AI agents and the AI Act after the Digital Omnibus: roles, duties, deadlines

Organisations introducing an AI agent soon ask which AI Act obligations apply, and from when. Last reviewed: October 2026. Our position: the AI Act regulates purposes, not agents. Most agents in law firms, tax practices, clinics and small businesses are not high-risk AI systems. Some duties already apply, and the Digital Omnibus delay is no reason to wait.

Provider or deployer: which role is yours?

An LLM-based agent fits the AI system definition in Article 3(1): it operates with some autonomy and infers from input how to generate outputs that can influence physical or virtual environments. The model underneath is usually a GPAI model, whose own obligations fall on the model provider.

What counts is your role. A provider develops an AI system, or has one developed, and places it on the market or puts it into service under its own name; putting into service includes supply for own use. A deployer uses an AI system under its authority in a professional capacity.

When an agent becomes a high-risk system

Autonomy does not make an agent high-risk. Its purpose does: either it falls under Annex III (Article 6(2)) or it is a safety component of a product covered by Annex I. For small businesses and practices, two areas matter most:

• Employment (point 4): screening or filtering applications, evaluating candidates, monitoring and evaluating the performance and behaviour of staff.
• Essential services (point 5): assessing the creditworthiness of individuals, risk assessment and pricing for life and health insurance, triage of patients in emergency care.

An agent that pre-codes receipts, records deadlines or researches case law generally falls under no Annex III area; point 8 (justice) concerns systems used by or for judicial authorities or in alternative dispute resolution. Under Article 6(3), even an Annex III system is not high-risk if, for instance, it only performs a narrow procedural or preparatory task and does not materially influence the outcome of decisions. Profiling of individuals is always high-risk, and a provider relying on the exception must document its assessment beforehand. Draft Commission guidelines on classification followed in May 2026.

What already applies

• AI literacy (Article 4), since 2 February 2025: providers and deployers must take measures to support the AI literacy of their staff, in line with their knowledge and the context of use. Since the Omnibus, no specific level has to be guaranteed, but the duty covers every agent. In practice: training on what the agent may do, where it errs and how to stop it.
• Prohibited practices (Article 5), since 2 February 2025: this includes inferring people's emotions in the workplace, except for medical or safety reasons, so an agent rating staff mood from call audio would be unlawful.
• Transparency (Article 50), since 2 August 2026: an agent that talks directly to people, such as a clinic's online appointment assistant, must be designed to tell them they are dealing with AI, unless that is obvious. Providers of generative systems must also mark outputs as AI-generated in machine-readable form.

What the Digital Omnibus postponed

The Digital Omnibus on AI, Regulation (EU) 2026/1744, entered into force on 27 July 2026. It postponed the high-risk requirements, including deployer obligations under Article 26; Articles 4, 5 and 50 largely stayed on schedule.

From 2 December 2027, deployers of Annex III systems must, among other things, assign oversight to competent people, keep automatically generated logs for at least six months and inform workers before workplace use.

What to do now

1. Inventory by purpose: record each agent's purpose and check it against Annex III. See our primer What is AI agent governance?
2. Record role and classification: note for each agent whether you are provider, deployer or both, and write down why it is not high-risk.
3. Evidence AI literacy: document your measures per staff group.
4. Check transparency: confirm the AI notice on agents that talk to clients, patients or customers.
5. Prepare oversight and logging: if you plan an agent in HR, build human approval and logging into the architecture now, not in 2027. Under Austria's NISG 2026, see also NIS2 and AI agents.

Obligations need a technical foundation

AEGIS, the locally operated agent system from oQueo, does not make you compliant. It does provide building blocks for it: limited and documented permissions through an allowlist with default deny, a check of every action by a separate authority before execution, a maximum action level per agent as an emergency brake, and an append-only audit trail as a SHA-256 hash chain. Classification and responsibility stay with you. More on our security page.

This article does not replace legal advice.