Connecting an agent to your software is not primarily a technical problem. The AI connectors that perform this task exist, they are documented, and an integrator can set them up in a matter of weeks. The real problem lies elsewhere: every connector forces you to finalize a rule that your organization previously had the luxury of leaving implicit.
The first part of this article explained why. To expose an action to an agent, you must name it, specify which data it concerns, and define what happens when it fails. This exercise almost systematically brings three things to the surface: decision rules that previously only existed in the heads of two people, exception handling that no one had documented, and processes with no defined stop condition.
Let’s assume you have completed this mapping. What remains is the legal framework that actually applies to you, the choice of a tool family, and seven key decisions. None of these are technical. All of them must be made by a steering committee.
The Applicable Framework in France, and What It Does Not Yet Impose
The European calendar changed this summer. Regulation (EU) 2026/1744, known as the AI Digital Omnibus Act, entered into force on July 27, 2026. It postpones the application of obligations for high-risk AI systems: to December 2, 2027, for those falling under Annex III, and to August 2, 2028, for those integrated into already regulated products under Annex I.
Hence, there is a hierarchy that the market often presents backward: for the project you are launching today, the GDPR (RGPD) is in command, and it has experienced no delays.
What Applies Today, Text by Text
| Text | What Applies Today | To Whom |
|---|---|---|
| GDPR (RGPD) | All principles and rights, including Article 22 on automated decision-making | Data Controller, therefore you |
| AI Act, Article 4 | AI literacy, reformulated as an obligation of means by Regulation 2026/1744 | Providers and Deployers |
| AI Act, Article 50 | Transparency, with a four-month transition for generative systems already on the market as of August 2, 2026 | Mainly Providers |
| AI Act, Chapter III | Postponed to Dec 2, 2027 (Annex III) and Aug 2, 2028 (Annex I) | Providers, and deployers per Article 26 |
| Civil Liability | French common law, as the EU AI Liability Directive was abandoned in 2025 | You, as defendant |
Two points deserve the attention of your executive committee.
Article 22 of the GDPR recognizes the right not to be subject to a decision based solely on automated processing, including profiling, which produces legal effects concerning them or similarly significantly affects them. The Court of Justice of the European Union clarified in its SCHUFA ruling of December 7, 2023 (Case C-634/21) that human intervention must be real, effective, and exert an influence on the final decision. The CNIL notes that in an agentic chain, human intervention placed only at the output stage does not necessarily rule out qualification under this article.
Article 4 of the AI Act has been in effect since February 2025 and is not concerned by the postponement. Regulation 2026/1744 rewrote it: you must take measures to promote AI literacy for your staff, without being required to guarantee a specific level for every individual. The obligation is lightened, but it does not disappear. Regarding cybersecurity, the Cyber France Reference Framework published by ANSSI on March 17, 2026, already provides the expected content for NIS2 compliance, regardless of the transposition timeline.
Four Families of Tools, Four Purchasing Logics
The market is organized into four families that cannot be compared one-to-one. Situate your need within one of them before meeting with a vendor: this saves you from an impressive demo that leads to a mismatched engagement.
| Family | What You Are Buying | Point of Vigilance |
|---|---|---|
| Unified System | An integrated application suite, driven by a proprietary orchestration engine | High dependency, reversibility must be negotiated in the contract |
| Visual Canvas | A composition interface accessible to business teams, with a connector library | Proliferation of ungoverned flows outside IT department control |
| Code-First Approach | An execution engine driven from your repositories and CI/CD pipeline | Assumes a team capable of maintaining it long-term |
| Meta-Harness | A supervision layer for autonomous agents already in place | Scope still narrow, especially regarding software engineering |
A common threat looms over all four: model providers are progressively integrating orchestration, memory, and tool calling into their own interfaces. A three-year commitment must therefore anticipate what happens to the contract if the purchased capability becomes native elsewhere.
Seven Decisions to Finalize
Each leads to a specific decision to be made, not just an intention to be displayed.
1. Classify Actions by Risk Level Before Choosing Tools
The CNIL recommends classifying possible actions on each service: data access, modification, deletion, transfer outside the system, and the nature of the data concerned.
Your decision is the short, written list of actions that require human validation before execution. This must be made by a committee, not in a configuration file.
2. Arbitrate Between the Three Access Routes Based on Nature of Need
A classic API remains the right choice when the action is stable, high-volume, and deterministic: predictable cost, trivial audit. MCP takes the lead when the agent must discover available tools and compose an unforeseen sequence. A2A handles dialogue with agents from a partner or third-party vendor.
Your decision consists of freezing these boundaries and not allowing a discovery protocol to be installed on critical transactions simply for ease of integration.
