How much does an AI project cost: what the budget actually covers

Four situations behind the same words

An executive asking for a budget and a vendor answering are not always talking about the same thing. The vocabulary is shared, the scopes have nothing in common.

• Rolling out an off-the-shelf tool. You buy licences, configure them and train your teams. Most of the cost is recurring and predictable.
• Automating a specific process. You connect two or three existing tools and remove double entry. The lever, incidentally, is not always AI.
• Building an assistant connected to your documents. You have to prepare the data, manage access rights and check the quality of the answers.
• Developing a business application with AI inside it. This is a full software project, in which AI is only one component.

These four situations carry neither the same price, nor the same timeline, nor the same risk. Before looking for a figure, name yours.

Where the figures in circulation come from

Few players publish figures. Those who do are worth reading, provided you look at who did the counting, and how.

A study published in August 2026 by Denis Atlan documents 200 AI deployments in French B2B companies. It puts the median budget for an SME at €19,482, and for a mid-cap at €84,112. Two medians given down to the euro: read them as benchmarks, not as price lists.

Its limits are stated openly, and that is what makes it useful. It is self-published, unaudited, and its author acted as the vendor on most of the cases. Its data remains open, and therefore open to discussion.

The studio Infinex publishes pilot envelopes by company size. From €5,000 to €15,000 for an organisation under twenty employees, from €20,000 to €50,000 above eighty. These are commercial estimates, not a market survey.

International guides, such as the one from Netguru, show markedly higher figures. They describe an English-speaking market and bigger projects. Their breakdown transfers well, their amounts do not.

None of these benchmarks tells you what your project will cost. They only tell you whether you are aiming at the right universe.

What you pay once, and what comes back every month

An AI budget reads in two columns. Confusing the two is what produces the most unpleasant surprises.

On one side, what you pay once: scoping, data preparation, development and integration, testing and security work, training the teams.

On the other, what comes back every month: model calls, hosting, application maintenance and monitoring the quality of the answers. AWS keeps these expenses separate from the initial build.

That leaves 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. Those are workshop and review hours, taken out of the weeks of the people who already own the process.

Netguru puts maintenance and monitoring at 25% to 30% of the budget. The same guide observes that cumulative operation ends up matching the original build investment. The budget you sign is therefore only half the bill.

What separates the prototype from software people actually use

A prototype built in a few days proves something real. It shows that the idea holds up, and that is a very good way to start.

It proves that the thing is feasible. Not that it is usable every day, by everyone, on your real data.

Between the demo and real use, precise work remains. That is what costs.

• Access rights, so that each person sees only what they are entitled to see.
• Real data, more varied and less clean than the test set.
• Error handling, because a generative system sometimes gets things wrong convincingly.
• Edge cases, the ones the business knows and the demo avoids.
• Service continuity, backups, recovery, and the person who maintains it.

The cost gap between the two is not abnormal. It becomes a problem the day nobody announced it.

At equal scope, two projects do not cost the same amount

At comparable scope, two projects do not land on the same budget.

• The state of your data. Netguru puts its preparation at 20% to 25% of the budget. This is the line item that moves the most.
• How open your software is. A tool with a documented API connects quickly. A closed tool calls for workarounds, more expensive and more fragile.
• The level of human checking. An answer reviewed by an operator costs time every day. That is a reliability choice, to be budgeted as such.
• Growth in volume. The build cost is fixed, the usage cost follows activity. Successful adoption has a price.
• Maintaining the integrations. Every connected tool evolves on its own schedule. Those changes have to be followed, and budgeted.

An assistant that answers by consulting a single database stays simple. The same assistant, if it has to cross-check the CRM, the billing history and the product documentation, changes in nature.

It no longer makes one call, it chains several. Each step calls a model, and the cost follows. Galadrim points out that a token produced costs 5 to 10 times more than a token read.

Measure the cost per request handled from the prototype stage, as AWS recommends. That is the number that tells you whether successful adoption is affordable.

The return calculation you can run yourself

This calculation needs neither a consultant nor a sophisticated spreadsheet.

• The current cost of the process. The number of cases handled per month, the time spent on each, a loaded hourly cost.
• The gain you can actually use. Not the time theoretically saved, but the time you can genuinely reallocate.
• The annual running cost. Models, hosting, maintenance, monitoring, and the internal time spent following it.
• The payback period. The initial investment divided by the net monthly gain.

The second line is the one that misleads most. Take twenty minutes saved per day across ten people. That does not make one role saved, it makes ten people slightly less under pressure.

That result has value, but not the same value. The hours freed improve daily work. The euros saved can be defended in front of a board.

The study cited above observes a positive return after a median of 259 days, and a majority of projects profitable within the first year. Declared, like the rest.

What to gather so that three quotes are comparable

A quote is only as good as the request behind it. On a vague request, three vendors price three different projects, and you end up comparing figures that are not about the same thing. Gather these items, and the proposals become comparable again.

• The target process, described end to end, with the person who owns it today.
• The volumes, in cases, requests or documents per month.
• The tools involved, and for each one whether it exposes an API.
• The data available, where it sits, in what format and in what real condition.
• The level of reliability expected, and who checks what before an answer goes out.

That takes a few hours. It is probably the best investment in the project. Without these items, any costing is a hunch, yours as much as your vendor's.

The real question is not how much AI costs. It is what exactly you are buying, and for how long.

A build figure is negotiated once. A cost structure stays with you for years. Look at which of the two you were actually shown.