You probably know the amount of your AI quotes. The licenses too. But the time spent internally preparing data, testing answers or correcting errors: where do you count that? And the maintenance six months after go-live? That is often where budgets become less obvious.
Four projects behind the same word
Before counting, you need to know which project you are talking about. “AI project” covers situations that have neither the same price, nor the same timeline, nor the same risk:
- Deploying an off-the-shelf tool. Licenses, configuration, training: most of the cost is recurring and predictable.
- Automating a targeted process. You connect two or three existing tools and remove double data entry. The lever is not always AI, by the way.
- Building an assistant connected to your documents. You need to prepare the data, manage access rights and check the quality of the answers.
- Developing a business application that includes AI. It is a full software project, in which AI is only one component.
Published benchmarks, and what they are worth
Few players publish amounts, and the ones in circulation are heterogeneous. A study self-published in August 2026 by Denis Atlan, not audited, covering 200 B2B deployments in France, most of which its author worked on, puts the median budget at around €19,500 for an SME and €84,000 for a mid-cap. The studio Infinex, for its part, quotes pilot budgets of €5,000 to €50,000 depending on company size. International guides, such as Netguru's, show much higher amounts, for a different market and different projects.
These are orders of magnitude, not market prices. They tell you whether you are in the right ballpark, not what your project will cost.
What you pay once, what keeps coming back, and your own time
What you pay once: scoping, data preparation, development and integration, testing and security, team training.
What comes back every month: model calls, hosting, application maintenance and monitoring the quality of the answers. AWS classifies these expenses separately from initial development, and Netguru puts maintenance and monitoring at 25 to 30% of the budget. That is a provider's benchmark, not a rule.
There is a third column, almost never written down: your own teams' time. Infinex estimates it at 30 to 60 hours for a €15,000 project run over three months. Hours of workshops, testing and review, taken from the weeks of the people who already run the process.
A prototype does not cost what production costs
A prototype built in a few days proves that the idea holds. Not that it is usable every day, by everyone, on your real data. Between the two, you need to add access rights, error handling (a generative system is sometimes convincingly wrong), the edge cases the business knows and the demo avoids, and then maintenance.
That gap is nothing abnormal. It becomes a problem the day nobody has mentioned it.
The price of AI does not change, the project does
Take an assistant that answers from a single document base. The project remains relatively simple to scope. Then ask it to query the CRM, check invoices and write to the business application: you have added integrations, access rights, possible errors and tests. The price of the AI has not necessarily changed; the project has.
Each new source means data to prepare (Netguru puts this item at 20 to 25% of the budget) and software that is more or less open: a tool with a documented API connects quickly, a closed tool requires workarounds that are more expensive and more fragile. The assistant no longer makes one request, it chains several, and Galadrim points out that an output token costs 5 to 10 times more than an input token. Hence AWS's advice: measure the cost per request processed from the prototype stage.
Time saved is not money saved
A first return calculation does not need a sophisticated spreadsheet:
- the current cost of the process (cases per month, time spent, fully loaded hourly cost);
- the gain you can actually use;
- the annual running cost, including internal follow-up time;
- the payback period: the initial investment divided by the net monthly gain.
The second line is the one that misleads most. Twenty minutes saved per day for ten employees is a worthwhile gain. But it is not automatically a position eliminated, nor the same amount of money back in the bank. You need to decide what to do with the time freed up in order to assess the project's return.
Before comparing three quotes
Given a vague request, three providers price three different projects. Give them the same elements: the target process and the person who runs it, the monthly volumes, the tools involved (with or without an API), the actual state of the data and the level of reliability expected.
Then ask them to specify what will be maintained, who will check the results and which costs will remain every month. Otherwise, you risk comparing three different projects.
Author · B.Coquisart - Project manager
Sources
- Denis Atlan, "How much does an AI deployment project cost in a company?" (August 3, 2026, self-published observational study, 200 deployments in France)
- Infinex, "What budget should an SME plan for AI?" (November 6, 2025)
- Netguru, "AI development cost: full budget guide" (updated July 21, 2026, international market)
- AWS, "Optimizing Cost for Generative AI with AWS" (March 18, 2025)
- Galadrim, "How much does an AI project cost in a company?" (June 17, 2026)
