---
title: "Cost of an AI project for an SME: items and levers"
description: "A third of the businesses in France that use AI say its cost holds them back. The six cost items of a project, what running a model really costs, the seven levers that move a quote, and eight questions to judge it. No house rates: public benchmarks, dated and sourced."
url: "https://gilabs.fr/en/guides/cout-projet-ia-pme/"
lang: "en"
author: "Yann Gilliot"
published: "2026-10-01"
tags: ["Cost", "SMEs", "AI project", "Return on investment", "Fixed fee"]
---
# What drives the cost of an AI project for an SME: cost items, levers and public benchmarks

By [Yann Gilliot](https://gilabs.fr/en/about/) · published on 2026-10-01

A third of the businesses in France that use AI say its cost holds them back. The six cost items of a project, what running a model really costs, the seven levers that move a quote, and eight questions to judge it. No house rates: public benchmarks, dated and sourced.

## The short version

- In 2025, 33% of businesses in France that use AI say they are held back by costs that are too high, and 32% of those that do not use it give that reason. Lack of expertise comes first in both cases (Insee Première no. 2120, July 2026).
- Running the model is rarely the largest item in an AI project. At the public price recorded on 1 October 2026, custom extraction in Azure Document Intelligence costs 30 dollars per 1,000 pages: 4,000 pages a month come to 120 dollars.
- The cost of an AI project lies mostly in people's time: scoping, getting the data into shape, connecting existing systems, organising human validation, then running it.
- On Malt, the average day rate recorded on 1 October 2026 in France is 576 euros for a freelance developer and 666 euros for a data expert. A day rate does not tell you the price of a project: you need the scope, the deliverables and who carries the risk.
- A quote is judged against what the process costs today. Without a baseline (volumes, lead times, time spent, errors), no price is either expensive or cheap.

This guide is for owners and managers of SMEs who have a quote in front of them, or are about to ask for one, and want to understand what it is made of. It gives no rates, not even mine: a price is set on a scope, not in an article. It gives what you need to read a price: the cost items, what pushes them up or down, and public benchmarks, dated and sourced. The figures come from French sources.

## Cost is a brake, but not the first one

Insee, the French national statistics office, puts the question each year to businesses that use AI and to those that do not.

| Brake cited in 2025 | Businesses using AI | Businesses not using AI |
|---|---|---|
| Lack of expertise | 53% | 54% |
| Costs too high | 33% | 32% |

One business in three cites cost, one in two cites expertise. The two are linked: without knowing what a project is made of, you cannot judge its price, and any figure looks too high. The France Num 2025 barometer adds a benchmark: three in four French small businesses and SMEs set aside a specific budget for their digital projects, and 26% use AI solutions, twice as many as in 2024.

## Three kinds of spending that do not compare

The word "AI" covers three purchases of a different nature. Putting them on the same line is the first source of misunderstanding.

| Spending | How it is paid | What it buys |
|---|---|---|
| A general-purpose assistant | A subscription per user per month | Time saved per person, at their desk |
| An AI feature in your software | An option or an upgrade from the vendor | One automated step in a tool you already have |
| A custom project | A fixed fee or days, then a running cost | A transformed process, with your rules and your systems |

The first two are running expenses. The third is an investment, and is judged as one: by what it returns, not by comparison with a subscription. That is the subject of ["It's expensive": a category problem, not a pricing problem](https://gilabs.fr/en/blog/outil-vs-transformation-categorie/). For Copilot's public prices and what it covers, see [Copilot or custom AI](https://gilabs.fr/en/guides/copilot-ou-ia-sur-mesure/).

The rest of this guide is about the third line.

## The six cost items of a custom project

| Item | What it covers | What pushes it up |
|---|---|---|
| Scoping and measurement | Choosing the process, measuring it, setting the expected result and the acceptance threshold | A process nobody has described, several variants across teams |
| Getting the data into shape | Gathering, cleaning and structuring what the AI will have to read | Paper documents, hand-kept files, reference data that contradicts itself |
| Development and integration | The workload itself and its connection to your tools | The number of systems to connect, a missing interface, permissions |
| Validation and adoption | The review screen, the validation rules, staff training | A very low tolerance for error, many special cases |
| Operations | Hosting, model usage, monitoring | Volume, mandated hosting, high availability |
| Maintenance and change | Fixes, model changes, new cases | A scope that grows without being renegotiated |

The first four are paid once. The last two come back every month. A quote that shows only the first four is not complete.

