How to choose an agentic AI consultant in France

I What an agentic AI consultant actually does

An agentic AI consultant decides where an AI agent brings measurable value, and where classic automation is enough, then designs, builds and puts that agent into production. An AI agent pursues a goal: it plans, calls tools and APIs, queries your data and acts inside your information systems, with guardrails and human validation on sensitive actions.

For the detailed comparison between a chatbot, automation and an agent, see what an agentic AI consultant does.

Agentic or conversational AI consultant?

The two overlap but are not the same. A conversational AI consultant builds assistants that answer: a support assistant, an internal helpdesk, usually on a RAG pipeline that retrieves answers from your knowledge base rather than inventing them. An agentic AI consultant goes one step further: the system acts. It creates records, triggers workflows and calls APIs. If you need reliable answers, a conversational assistant is often the right first step. If the value lies in actions across your systems, you need agentic expertise, including permissions and human validation.

Two examples from our own projects: a GenAI tier-1 support assistant cut resolution time by 45% and tier-1 load by 20%; for INFPC in Luxembourg, three conversational assistants were built on a shared RAG architecture, two of which are in production.

II The four types of providers in France

Searching for “agentic AI companies in France” returns structures that work very differently. The right choice depends on the size of the project, the maturity of your teams and how much of the design you want to keep in-house.

A) The AI agency

Strengths: a multidisciplinary team (design, development, data) available quickly, and a good ability to produce a demonstrable interface. Watch out for: the business model sometimes favours attractive demonstrators over integrated systems. Check who actually designs the architecture and the evaluation.

B) The specialist consultancy

Strengths: deep AI expertise, the ability to connect strategic framing, architecture and production, and a discourse suited to the executive committee. Watch out for: in large generalist firms, the team presented during the sale can differ significantly from the team that delivers the assignment.

C) The IT services company (ESN)

In France, an ESN (entreprise de services du numérique) is an IT services and staffing company. Strengths: strong delivery capacity, knowledge of your information system if it already works on it, and a contractual framework your procurement team knows. Watch out for: AI expertise varies a lot from one profile to another, and time-and-materials staffing fits poorly with a PoC that must lead to a decision within a few weeks. An ESN is often more relevant for running and maintaining the system than for the PoC.

D) The independent consultant

Strengths: a single senior point of contact who designs and builds, with no loss of information between the sale and the delivery, a lighter cost structure and high responsiveness. Watch out for: limited capacity for projects that need several profiles in parallel. Make sure they can bring in reinforcements and hand the system over to your teams.

III Seven criteria to evaluate a provider

Whatever the type of provider, compare them on the same seven criteria.

1. Production evidence — Which AI systems has the provider put into production, and which indicators were measured at delivery? A demo is not a reference.
2. Evaluation method — How is the quality of the agent measured: evaluation sets built from your real cases, robustness tests, regression checks before each release?
3. Architecture choices — Single agent or multi-agent, RAG or GraphRAG, tools exposed through MCP, coordination through A2A: the provider should justify each choice for your context, not apply it by default.
4. Guardrails and governance — Scoped access rights, human validation on sensitive actions, logging, GDPR and EU AI Act alignment. An agent acts in your systems: none of this is optional.
5. Who does the work — Will the people you meet in the first meeting design and build the system themselves?
6. What you own at the end — The code, the evaluation sets, the documentation and the CI/CD pipelines should be yours.
7. Knowledge transfer and run — Who operates the system after go-live, and how will your teams be trained to take it over?

IV Questions to ask in the first call

Eight questions, one per criterion plus the step from PoC to production:

• Which AI agents have you put into production, and what did you measure at delivery?
• How will you evaluate the agent before go-live, and after?
• Why this architecture for our case, rather than a simpler one?
• Which actions will the agent take on its own, and which will require human validation?
• How do you handle GDPR and EU AI Act requirements?
• Who exactly will work on the project, and for how many days?
• What will we own at the end: code, evaluation sets, documentation, pipelines?
• What is the plan from PoC to production, and who runs the system afterwards?

Red flags

• A demo with no evaluation plan.
• Prototypes only, and no reference in production.
• A quote with no success KPI defined at framing.
• Broad access to your systems, with no human validation.

V Timelines and what drives the budget

Most engagements follow the same sequence: audit and framing, proof of concept, hardening and testing, then production. For a well-scoped project, a proof of concept takes 2 to 4 weeks, an MVP used by real users 6 to 12 weeks, and turning a validated MVP into a stable production system 8 to 10 weeks.

Rather than a price list, look at what drives the budget: the quality and accessibility of your data, the number of systems the agent must connect to, the level of security and compliance required, and how much of the run your teams will take over. A quote given before these points are understood is a guess.

Most of the risk sits in the step from PoC to production: here is why most AI PoCs never reach production.

VI Where Omicron AI Labs fits

Omicron AI Labs sits between two of these worlds: an independent agentic AI consulting firm, led by a senior consultant who personally works on each project from framing to production. Elhadji Ngom brings over 8 years of AI/ML experience and nearly 4 years of GenAI in production, with assignments in Paris, Luxembourg and New York. We work on site across Paris and Île-de-France, and remotely.

Delivered systems include a GenAI support assistant (−45% resolution time) and a natural-language-to-SQL Data Analyst Agent (−60% query time). More about our agentic AI consultant based in Paris.