AI project budgets surprise in both directions: the model costs less than people imagine, and everything else costs more. Understanding the split avoids the two classic mistakes — under-funding integration, and over-paying for the "intelligent" part.
The real breakdown
Across our projects of the last three years, the structure recurs with striking regularity:
- 20% — the AI core. Model selection or training, prompts, evaluation. The most visible part, and the shortest.
- 30% — the data. Access, pipelines, use-case-driven cleanup, annotation. Incompressible, and where optimistic quotes lie.
- 30% — integration. Wiring into existing systems, error handling, interface, security, access rights. An AI system without integration is a demo.
- 20% — hardening. Monitoring, drift alerts, load testing, documentation, training. This is what separates a system that runs for two years from one that dies at the first incident.
Why this is good news
This split means most of the budget goes into classic, predictable, estimable engineering skills. The real technology risk — "can AI even do this?" — resolves in the first weeks, on the cheapest slice.
Hence our insistence on short prototyping: a few weeks and a fraction of the budget are enough to know whether the core works. If yes, the rest is execution. If no, you spent 15% to avoid losing 100.
The recurring costs everyone forgets
A production system then costs 15–25% of its build cost per year: inference, retraining, integration maintenance, review of escalated cases. A budget without that line is not a system budget — it is a demo budget.
Ask every vendor for that number. Its absence from a proposal says a great deal about what will be delivered.