The second item is the one most often forgotten. If your documents are in binders or your data in files nobody keeps up to date, the first job is not an AI job, and it is better to know before signing: see [Your problem is not (yet) an AI problem](https://gilabs.fr/en/blog/socle-data-avant-ia/).

## Model usage: the item people overestimate

AI is often assumed to be "expensive to run". Public prices say otherwise. Recorded on 1 October 2026, in US dollars, excluding tax, default region for Azure.

| Service | Unit | Public price |
|---|---|---|
| Azure Document Intelligence, read | 1,000 pages | $1.50 |
| Azure Document Intelligence, prebuilt models | 1,000 pages | $10 |
| Azure Document Intelligence, custom extraction | 1,000 pages | $30 |
| Mistral Large, input | 1 million tokens | $0.50 |
| Mistral Large, output | 1 million tokens | $1.50 |

A simple calculation gives the order of magnitude. An SME that processes 2,000 two-page documents a month uses 4,000 pages. With custom extraction, that comes to 120 dollars a month. Microsoft states that these prices are estimates and that they vary with the contract and the region.

Three caveats. An agent that chains several calls per case uses more than a plain extraction. Dedicated or on-premises hosting replaces this usage cost with a higher fixed cost. And model prices change fast: this table will be revised. Even so, at an SME's volumes, model usage is rarely what weighs in the total.

## People's time: the real item

What weighs is the work: understanding the process, connecting the systems, handling the cases nobody had foreseen. The freelance market gives a public benchmark. On Malt, average day rates recorded on 1 October 2026, in France, for freelancers active over the last three months:

| Profile | Average day rate |
|---|---|
| Developers | €576 |
| Data experts | €666 |
| Experienced data scientists | €681 |

These figures are platform averages, for freelancers. They do not describe the rates of a consultancy or an IT services company. And a day rate does not tell you the price of a project: the number of days is missing, and above all who carries the risk if that number is exceeded.

| Criterion | Time and materials | Fixed fee |
|---|---|---|
| What you buy | Days | A result described in writing |
| Who carries the overrun | You | The supplier |
| What it takes to start | A need, even a vague one | A scope, deliverables, an acceptance criterion |
| What you compare between two offers | The day rate | The scope and the result |

Neither mode is better in absolute terms. Time and materials suits exploration. A fixed fee suits a result that can be described. A fixed fee without a written scope is not one.

## Seven levers that move a quote

| Lever | The price goes down when | The price goes up when |
|---|---|---|
| Number of systems to connect | One tool, with a documented interface | Several tools, one of them without an interface |
| State of the data | Digital, structured, up to date | Paper, scattered files, duplicates |
| Variety of cases | One type of document, one route | Ten formats, frequent exceptions |
| Reliability required | An error is caught internally | An error leaves the company or creates liability |
| Hosting | A standard online service, in the European Union | Dedicated or on-premises hosting, required by the business |
| Volume | Stable and predictable | Highly variable, with peaks |
| Operations | Your teams take over | The supplier runs it and answers for availability |

The first and the fourth weigh most. Each extra system adds a connection, access rights and error cases. And taking a workload from "often right" to "reliable" requires more work than the first version: that is the difference between a demonstration and a system in production, covered in [What keeps a document pipeline alive in production](https://gilabs.fr/en/blog/ce-qui-tient-en-production/).

The most effective lever for bringing a price down is not negotiation. It is scope: one process, one type of document, one system, then widen.

## The cost you are not invoiced for

A project also takes up your teams: interviews during scoping, examples to provide, cases to validate while the system is tuned, training. That time appears on no quote and it conditions the result. Ask for it in writing. For an order of magnitude on the diagnostic phase, see [A 30-day AI diagnostic, one day of your time](https://gilabs.fr/en/blog/ce-que-le-diagnostic-demande-aux-equipes/).

