What an AI Project Actually Costs
Last updated:
Three costs, not one
Traditional software has a build cost and a hosting cost. AI features have a third: the work of keeping quality where it was on launch day, as models change, content drifts and edge cases accumulate.
Leaving that out is how an AI feature that launched well is quietly disliked a year later, with nobody able to say when it got worse.
Build cost ranges
| Project | Range | Duration |
|---|---|---|
| Classification or routing | £6k–£15k | 3–5 weeks |
| Document extraction, one format family | £12k–£30k | 5–9 weeks |
| Grounded chatbot with escalation | £15k–£40k | 5–10 weeks |
| Retrieval over a document set | £18k–£45k | 6–12 weeks |
| Production agent with tools and approvals | £35k–£90k | 10–20 weeks |
What moves a project within its band is rarely the model work. It is the number of source formats, the state of the content, and how many systems the feature has to talk to.
Running cost, and how to keep it sane
Per-use cost is genuinely small per transaction and genuinely surprising in aggregate. The variable that dominates is context size — how much text you send with each request — not the number of requests.
- Send the passages you need, not the whole document
- Cache identical and near-identical requests; repetition is higher than people expect
- Route easy cases to a cheaper path before invoking the expensive one
- Batch anything that does not need an immediate answer
- Cap spend per day, and alert well before the cap
For most of our clients the monthly running cost lands between £80 and £900 depending on volume. It is almost never the largest line, but it deserves a dashboard rather than a surprise.
The third line: evaluation and maintenance
Budget 15–20% of the build cost per year. It covers re-running the evaluation set when providers ship changes, reviewing the cases the system got wrong, updating content, and adjusting prompts as your business changes.
An AI feature is a garden, not a monument. The maintenance is not a failure of the original build; it is the shape of the thing you bought.
Where the money actually goes on a build
| Component | Share of build effort |
|---|---|
| Data preparation and evaluation set | 25–35% |
| Integration with your systems | 20–30% |
| Prompting, retrieval and model work | 15–25% |
| Failure handling, monitoring, cost controls | 15–20% |
| Interface and review workflow | 10–15% |
The model work is the smallest line on almost every project we run. Clients are consistently surprised by this, and it is the most useful single fact about AI budgeting.
How to spend less without getting less
Narrow the scope to one document family, one language or one decision type. Accuracy on a narrow task is far easier to reach and demonstrate, and the second family costs a fraction of the first because the pipeline already exists.
The other reliable saving is accepting a human review step in version one. It cuts the engineering needed for edge cases substantially, and you can automate the confident cases later once you have data on where confidence is justified.
Frequently asked questions
Is there an ongoing licence fee to you?
What is the cheapest useful AI project?
Do costs fall over time?
Can we start with a pilot?
Wondering whether AI would actually help here?
Describe the task and we will tell you plainly whether it is a good fit, including when the honest answer is that ordinary software would serve you better.