How much does an AI project cost a company?
Scoping, build, hosting, cost per request, maintenance. What an AI project costs, what moves the bill, and the running costs most quotes leave out.
6 min read
The build is rarely the part that surprises a CFO. The surprise comes six months later, when the invoice from the model provider arrives and nobody budgeted for it.
Or it comes earlier, when two quotes for the same project sit on your desk, one three times the other, and nothing in them tells you why.
An AI project has five cost lines. Two are paid once. Three are paid every month, for as long as it runs.
The five lines on the bill
Scoping. Someone has to decide what gets built, with what data, and what "working" means. It is the smallest line on the bill and the one that moves all the others. Projects that skip it get scoped anyway, halfway through the build, at build rates.
Build. The product itself: the agent or assistant, the screens around it, the connections to your CRM, ERP or helpdesk. In most projects, the AI part is the smaller share. The integrations and the interface take the time.
Hosting. Servers, databases, storage, logs. For a typical internal tool this is a modest monthly line. It grows when your data has to stay in a given country or inside your own walls.
Hosting your own model rarely pays at low volume: the servers cost the same whether anyone asks a question or not.
Cost per request. Every time the system calls a language model, you pay for it, by the volume of text sent and received. The estimate is simple arithmetic: requests per month × average text per request × the provider's price per volume of text, which every provider publishes on its pricing page (Anthropic, OpenAI). Long documents in every request, not the number of users, are what make this line grow.
Example: an assistant that answers 500 questions a day, each sending a 10-page document, is about 15,000 requests a month. Multiply by the text in each request and the provider's current price, and you have the monthly line. Send a summary instead of the full document and the line drops with the volume of text you send.
Maintenance. Your provider retires a model, your CRM changes its API, your team changes the process. Someone watches the system, fixes what breaks and improves what works. It costs less when the team that built the system also runs it: nobody has to learn the code again before fixing it.
What makes the bill move
Two projects with the same name, "a customer support AI agent", can differ several times over in cost. The difference is almost never the model. It is everything around it.
The drivers we see most often:
- The state of your data. Data spread across PDFs, inboxes and three versions of a spreadsheet has to be gathered and cleaned first, and that work is paid for.
- The number of systems to connect. An old ERP with no API, or a vendor that charges for access, can take longer than the AI itself.
- The cost of a wrong answer. An assistant that drafts internal emails can be wrong now and then. One that quotes prices to customers or reads medical documents needs review steps, tests and human checks. Each of those is build time.
- Volume. Ten users a day and ten thousand customers a day do not need the same architecture, and do not produce the same monthly bill.
- Where the data must live. A public cloud API is the cheapest option. Hosting in Europe, in your own cloud or on-premise is possible, and costs more to set up and to run.
What pushes the budget up, by type of project
The table covers a first version in production, not a prototype.
| Type of project | What pushes it up |
|---|---|
| AI agent for your customers, in your app or site | CRM and order integration, handover to a human, volume of conversations |
| Internal assistant on your documents | Number of document sources, access rights, document length |
| Automation of a back-office task | Tools without an API, formats that vary, human validation steps |
| AI inside an existing product | State of your codebase, latency targets, number of users |
We do not publish ranges we cannot stand behind for your case. The four answers further down are what it takes to get one.
On timing, we can be more precise. With us, a first version usually reaches production within 3 to 6 weeks of kick-off, then improves in iterations.
The running costs most quotes leave out
A quote for an AI project often stops at launch. The product does not.
Model usage at real volume. The demo ran on fifty test questions. Real users send longer messages, attach files and ask the same thing three times. Estimate usage on expected volume, then watch it in the first weeks.
Evaluation. You need a way to know the system still answers well after a change. That means a set of test cases, rerun on every update.
Model changes. Providers retire models and release new ones. A change can improve quality, cut cost, or quietly break a prompt that worked. Someone has to test and switch.
Human review. If a person checks part of the output, their time is a cost of the system. It belongs in the budget from day one.
The next iteration. Once the first version runs, your teams will ask for more. Plan a budget line for it now, not when the first request arrives.
A reasonable rule: plan a monthly budget for running and improving the product, not just for hosting.
Five signs a quote is incomplete
Send this list to whoever reads the vendor's PDF with you.
- No monthly line after launch.
- Model usage priced on the demo, not on your volume.
- No test set, so nobody can tell whether an update made things worse.
- Nothing about who switches models when the provider retires one.
- A timeline with no first version in production before month six: nothing in it tells you when you will see a result.
How to budget before the first call
You do not need a precise figure to start. You need four answers, and most companies already have them.
- The task. Which job the system takes over or speeds up, and how many hours it takes today.
- The volume. How many requests, documents or users per day, now and in a year.
- The data. Where it lives, who owns it, and what rules apply to it.
- The cost of an error. What happens when the system is wrong, and who catches it.
With these, anyone serious can give you a range for the build and the running cost. Without them, any figure you receive is a guess, including ours.
The hours saved give you the ceiling. If a task takes one person two days a week, that is roughly 90 working days a year: multiply by what a day of that person costs you, salary and charges included. The build plus a year of running has to come in under that figure, or bring something those days do not.
Some projects do not pass that test. We say so in the audit. Not building is also a result, and it is the cheapest one.
Start with the four answers
The free audit is one hour on your business, then a short document: what is worth building, and what is not. Book a free audit and bring the four answers above. If you only have two, come anyway: the call is where the other two get worked out.