3. Treat Protocol Maturity as a Contractual Criterion
The MCP specification of July 28, 2026, brings three elements that interest you: a stateless architecture executable behind your load balancers and existing tooling, a guaranteed twelve-month delay between the announcement of a feature's obsolescence and its withdrawal, and authorization manageable via a central identity provider.
Your decision concerns the clause in your vendor contract: which version, what notice period, and who pays for migration.
4. Have the Connector, Not the Agent, Carry Identity
The same version validates the issuer parameter in authorization responses, which protects against confusion between OAuth providers when a client connects to multiple servers. The broader subject of an agent's own identity and delegation of authority is on the roadmap for the A2A project and is not yet stabilized.
Your decision: Does the agent act under its own identity, traced and revocable, or does it borrow the user's? The second choice is faster but makes any subsequent attribution impossible.
5. Partition Memory and Set Durations
The CNIL recommends that each agent have a dedicated and isolated memory, limited in size, with automatic expiration, and that the execution environment also be partitioned.
Your decision is the retention period selected for each memory space, inscribed in your record of processing activities rather than deduced from the tool's default setting.
6. Log at the Action Level, Not the Request Level
The CNIL memo describes what must be reconstructible for each task: data utilized, agents involved, third-party services solicited, exchanges, and their chronology. This requirement is not yet enforceable under the AI Act, for which logging falls under the postponed chapter. It is enforceable under the accountability principle of the GDPR and remains the only way to isolate a failed agent without stopping the entire chain.
Your decision is the scope of the log and the person authorized to read it.
7. Keep Orchestration Competence In-House, Including Business Units
Among French companies already using AI, 53% declare themselves held back by a lack of expertise, and this motive dominates among the largest firms. The orchestration layer shifts technical competence toward process competence: your business managers must know what an agent can do, what it must not do, and how to read an execution trace. This topic connects to the broader issue of the cost of IT outsourcing for industrial firms.
Your decision is the number of people, combining IT and business units, capable of challenging an architecture proposed by a provider. Below two people, the layer is subcontracted, not governed.
What This Demands of Your Organization
Reread the list. Not a single one of these seven decisions can be made in front of a screen. Each requires you to finalize a rule that the organization was previously bypassing quite well, which explains why these projects rarely fail for engineering reasons.
Therefore, the RESONANCE™ framework treats this phase as a preliminary audit: the question is not whether you can automate, but whether you know how to describe what you do today. The industrial ETIs in the Occitanie region that we advise almost all stumble on the same realization mentioned in the first part: an application landscape that is more heterogeneous than the official mapping suggests.
Your window of opportunity is favorable. The heaviest obligations of the European AI Act do not apply until December 2027, and the dominant protocol has just adopted an actionable lifecycle policy. You have a little over a year to specify your processes under good conditions rather than under calendar pressure.
These Seven Decisions Commit You for the Duration of the Contract
They are made once. Our consulting interventions focus on this precise phase: arbitration, not tooling.
Frequently Asked Questions
What is the immediate regulatory constraint on an orchestration project in France?
The GDPR (RGPD), not the European Artificial Intelligence Act. Regulation (EU) 2026/1744 postponed the obligations for high-risk systems under Annex III to December 2, 2027, and those under Annex I to August 2, 2028. The GDPR, however, applies fully starting today, including its Article 22 on automated decision-making.
Should we expose our systems via a classic API, MCP, or A2A?
Via a classic API when the action is stable, high-volume, and deterministic: the cost is predictable, and the audit is trivial. Use MCP when the agent must discover available tools and compose a sequence that was not planned. A2A answers another question: dialogue between your agents and those of a distinct organization. The rule of prudence is not to allow a discovery protocol to be installed on your critical transactions simply for ease of integration.
How long must we keep execution logs?
The six-month duration often cited appears in Articles 19 and 26(6) of the AI Act; it concerns only high-risk systems, and it is not required before December 2027. Today, your retention period falls under the GDPR's accountability principle and your own decision, recorded in your processing register.
Sources Log
All addresses below were accessed and verified in August 2026.
- Regulation (EU) 2026/1744 of July 8, 2026, on digital omnibus AI legislative package, full text. eur-lex.europa.eu
- Distribution of obligations between provider and deployer, Articles 12, 19, and 26 of the AI Act. village-justice.com (in French)
- 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 (in French)
- CJUE, December 7, 2023, Case C-634/21, SCHUFA Holding, cited and analyzed in the memo above.
- 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 (in French)
- "MCP gets an enterprise makeover," The Register, July 29, 2026, regarding the specification of July 28, 2026. theregister.com
- Google Cloud, "Google Cloud donates A2A to Linux Foundation," June 23, 2025. developers.googleblog.com
