Early AI business cases often count the build cost and the model usage fee, then claim a large saving. Six months later the numbers look different, because a set of costs that didn't appear in the proposal are now visible. None of them is a surprise to people who have run such projects. All of them can be budgeted up front.
The costs that get missed
- Human review — someone still checks outputs, handles exceptions and approves actions. Review time often becomes the largest recurring cost and the main limit on savings.
- Integration — connecting to the systems of record, handling their quirks and keeping the connections working when those systems change.
- Data preparation — cleaning, organizing and maintaining the documents and records the system relies on. Knowledge bases decay without owners.
- Evaluation and monitoring — test sets, sampling, alerts and the time to investigate failures.
- Security and compliance — vendor assessments, access reviews, audit support and legal review.
- Change management — training people, adjusting processes and handling the period when old and new ways run together.
- Model and vendor change — models are updated or retired, prices change, and prompts and tests need rework.
Budget by lifecycle, not by build
Split the cost into three parts: getting to the first decision point, getting to reliable production, and running it each month. Present all three, with the monthly run cost per completed task. A project that looks attractive on build cost and unattractive per task at normal volume is better discovered in the proposal than in the review.
A useful check: can you state the cost of one completed task, including the human minutes? If not, you don't yet know whether the project saves money.
What to do about it
- Measure review time in the pilot — and set a target for reducing it as quality improves.
- Name owners for the knowledge and data — with a review schedule.
- Reserve a maintenance allowance — a fixed share of the build budget per year is a common planning approach; set yours from your own history.
- Re-run the numbers at each gate — with measured costs replacing estimates.
Projects with honest cost models make fewer promises and keep more of them. They are also the ones that survive a finance review, because the numbers at the end look like the numbers at the start.