GuidesAI budget
How much an AI project costs: what drives the price
Two budgets, not one
People think of the build cost. They forget the running cost. An AI project has both, and the second one lasts.
- The buildOnce, to design, develop, test and ship.
- The runningEvery month, for models, hosting and maintenance.
We give no price grid here: it would be wrong for your case. Here is what pushes each item up or down.
What drives the build price
| Item | Pushes the price up | Pushes the price down |
|---|---|---|
| Data | Scattered, no API, needs cleaning | One central tool with an API |
| Scope | Several use cases at once | A single, well-bounded one |
| Actions | The agent writes to your systems | Read-only |
| Interface | A dedicated application | An email report, an existing tool |
| Stakes | Costly errors, heavy testing | Harmless errors |
| Compliance | Sensitive data, imposed hosting | Public data |
The first item dominates. Two projects identical on paper can cost one to three times as much depending on the state of the data.
What drives the running cost
- VolumeNumber of requests, size of the documents read each time.
- The modelA powerful model for everything, or a light model for sorting and a powerful one for reasoning.
- FrequencyAn agent running every fifteen minutes does not cost what a weekly report costs.
- HostingOf the application and the data.
- MaintenanceAPIs change, models are replaced, content evolves.
The levers to reduce it
- 1Rules before a modelWhat a rule handles needs no AI.
- 2The right model per task
- 3CachingWhat repeats is paid for once.
- 4Short contextsGive the model the useful passages, not the whole file.
- 5A capAlfred’s agent works under a capped budget: a runaway loop stops by itself.
How to read a quote
A serious quote clearly separates:
- scoping and audit;
- the build, with the scope of the first version;
- the estimated monthly running cost, with the volume assumptions;
- maintenance;
- what is not included.
Be wary of a price with no volume assumption, and of a flat fee that says nothing about running costs.
Working out the return
The calculation is simple and should be done beforehand.
- The time spent on the task today, multiplied by its hourly cost.
- The share the automation really frees up, review included.
- Minus the monthly running cost.
If the result does not pay back the build within a reasonable time, change use case. Our guide to the AI audit helps you choose.
Start small
A first read-only project, on a single use case, limits the risk and gives real running figures. The next step is sized on those figures. For choosing a provider, see AI agency, freelancer or in-house team.
Frequently asked questions
- What costs the most in an AI project?
- Rarely the model. The biggest item is almost always access to the data and putting it in order, then testing on real cases. The model call itself weighs little at the start.
- How are AI models billed?
- By usage, according to the volume of text sent and produced, measured in tokens. The most powerful models cost several times more than light ones. Choosing the model per task is therefore a cost lever.
- How do you avoid a runaway bill?
- With a spending cap per agent and per day, alerts, light models for simple tasks, and caching for whatever repeats.
