AI Act kit for AI agents

What the kit contains

The kit is made of two files, in French: a 9-page PDF guide and an Excel register ready to fill in. It brings together four tools:

• a four-question classification grid, to place each agent and each use;
• the deployer checklists: the common baseline and transparency, which apply today, then high risk, to prepare for December 2027;
• the provider checklist, in case an agent turns you into one;
• the register of AI agents, which suggests a risk level for each agent and dates its reviews.

The kit is designed to help with compliance: it organises the work, it does not do it for you. This kit does not constitute legal advice.

The timeline

The Digital Omnibus on AI, in force since 27 July 2026, postponed the high-risk obligations. The delay leaves a year to get organised, not an exemption: the register and the classification are done now.

Situation as of 9 October 2026. Source: Regulation (EU) 2026/1744, known as the Digital Omnibus on AI, published in the Official Journal of the European Union on 24 July 2026.

Classify an agent in four questions

Ask the questions in order, for each agent and each use: the level depends on what the agent does, not on the technology.

1. Is the use a prohibited practice under Article 5? If so, the use is banned: stop it. If not, go to question 2.
2. Does it fall within an Annex III area? If not, go straight to question 4.
3. Does it contribute to a decision about a person, or profile them? If so, it is high risk: Article 26 checklist, from 2 December 2027. If not, and if the Article 6(3) exception is documented, go to question 4.
4. Does it interact with people or publish generated content? If so, transparency applies (Article 50, in force). If not, its risk is minimal: the common baseline of the checklist is enough.

Annex III areas: biometrics, critical infrastructure, education, employment and HR, access to essential services (credit, life and health insurance, public benefits), law enforcement, migration, justice and democratic processes.

Levels add up: a high-risk agent that interacts with people is also subject to transparency, and both checklists apply.

The Article 6(3) exception covers, among other cases, a system that only performs a narrow procedural or preparatory task and does not materially influence the decision. It never applies to profiling, and the provider documents its assessment before placing the system on the market or putting it into service.

Obligations by role

The role is decided use by use. By default, a company that configures agents on a platform for its teams is a deployer (Article 3(4)): it uses an AI system under its own authority, in a professional context. It is a provider (Articles 3(3) and 25) if it develops a system and puts it into service under its own name; for a high-risk system, it also becomes one by putting its trademark on it, substantially modifying it or changing its intended purpose.

Deployer: the common baseline, for every agent

• The agent is in the register, with a named owner.
• Users have been made aware of AI, its limits and its risks (Article 4).
• The use is not a prohibited practice (Article 5).
• Personal data is covered by the GDPR record of processing, with an impact assessment where needed.
• A review date is set, and the agent is reassessed whenever its use changes.

Transparency by role (Article 50, in force since 2 August 2026)

• Telling people they are interacting with an AI, from the first exchange, unless it is obvious, is a provider obligation (Article 50(1)): check it with your platform, or ensure it if you build the agent yourself.
• Machine-readable marking of generated content is also a provider obligation (Article 50(2)); systems already on the market before 2 August 2026 have until 2 December 2026.
• The deployer informs people exposed to emotion recognition or biometric categorisation (Article 50(3)).
• The deployer discloses deep fake images, audio or video; it also discloses generated text published to inform the public on matters of public interest, unless that text has undergone human review under editorial responsibility (Article 50(4)).

Deployer of a high-risk system (Article 26, from 2 December 2027)

• The agent is used in line with the provider’s instructions for use (26(1)).
• A competent, trained person with the necessary authority provides human oversight (26(2)).
• The input data you control is relevant and representative of the intended use (26(4)).
• Operation is monitored; in case of a risk or a serious incident, use is suspended and the provider informed (26(5)).
• The logs generated are kept for at least six months (26(6)).
• Affected workers and their representatives are informed before the system is put into service or used at the workplace (26(7)).
• The people concerned are informed when the agent contributes to a decision about them (26(11)).
• A fundamental rights impact assessment is carried out before use if you are a public body or provide a public service, or if the agent assesses creditworthiness or prices life or health insurance (Article 27). It is notified to the market surveillance authority (Article 27(3)). It can reuse parts of the GDPR impact assessment, without being replaced by it.

Provider of a high-risk system

The obligations are much heavier: if an agent turns you into a provider, get support. The kit’s list, below, is not exhaustive.

• Risk management over the whole life cycle (Article 9).
• Governance of training, validation and testing data (Article 10).
• Technical documentation (Article 11 and Annex IV).
• Automatic recording of events (Article 12).
• Instructions for use for deployers (Article 13).
• Human oversight built into the system (Article 14).
• Accuracy, robustness and cybersecurity (Article 15).
• Quality management system (Article 17).
• Conformity assessment, EU declaration and CE marking (Articles 43, 47 and 48).
• Registration in the EU database (Article 49).
• Post-market monitoring and reporting of serious incidents (Articles 72 and 73).

The regulation sets out others, notably in Articles 16 and 18 to 22.

The agent trap

An agent built to summarise documents, reconfigured to sort job applications, changes its intended purpose. It becomes high risk, and you become its provider.

The register

The register is the first document of an audit: one row per agent and per use. It proves that the company knows which agents it uses and what they do, which is the first thing an auditor or a large customer asks for. The Excel file arranges its columns in six groups:

• Identity: name, use, department, owner, status;
• Role and technology: your role, platform, platform provider, models;
• People and data: audience, Annex III area, decision or profiling, interaction with people, published content, sensitive data;
• Classification: suggested level, calculated; retained level and its justification;
• Control: human approval, logs, awareness, impact assessments;
• Follow-up: last review, next review six months later, in red once overdue, and status.

The grey columns calculate themselves. The suggested level applies questions 2 to 4 of the grid and stays empty until the columns it depends on are filled in; the next review falls six months after the last one. The register does not detect prohibited practices: check Article 5 for each agent. Three fictitious examples show how to fill it in.

Frequently asked questions