What the AI Act delay changes for companies deploying agents

This note is our reading, as engineers who build agent systems, of published summaries of the 2026 amendments to the EU AI Act. It is not legal advice. Whether a given system is high-risk, and which role your company holds under the Act, are questions for your counsel. Dates below are as of 5 November 2026.

What happened

The EU reached a provisional political agreement on the Digital Omnibus on AI on 6 May 2026. The final text, Regulation (EU) 2026/1744, was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026. It amends the AI Act (Regulation (EU) 2024/1689) mainly by moving the dates on which the high-risk requirements apply.

What moved

Obligation	Original date	Date after the omnibus
High-risk systems in Annex III areas (for example employment, education, credit scoring, law enforcement)	2 August 2026	2 December 2027
High-risk AI embedded in regulated products under Annex I (for example medical devices, machinery)	2 August 2027	2 August 2028
Machine-readable marking of AI-generated content, for generative systems already on the market before 2 August 2026	2 August 2026	2 December 2026

The omnibus also changes the AI literacy duty in Article 4. According to Gibson Dunn's summary of the agreement, providers and deployers now have to support the development of AI literacy among their staff, rather than ensure a particular level of it.

What did not move

Obligations for general-purpose AI models have applied since 2 August 2025, and the omnibus did not change that date. Those duties fall mainly on model providers. For a company deploying agents on top of such models, they matter in two ways: the documentation you should be able to get from your model supplier, and the possibility that modifying a model substantially changes your own role under the Act. The second point depends on facts and is worth raising with counsel early.

The transparency obligations in Article 50 apply from 2 August 2026, as originally planned. Among other things, systems that interact with people have to make it clear they are AI unless that is obvious, and deepfakes have to be disclosed. Only the machine-readable marking for generative systems already on the market got the short grace period listed above.

The prohibited practices in Article 5 have applied since 2 February 2025. The omnibus adds a new prohibition on systems that generate or manipulate non-consensual intimate imagery or child sexual abuse material, applying from 2 December 2026.

What this means for agent deployments

Most agent use cases companies discuss today (code changes, document processing, reconciliations, internal reporting) is unlikely to fall under Annex III. Some of it plainly can. An agent that screens CVs or ranks candidates touches employment. An agent that feeds a creditworthiness decision touches credit scoring. For those, the high-risk requirements now apply from 2 December 2027, about thirteen months after this note was published.

Article 50 is already live, and agents are exactly the kind of system it is written for. An agent that emails suppliers, answers customers or negotiates appointments interacts with people outside the company, and those people have a right to know they are dealing with AI. Article 50(1) is a provider duty: the provider has to design the system so people know they are dealing with AI. If you build the agent yourself, that provider may be you.

For high-risk systems, the AI Act's deployer obligations in Article 26 remain the reference point. They include using the system according to the provider's instructions, assigning human oversight to people with the competence and authority to exercise it, keeping the automatically generated logs for at least six months, and informing workers before a high-risk system is put into use at the workplace. Providers carry the design-side duties on record-keeping (Article 12) and human oversight (Article 14).

Why the delay is not a reason to wait

The omnibus changed when the high-risk requirements apply. In the summaries we have read, it did not remove what they require: risk management, logging, human oversight, documentation. Thirteen months is not long for that work in a large company.

The slow part is rarely writing policy. It is finding out where agents already run, what data and tools they touch, and who is affected by their output. Then comes classifying each use, and only after that the engineering: traces that capture every tool call, retention that matches the rules, oversight points with named people who can stop a task. Adding logging to a system that is already in production takes far longer than building it in from the start.

There is also an engineering reason that has nothing to do with Brussels. The same traces and approval records that a regulator may ask about are what you need to debug a failed task or run an incident review. Teams that skip them to save time usually rebuild them after the first serious incident.

Finally, the questions are already arriving from other directions. Procurement and internal audit ask how agents are supervised and what they log, whatever the regulatory calendar says.

Our defaults for every deployment

We treat these as default engineering requirements, independent of how a system is eventually classified:

• An inventory entry for each agent: what it does, what data and tools it can reach, who is affected by its output, and a provisional classification for counsel to confirm.
• Traces of every agent action stored inside the client's infrastructure, with retention agreed with the client's legal team. For anything that could be high-risk, six months is the floor we design for.
• Human sign-off at defined points, with named roles, recorded in the same trace as the agent work.
• Disclosure wherever an agent communicates with people outside the company.
• A record of the model versions in use and the documentation obtained from each model supplier.

None of this guarantees compliance, and we do not claim it does. It means that when counsel classifies a system, the evidence they need already exists, and the remaining work is a decision rather than a rebuild.