The first question every executive asks about an AI project is “what will it cost?” — and the honest answer is that the number they are imagining is almost always the wrong number. They are thinking about the model licence or the API bill. That is the smallest line on the invoice. The cost of enterprise AI lives in the work around the model: data, integration, evaluation, oversight, and the run-rate that never goes away.
This is a grounded 2026 breakdown of where the money actually goes, with a three-year total-cost-of-ownership model you can adapt. The figures are ranges, not quotes — your numbers will vary with industry, data quality, and regulatory load — but the shape of the spend is consistent across almost every engagement we run.
The five cost centres of a production AI system
For a single production use case, the spend distributes roughly like this:
| Cost centre | Share of first-year spend | What it covers |
|---|---|---|
| Data preparation | 25–35% | Sourcing, cleaning, labelling, access controls, pipelines |
| Engineering & integration | 25–35% | Building the system, wiring it into your stack, UI/UX |
| Evaluation & safety | 10–20% | Golden sets, eval harnesses, red-teaming, oversight design |
| Run-rate (inference + infra) | 10–20% | Model calls, vector stores, hosting, observability |
| Model / licence fees | 5–15% | Foundation-model API or self-hosting GPU cost |
Read that table twice. The line everyone fixates on — model fees — is the smallest one. A platform with a steep learning curve or heavy configuration can cost two to three times its licence fee in implementation alone. If a vendor quotes you a price that is mostly licence, they are either hiding the implementation cost or they have not thought about it.
What a first build actually costs in 2026
Typical first-build ranges we see for a single, well-scoped production use case:
- Lightweight assistive use case (internal copilot, document Q&A over a bounded corpus): US$120k–$220k build, then 20–30% of that per year to run.
- Customer-facing or workflow-embedded system (support automation, claims triage, agentic process): US$250k–$500k build, then 25–40% per year.
- Regulated or high-risk system (credit, healthcare, anything under EU AI Act high-risk obligations): add 30–50% for documentation, traceability, and conformity work.
These assume you are building one use case properly, not buying a platform licence and hoping. The temptation is to spend big on a platform and treat use cases as free add-ons. In practice every use case carries its own data, integration, and evaluation cost — the platform amortises infrastructure, not the work.
The run-rate trap
The build is a project. The run-rate is forever, and it is where budgets quietly fail. Budget 20–40% of build cost per year to keep a system healthy:
- Inference cost, which scales with usage — model this at 1x, 2x, and 5x volume
- Monitoring and observability tooling
- Periodic re-evaluation as the world and your data drift
- Occasional retraining or prompt/retrieval tuning
- Human oversight time for high-stakes outputs
The most common budgeting mistake is treating inference as a fixed cost. It is a variable cost that grows with success. A system nobody uses is cheap to run; a system everyone uses can cost more in inference than it cost to build. Model the usage curve before you launch.
A three-year TCO model you can reuse
Organisations that build a three-year TCO before vendor selection are 2.8x more likely to stay in budget. The structure is simple:
- Year 0 (build): data + engineering + evaluation + initial infra. The one-time number.
- Years 1–3 (run): for each year, sum inference (at your projected volume), infra, monitoring, re-evaluation, retraining, and oversight. Grow inference with your adoption forecast.
- Switching reserve: add a line for the cost of changing vendors — 40–60% of build — and use it to value portability. A portable system has a smaller reserve, which is real money.
Sum it. The headline build quote is usually 40–55% of three-year TCO. If a vendor will only discuss the build number, you are looking at half the picture.
When does it pay back?
Honestly: not as fast as the pitch decks suggest. Typical enterprise AI ROI takes two to four years, and only about 6% of organisations see significant financial impact in under a year. The teams that get there fastest do the same three things: they scope one high-value use case rather than a portfolio, they define a baseline metric before launch so the gain is measurable, and they expand only after the first system has paid back. Spread thin across five pilots, the payback never arrives — which is exactly why most AI budgets disappoint.
The cheapest AI system is the one you scoped correctly and measured honestly. The most expensive is the one you bought on a licence quote and discovered the real cost of after you had committed.