budgeting guide"

"What will the API bill be?" is the first question I get from almost every founder budgeting an AI feature. It is a reasonable question with a surprising answer: in nearly every system I have shipped, the model API is the smallest line in the budget. If your spreadsheet has one row for "AI costs" and it points at token pricing, you are budgeting for the wrong thing.

A production AI feature has five cost categories: the build, the evaluation suite, human review time, maintenance, and, last and least, the API bill. Miss any of the first four and the project either overruns or quietly degrades after launch. Let me take them in order, with rough proportions from real projects.

The build is the engineering around the model, and it dwarfs the prompt work. Getting a model to extract fields from a document takes an afternoon. Making that extraction survive malformed PDFs, timeouts, rate limits, hallucinated values and a run at 3am with nobody watching takes weeks: input validation, schema enforcement on outputs, retries with idempotency, queuing, fallbacks, logging. Expect 70 to 80 per cent of the build effort to be this plumbing, exactly as it would be for any integration-heavy software.

The eval suite is the line most buyers have never budgeted, and skipping it is the most expensive mistake on this list. An eval suite is a set of real, labelled examples the system is scored against, your equivalent of a test suite, except the thing under test is probabilistic. Someone has to assemble representative inputs, label the correct outputs, and build the harness that runs and scores them. Budget 20 to 30 per cent of build cost. Without it, every prompt tweak and every model upgrade is a blind bet, and you find regressions through customer complaints.

Human review time is an operating cost, not a build cost, and it belongs in the budget for as long as the feature runs. No serious system ships at 100 per cent automation on day one. If the model handles 90 per cent of cases cleanly, someone on your team reviews the other 10 per cent, and someone spot-checks the 90 too, at least initially. Price that person's hours. Often this is still a large saving over the fully manual process, but it is not zero, and pretending it is zero is how "AI project" becomes "disappointed operations team".

Maintenance is the category that surprises people who think of AI features as finished software. Models get deprecated on providers' schedules, not yours, some platforms guarantee only twelve months of availability from a model's launch, and retirement notice can be as short as sixty days. Expect a forced migration roughly every 12 to 18 months, and each migration means re-running evals and adjusting prompts. Prompts drift out of tune as your inputs change. Costs and latency shift with pricing updates. A sensible retainer is 10 to 15 per cent of build cost per year, which buys you eval re-runs on model changes and small corrections before they become incidents.

And then the API bill. For a document-heavy workflow processing tens of thousands of items a month, the model spend is typically tens to low hundreds of pounds, often less than the observability tooling watching it. There are exceptions: consumer-scale chat products, heavy multimodal processing, long-context work at volume. But for the internal workflow automation most companies are actually buying, tokens are cheap and getting cheaper. The one caveat is variance: put a hard monthly cap and alerting in place, because a retry loop bug can multiply a small bill quickly.

Put together, a realistic budget for a mid-sized feature looks something like this: if the build is X, expect roughly 0.25X for evals, 0.1 to 0.15X per year in maintenance, a recurring line for review hours that declines as confidence grows, and an API line so small you should mostly ignore it. When a proposal shows you a big build number and nothing else, the missing lines have not disappeared, they have been deferred to the point where they are yours. This is not a hypothetical failure mode: S&P Global found in 2025 that 42 per cent of companies were abandoning most of their AI initiatives before production, with escalating costs high on the list of reasons, which is what happens when the missing lines surface mid-project instead of in the budget.

The honest way to budget is in ranges, narrowed by a short discovery: a week with your real data tells you the likely automation rate, which drives the review-time line, which is usually the swing factor in whether the economics work. If you are pricing an AI feature and want a second pair of eyes on the numbers before you commit, a one-hour conversation is usually enough to stress-test the budget, get in touch via the contact form at afareed.com.