Select your language

+33 6 13 04 56 39
EXECUTIVE INSIGHTS | 8 min read | 14 Nov 2026

You manage a mid-market company (ETI) with 200 to 2,000 employees. Your ERP has been running for eight years, your CRM was deployed in 2022, document management lives in one silo, payroll in another, and three spreadsheets act as the glue because no one ever had the budget to connect it all properly.

Does this scenario sound familiar? It describes most digital departments of French ETIs that I advise from Toulouse. And this is precisely the environment where AI agent orchestration changes the game.

The principle is simple. Rather than migrating your systems to a single, monolithic platform, you place a connection layer on top of them: adapters, called connectors, that explain to a program what each application can do, on which data, and what it is permitted to do. The agent then executes the sequence of steps that your teams used to perform manually.

That’s the promise. Here is the surprise. Before automating anything, this layer functions as a measurement instrument: to plug in a connector, you must write down in black and white exactly what each process does and what to do when it fails. Many of your processes will not survive this exercise, and this is where the first value of the project lies.

A quick clarification upfront, contrary to popular market belief. The primary regulatory constraint right now is not the European AI Act, the bulk of which has been postponed to December 2, 2027, by Regulation (EU) 2026/1744. It is the GDPR (RGPD), which has never ceased to apply. The second part of this series covers this in detail.

How Agentic AI Actually Works

In July 2026, the CNIL (French Data Protection Authority) and the AI and Digital Council published a joint memo describing the architecture, stripped of marketing hype. It consists of three components. An orchestrator agent, whose generative model serves as a natural language interface. Specialized agents that it coordinates, each assigned a specific task. And external services—applications, databases, research tools—that they mobilize to take action.

Two data management mechanisms coexist, and the distinction matters to you. Context carries the history of a specific process execution and disappears with it. Memory is persistent and independent of individual executions. Each agent has its own, and a shared memory may also exist.

This brings us to the question that interests you: how action is executed. Interactions between the orchestrator, specialized agents, and third-party services rely on standardized exchange protocols. The memo explicitly cites MCP, developed by Anthropic, and ACP, developed by OpenAI. MCP is now hosted by the Agentic AI Foundation, a structure of the Linux Foundation, which takes it out of the purview of a single vendor.

A third protocol, A2A, addresses a distinct problem: enabling dialogue between agents belonging to different organizations without them sharing their systems. Google Cloud donated it to the Linux Foundation in June 2025, alongside Amazon Web Services, Cisco, Microsoft, Salesforce, SAP, and ServiceNow.

However, a connector is not just a pipe. It is a declaration. To expose an action to an agent, you must name it, describe its expected inputs, describe its response, and specify what happens when it fails. In other words, you must fully specify the process.

A process that you cannot describe to a connector was never a process. It was a shared habit.

What the Connection Layer Uncovers

On July 21, 2026, Insee (French National Institute for Statistics) published the results of its 2025 survey on information technologies in companies. Among French companies that do not use any AI technology, 38% cite incompatibility with their existing equipment, software, or systems. Among those with 250 employees or more, this figure rises to 46%.

Therefore, the obstacle cited by companies themselves is not the AI model. It is the plumbing, and this is good news: the work shifts to where you have control. You will see three things surface.

Unwritten Decision Rules

Approving a customer credit memo, unblocking an order, granting a pricing exception: the threshold exists in the heads of two people and nowhere in the system. As long as a human arbitrates, the absence of a formal rule costs nothing. As soon as you must expose it to an agent, the rule must be defined definitively.

Exception Handling

Agents fail on edge cases, and edge cases are exactly where your teams were silently adding value. An automation that gets stuck on 30% of cases does not mean the technology is immature, but rather that 30% of the workflow relied on judgment that was never documented.

The Absence of a Terminal State

The CNIL memo notes that exchanges between agents multiply in case of repeated failures and can lead to massive data transfers. A process without an explicit stop condition creates a loop. On a spreadsheet, this loop is called a follow-up email. In an agentic chain, it is called a consumption invoice, and the only reliable safeguard is a predefined ceiling: a token budget, maximum number of calls, or a duration beyond which the process is interrupted.

