An AI agency scopes, builds and ships artificial intelligence use cases inside a company. The French market holds hundreds of them. No public, independent ranking separates them today. This article therefore publishes the criteria first, the families of players second, and the questions to ask before signing.
Tandem publishes this comparison and appears in it. Better to write that at the top than in a footnote. It is also true of most pages that occupy this search, without always saying so. The only way to make the exercise useful is to set verifiable criteria, then submit to them.
The gap between AI installed and AI that pays
The first number to know before choosing an AI agency is not an agency number. In 2025, 18 % of French companies with 10 employees or more reported using at least one AI technology, against 6 % in 2023 (INSEE, Insee Première no. 2120, 21 July 2026). The rate reaches 58 % above 250 employees. The European Union average sits at 20 %.
Among smaller companies, usage is climbing fast. Bpifrance Le Lab measured 55 % of small and mid-sized firms using generative AI at the end of 2025, against 31 % a year earlier (barometer published on 13 January 2026, 4,722 responses). But only 17 % use it regularly. The whole gap between trial and installed use sits in that difference.
McKinsey named this gap in June 2025. In its report « Seizing the agentic AI advantage », the firm observes that nearly eight companies in ten use generative AI, and just as many report no significant effect on their bottom line. Around 90 % of function-specific use cases remain stuck in pilot. An AI agency is judged exactly there.
We detailed this mechanism in our analysis 80 % of companies see no return on their AI. The problem is rarely the technology. It is the choice of use cases, and the missing production step behind them.
Why AI agency rankings are not citable
As of 20 August 2026, the first page of Google on this search is almost entirely made of league tables published by market players. Agencies, schools, consultancies. Several place themselves at the top of their own ranking. None rests on an independent audit.
This is not illegal, it is regulated. Article L122-1 of the French Consumer Code allows comparative advertising under conditions. It must not be misleading. It must cover services meeting the same needs. It must compare essential, relevant, verifiable and representative characteristics. The advertiser must also be able to prove its claims at short notice.
The practical consequence is simple. A ranking with no stated criteria remains an opinion, not a comparison. And an opinion is not citable, neither by a reader who has to decide, nor by an answer engine looking for a source.
Six verifiable criteria for choosing an AI agency
The six criteria for selecting an AI agency share one property. Each can be checked before signing, with a precise question and a written answer.
- Cases in production. Ask for a process still running six months after delivery, not a filmed demonstration.
- Named clients. One company name, one reachable contact and one owned figure beat twenty anonymous logos.
- A dated first deliverable. The go-live date belongs in the proposal, not the theoretical project duration.
- Usage costs quoted. Model, hosting and connector costs are priced separately from the agency fees.
- Skills transfer. Your teams must be able to change what was delivered, otherwise every tweak goes back through a quote.
- A planned exit. The code, the credentials and the documentation stay with you, and the clause is in the contract.
One missing criterion is not a deal breaker. Six missing criteria are.
Five families of players share the French market
The French applied AI market is not a single block. Five families coexist inside it, with different economics. Choosing an AI agency means first choosing a family.
Large consultancies and IT services firms have moved in heavily. Capgemini bought WNS for about 3 billion euros, a deal announced on 6 July 2025 and closed on 17 October 2025, with the stated ambition of creating a leader in agentic AI-powered operations. Accenture reported around 2.5 billion euros of generative AI revenue for its 2025 fiscal year, published in September 2025. These players know how to run a multi-year programme. They are not built to ship a first workflow in three weeks.
Data and AI pure players form the second family. Players such as Artefact, Ekimetrics or Onepoint work on data, models and their industrialisation. This family is strong on quantitative topics and on governance. It intervenes less often on day-to-day operational automation.
