The EU AI Act after the Omnibus: the new timeline for agent deployments

RegulationEngagement lead, ZürichUpdated 9 November 20268 min read

Most AI Act commentary is written for model providers. If you are a bank, an insurer or a utility putting agents into production, you are a deployer, and the obligations that matter to you are different, smaller, and closer than you think. They also moved this summer: the Digital Omnibus on AI, Regulation (EU) 2026/1744, was adopted on 8 July 2026 and entered into force on 27 July 2026.

The dates that matter to a deployer, as amended

DateWhat appliesWhat it means for an agent in production
2 February 2025Prohibited practices; AI literacy (Art. 4)Staff who operate or oversee agents need suitable training. The Omnibus softened Art. 4 so it no longer requires a guaranteed literacy level, but documented training is still the evidence you will want.
2 August 2025General-purpose AI model obligations; governanceMostly on providers. Ask your model vendors whether they signed the GPAI Code of Practice and keep their documentation on file.
2 August 2026General application and enforcement; transparency duties (Art. 50)If your agent talks to customers, it must say it is an AI. Providers of generative systems already on the market have until 2 December 2026 to mark AI-generated output.
2 December 2026New prohibitions on non-consensual intimate deepfakes and child sexual abuse materialNot an enterprise use case, but content filters on customer-facing generation must cover them.
2 December 2027High-risk obligations for Annex III use cases (was 2 August 2026)Credit scoring, life and health insurance pricing, HR decisions, safety components of critical infrastructure. Deployer duties under Art. 26 apply.
2 August 2028High-risk AI embedded in Annex I products (was 2 August 2027)Agents inside machinery, medical devices or vehicles. AI safety requirements for machinery move into the Machinery Regulation.

The Omnibus moved the dates; it did not change which systems are high-risk. Use the extra sixteen months to build evidence, not to wait. High-risk systems already on the market are caught only if their design changes significantly after the deadline, and systems used by public authorities have until 2 August 2030.

Who supervises you

In Germany, the AI Act implementation law (KI-MIG) of July 2026 makes the Bundesnetzagentur the default market surveillance authority and BaFin the authority for AI linked to regulated financial activity. BaFin has said it starts now on transparency, prohibited practices and AI literacy, and takes on high-risk systems such as creditworthiness assessment from December 2027. Switzerland has no horizontal AI law: the Federal Council plans a consultation on sector-specific rules in spring 2027, and the Federal Data Protection Act already applies to AI processing, including disclosure when people are talking to a machine.

Which of your agents are high-risk

Annex III is the list to read. In regulated sectors the hits are predictable: creditworthiness and credit scoring, risk assessment and pricing in life and health insurance, recruitment and promotion decisions, access to essential services, and AI used as a safety component in critical infrastructure such as electricity, gas, water and digital networks. An agent that supports a human decision in these areas is still in scope if it materially influences the outcome.

An agent that resolves billing disputes is not high-risk. An agent that decides whether to grant a payment plan to a customer in arrears may be. The difference is not the technology, it is the decision.

Deployer duties under Article 26, translated

  • Use it as instructed. Keep the provider's instructions for use and show that your configuration stays within them. For a platform like ours, that is the evaluation suite plus the policy configuration, versioned.
  • Human oversight by people with authority. The person overseeing the agent must be able to stop it, override it and understand its limits. This is why we design the loop before the agent.
  • Input data relevance. If you control the input data, you must make sure it is relevant and representative. Retrieval sources and memory stores count.
  • Logging. Keep the logs the system generates for at least six months. The gateway does this by default, inside your tenant.
  • Inform workers. If the agent affects employees, their representatives must be informed before deployment.
  • Transparency to affected persons. People subject to a high-risk decision must be told an AI system was involved.

What we build into every agent

  1. A one-page risk classification per use case, signed by the process owner and your DPO.
  2. An oversight design: who approves, who reviews, who observes, with thresholds written down.
  3. Evaluation results attached to every version, so "used as instructed" is provable.
  4. Gateway logging and PII redaction on by default, retention set to your policy.
  5. A disclosure line in every customer-facing channel, in the customer's language.
The Act does not ask you to slow down. It asks you to write down what you were going to do anyway, and to be able to stop the machine.

If you want the German version of this table for your Betriebsrat or your Datenschutzbeauftragte, ask us and we will send it.

Sources

RegulationEUDACHEnterprise AI
Engagement lead, Zürichunus · Applied AI for the European enterprise

Talk to the people who wrote this.