Where Value Materializes

Still according to Insee, the organization of corporate administrative processes has advanced by 21 points in one year to become the second most declared purpose of AI, at 38% of the salaried workforce, just behind IT security (39%). Among user companies, 73% cite cost optimization, including the automation of repetitive tasks, and 64% cite improved quality and reliability of processes.

These figures are based on declared usage, not measured gains. No primary source publishes average ROI for agent orchestration, and the ranges circulating come from vendors without a measurement protocol. Produce your own: end-to-end cycle time, manual rework rate, number of re-entries from one application to another for the same business object. Three indicators are sufficient, provided you record them before deploying.

Your most profitable area is repetitive tasks that cross two or three applications without any single one being master of the process: reconciling a purchase order, a delivery note, and an invoice, or updating a customer record between the CRM and the ERP. The agent does not replace your ERP; it bridges the gap. Industrial ETIs almost all encounter the same obstacle here—an application landscape more heterogeneous than the official mapping suggests, a subject addressed in our market analysis of industrial ERP software in France.

The decisive argument when facing a heavy integration project can be summed up in one word: reversibility. Unplugging a connector does not destroy the source data, whereas a migration commits you for ten years.

Two architectural effects weigh on your ROI analysis. In your favor: the orchestrator assigns each sub-task to the least expensive model capable of executing it correctly, meaning costs are managed by design rather than by tariff negotiation. Against you: passing context from one agent to another adds latency to each jump. Your first processes to orchestrate should therefore be those that tolerate a few seconds delay, overnight processing, or a queue, rather than a live customer response.

Where to Start

The first deliverable of an orchestration project is not an agent in production. It is a list: processes whose rules you know how to describe, those you do not, and those that no one realized relied on a shared spreadsheet.

This list is produced in a few weeks and costs only scoping time. Above all, it holds value independently of what follows: even if you orchestrate nothing, you will have documented your information system, something ten years of successive projects had failed to achieve.

Decisions remain. Which protocol for which usage, which family of tools, which obligations genuinely weigh on you rather than your vendor, and who maintains control internally. This is the subject of the second part: seven decisions before you wire in an agent.

Assess the maturity of your processes before automating

Our digital transformation diagnostic locates areas lacking specification in just a few minutes and indicates which process to start with.

Start the free Readiness Assessment

Frequently Asked Questions

Practically speaking, what is a connector?

It is a description, exploitable by a program, of what an application can do: the name of the action, the information it expects, the response it returns, and what happens when it fails. A connector does not move your data to a new tool; it makes an existing action callable from the outside.

Does orchestration replace our Enterprise Resource Planning (ERP) system?

No. It bridges systems that were never designed to talk to each other. The decisive advantage over a heavy integration project is reversibility: unplugging a connector does not destroy the source data, whereas a migration commits the company for ten years.

How can I prevent an agentic chain from entering an infinite loop?

By setting ceilings before deployment, not after an incident: token budget, maximum number of calls, execution duration beyond which the process is interrupted. The memo from the CNIL and the AI and Digital Council notes that exchanges between agents multiply in case of repeated failures and can lead to massive data transfers between services.

Sources Log

All addresses below were accessed and verified in August 2026.

  1. CNIL and AI and Digital Council, "Agentic AI and personal data protection" ("IA agentique et protection des données personnelles"), memo from July 2026. cnil.fr
  2. Insee Première n° 2120, "Information and communication technologies in enterprises in 2025" ("Les technologies de l'information et de la communication dans les entreprises en 2025"), July 21, 2026. insee.fr
  3. Regulation (EU) 2026/1744 of July 8, 2026, digital omnibus AI legislative package, full text. eur-lex.europa.eu
  4. Google Cloud, "Google Cloud donates A2A to Linux Foundation," June 23, 2025. developers.googleblog.com
Cookies user preferences
We use cookies to ensure you to get the best experience on our website. If you decline the use of cookies, this website may not function as expected.
Accept all
Decline all
Read more
Marketing
Set of techniques which have for object the commercial strategy and in particular the market study.
Quantcast
Accept
Decline
Unknown
Unknown
Accept
Decline
Analytics
Tools used to analyze the data to measure the effectiveness of a website and to understand how it works.
Google Analytics
Accept
Decline
Advertisement
If you accept, the ads on the page will be adapted to your preferences.
Google Ad
Accept
Decline
Save