Then come the automation studios and agencies, Tandem among them. Their promise is fast production of workflows and agents. The fourth family gathers AI vendors and startups. France Digitale counted 1,114 of them in France on 10 February 2026, for nearly 16 billion euros raised and more than 45,000 jobs, in a mapping co-produced with Sopra Steria Ventures. The fifth family, freelancers and marketplaces, answers a single isolated need very well.
| Family | What you get | Best for | What to watch |
|---|---|---|---|
| Large consultancies | Scoping, governance, multi-year plan | Large accounts and groups | Time until the first go-live |
| IT services and integrators | Dedicated teams, long projects | Structured IT departments | Long-term staffing dependency |
| Data and AI pure players | Models, data, industrialisation | Data-rich companies | Distance from business processes |
| Automation studios and agencies | Workflows and agents in production | SMEs, mid-caps, fund holdings | Ability to last over time |
| Freelancers and marketplaces | One building block shipped fast | A single isolated need | No overall picture |
One last marker lights up this landscape. Bpifrance funded 460 Data AI diagnostics between 2023 and 2025, plus 205 strategic advisory assignments, according to its review published on 10 February 2026. Compare that volume with the 55 % of small and mid-sized firms reporting generative AI use. The vast majority of deployments therefore happen with no structured scoping at all.
What an AI agency costs, and where the money goes
Three lines make up the budget of an AI project. Agency fees, licences, and the usage cost of the models. The first two are visible from the quote. The third holds the surprises.
Voice agents offer the clearest example. The Vapi platform layer costs around 4 euro cents per minute as of 20 August 2026, and the models are billed at their real cost on top. In a video on AI opportunities in business, I explained that many providers resell that minute with a factor of two. The service delivered can justify it. You need to know it before signing.
On licences, an enterprise copilot sits around 30 to 40 euros per user per month depending on the solution, as of spring 2026. One hour saved per month already pays for the line. On scoping, budget 5,000 to 15,000 euros excluding tax for an AI audit, part of which the Bpifrance Diag Data IA funds.
The tools an AI agency should already have broken
An AI agency that never says « this tool is not mature » is selling rather than advising. In May 2026 I ranked 37 AI tools from S to E in a LinkedIn post, based on what Tandem actually runs in production. Claude Code, n8n and Clay hold up. Others shut down or disappoint, and saying so is part of the job.
The test is easy to run in a meeting. Ask for three tools the agency removed from its recommendations in the past twelve months, and why. A precise answer points to someone who ships. A vague answer points to someone who resells. On orchestration, our position is public and owned on our n8n agency page.
What can be verified on Tandem's side
Tandem supports more than 40 companies across more than 60 AI projects, in SMEs, mid-caps and investment fund holdings. The first deliverable lands in about three weeks. On deployments run inside fund holdings, payback sits under six months, sometimes under two.
The cases are published with the client's name, which is the best test of criterion number two. InfraVia equipped more than 100 employees with AI tools. Nomination runs millions of AI executions per year, from the edges to the core of its business. Groupe Sprint processes thousands of quotes per month with no manual entry. Each page gives the context, the solution and the numbers.
Our limits are public too. Tandem is not the right choice for a five-year transformation programme steered by the executive committee of a listed group. Large consultancies do that better. We work on scopes where shipping matters more than planning.
The five questions to ask before signing
- Which process still runs at a client, six months after your delivery? A good answer names the company, the process and the volume handled.
- On what date will the first deliverable be live? A date commits, a duration commits to nothing.
- What will monthly usage cost, excluding your fees? Models, hosting and connectors can be priced, even approximately.
- Who will be able to change what you deliver, in six months? If the answer is « us », skills transfer was never planned.
- Which tools did you drop from your recommendations this year? This question separates those who ship from those who resell.
The lead times given in this video are a useful marker for the second question. On a mature AI use case, the first version takes days or weeks, not quarters. A quarterly plan on a simple topic signals scoping that never happened.
Which AI agency should your company choose?
There is no single best AI agency in France, but there is a best choice per profile. A listed group that must align governance and a multi-year plan should meet a large consultancy. A data-rich company that wants to industrialise models should meet a pure player. An SME or mid-cap that wants an automated process live this quarter should meet an automation studio.
In every case, the fastest sorting criterion stays the same. Demand one case in production, named, dated and quantified. Agencies that have one produce it in two minutes. The others talk about methodology.