## What to compare against: the cost of the process today

A price only makes sense against what the process already costs. The calculation takes four lines.

1. The volume: how many cases, documents or requests per month.
2. The time spent per unit, measured rather than estimated.
3. The fully loaded hourly cost of the people who do it.
4. What errors and delays cost: credit notes, penalties, chasing, cash tied up.

An example, with made-up figures. 400 cases a month, 12 minutes each, at 35 euros per fully loaded hour: 80 hours, or 2,800 euros a month, before errors and delays. It is against that figure, measured in your business, that a quote is read. The fourth point is often the heaviest and the least looked at: [the cost of a process lies in the time that elapses, not in the time worked on it](https://gilabs.fr/en/blog/copilot-deploye-rien-na-change/).

## Public funding in France

Bpifrance, the French public investment bank, offers the Diag Data IA, funded by France 2030 under the Osez l'IA plan: eight days of work by an approved expert, spread over three months at most, for an assessment, the identification of use cases and a roadmap. It is open to independent SMEs and mid-caps with 10 to 2,000 full-time equivalent staff, at least one million euros in turnover and more than one year in existence. The subsidy is 40% for eligible businesses. The current amount and conditions are on the Bpifrance page listed in the sources.

Your region and your chamber of commerce may offer other schemes. They change often: check them at source before building them into a budget.

## Eight questions to judge a quote

1. **What result is promised, and how will it be measured?** A figure, a method, a date.
2. **What is included, and what is not?** Getting the data into shape, connection to your tools, training.
3. **How much will the system cost each month once in service?** Hosting, model usage, monitoring, maintenance.
4. **Who carries the overrun?** On a fixed fee, the supplier. On time and materials, you.
5. **What happens when the AI gets it wrong?** A threshold, human validation, a named person.
6. **Where is your data hosted, and is it used to train a model?**
7. **Who owns what is delivered?** The code, the rules, the training data. And what does it take to change supplier?
8. **How much of your teams' time does the project need?** Per week, and at which stages.

A supplier who answers all eight in writing gives you something to compare. One who answers only the first gives you a price.

## Going further

Questions 1 and 3 assume a baseline figure that most SMEs do not have. That is the purpose of the [30-day AI diagnostic](https://gilabs.fr/en/services/ai-diagnostic/): your processes mapped, the gain costed in euros per priority case, and a 90-day roadmap, including when the right answer is a feature of the tools you already have. It is done on a fixed fee, set in writing before the work starts.

## Sources and references

- [Les technologies de l'information et de la communication dans les entreprises en 2025, Insee Première no. 2120 (21 July 2026)](https://www.insee.fr/fr/statistiques/9025878), Insee
- [Azure AI Document Intelligence pricing (public prices recorded on 1 October 2026)](https://azure.microsoft.com/en-gb/pricing/details/document-intelligence/), Microsoft
- [Mistral API pricing (recorded on 1 October 2026)](https://mistral.ai/pricing), Mistral AI
- [Freelance day rate barometer by profession, France (recorded on 1 October 2026)](https://www.malt.fr/t/barometre-tarifs), Malt
- [Diag Data IA: eligibility, terms and subsidy](https://www.bpifrance.fr/catalogue-offres/diag-data-ia), Bpifrance
- [France Num barometer 2025, 6th edition (15 September 2025)](https://www.entreprises.gouv.fr/espace-presse/france-num-presente-la-6e-edition-de-son-barometre-annuel-sur-la-transformation), Direction générale des Entreprises

Topics: Cost, SMEs, AI project, Return on investment, Fixed fee

## Take it further

Do you know what the process you want to hand to an AI costs you today?

Before you spend a euro on AI, you know what it will return. In 30 days: your processes mapped, the real end-to-end delay of each one, a return in hard euros per priority case and a payback measured in months. My method has three steps, map it, prioritise it, then automate it, in that order: automating a cluttered process does not speed it up, it multiplies it. That is why part of the gain I quantify for you cannot be bought, least of all from me. Quantified by the person who will build it.

[30-day AI diagnostic](https://gilabs.fr/en/services/ai-diagnostic/)

**Book a 30-minute call**: [Calendly](https://calendly.com/yann-gilabs/30min)