Regulatory · AI Regulation

The EU AI Act and AI Agents: What Teams Need to Know in 2026

What the EU AI Act means for teams running AI agents after the 2026 Digital Omnibus: new dates, human oversight, logs, transparency and a starter checklist.

Author
Quaspar Research
Published
November 18, 2026
Last reviewed
Last reviewed November 18, 2026
Reading time
11 min read
Article contents

Regulatory information can change. This resource is educational and should not be treated as legal, clinical, or regulatory advice. Consult the applicable primary guidance and qualified professionals for decisions affecting your organization. Last reviewed November 18, 2026.

Where the AI Act stands in November 20261,2

The AI Act entered into force on 1 August 2024 and applies in stages. Bans on prohibited practices have applied since February 2025, and obligations for providers of general-purpose AI models since August 2025.

In July 2026 the EU changed the timetable. The Digital Omnibus on AI, Regulation (EU) 2026/1744, entered into force on 27 July 2026, a few days before the high-risk rules were due to start. It moved the rules for high-risk uses listed in Annex III, such as hiring, credit and access to essential services, to 2 December 2027, and the rules for AI built into regulated products to 2 August 2028.

What did not move: the duty to tell people they are interacting with an AI system, and the bans on prohibited practices. The core high-risk obligations were also not rewritten. The EU bought time for standards and support tools; it did not lower the bar.

Key dates at a glance1,2,8

Dates as published by the European Commission after the Digital Omnibus. Check the official text before planning around them.

DateWhat applies
1 August 2024The AI Act enters into force
2 February 2025Prohibited AI practices are banned
2 August 2025Obligations for providers of general-purpose AI models
27 July 2026Digital Omnibus enters into force and moves the high-risk dates
2 August 2026Transparency duties, including telling people they are talking to an AI
2 December 2027Rules for high-risk uses listed in Annex III
2 August 2028Rules for high-risk AI built into regulated products

Agents are not a category. Uses are.3,4

Whether an agent is high-risk depends on its intended purpose, not on how clever or autonomous it is. An agent that drafts marketing emails is unlikely to be high-risk. An agent that filters job applications, allocates tasks based on people's behaviour, evaluates creditworthiness, or prices life and health insurance falls into areas listed in Annex III.

The same model, the same framework and the same code can therefore carry very different obligations depending on where the agent is pointed. That is why the first step is an inventory of what each agent is actually allowed to do, not a review of which model it uses.

Roles matter too. A team that builds an agent and puts it into service under its own name can be the provider of an AI system. A team that uses an AI system under its authority is a deployer. Many teams are both.

Four questions to ask about every agent

Is anything it does prohibited?

Rare for business agents, but check the Article 5 list, for example manipulative techniques or social scoring, before anything else.

Does it touch a high-risk use?

Hiring, managing workers, credit, insurance pricing, essential services, education and the other Annex III areas. If yes, the December 2027 rules apply.

Does it talk to people or create content?

Customer-facing agents and agents that generate text, images or audio carry transparency duties that already apply.

Are you the provider, the deployer, or both?

Providers design in the controls. Deployers must use them, assign people to oversee, monitor and keep logs.

What human oversight means in Article 145

Article 14 says high-risk AI systems must be designed so that people can effectively oversee them while they are in use. The measures must match the risks, the context of use and, explicitly, the system's level of autonomy. Autonomy is exactly what agents add.

The article lists what the people assigned to oversight must be able to do: understand the system's capabilities and limits and monitor it; stay aware of the tendency to over-rely on its output, which the Act calls automation bias; interpret its output correctly; decide not to use it, or override or reverse what it produced; and interrupt it with a stop button or similar procedure that brings it to a halt in a safe state.

Article 14, translated for agents5

Our reading of what each oversight ability looks like when the AI system takes actions instead of only answering.

Article 14(4) asks that people can…For an agent, that means…
Understand its limits and monitor its operationA live view of each agent's permissions, current runs and recent actions
Stay aware of automation biasApproval queues short enough to read, and tracking of how often reviewers simply approve
Correctly interpret its outputReviewers see the exact action, the sources behind it and the rule that made it stop
Override or reverse its outputApprove, edit or reject before the action, and an undo path after it
Interrupt it and bring it to a safe haltCancel any run and pause any agent instantly, without asking the agent

Deployers have their own duties6,7

Article 26 applies to organizations that use high-risk AI systems. They must assign human oversight to people with the necessary competence, training and authority. They must monitor the system's operation and suspend its use if they believe it presents a risk. They must keep the logs the system generates automatically, to the extent those logs are under their control, for at least six months unless other law says otherwise. And they must inform workers' representatives and affected workers before using such a system at work.

Article 12 sets the other side: high-risk systems must technically allow automatic recording of events over their lifetime, so that risky situations can be identified and the system's operation can be monitored. In practice, if you cannot show who an agent acted for, what it did and who approved it, you cannot meet either article.

Transparency applies to far more agents8

Most business agents will never be high-risk. Many will still meet Article 50. Providers must design AI systems that interact directly with people so that those people are told they are dealing with an AI, unless it is obvious from the context. AI-generated content must be marked as such, with specific rules for deepfakes and published text.

For a team running a support agent that answers customers by email or chat, this is usually the first EU rule that applies, and it applies now. Some content-marking duties have a short grace period.

What is at stake9

Member States set the penalties, within limits the Act defines. Using a prohibited practice can bring fines of up to EUR 35 million or 7% of worldwide annual turnover, whichever is higher. Breaching most other obligations, including the deployer and transparency duties, can bring fines of up to EUR 15 million or 3%. The Act asks authorities to take the situation of smaller companies into account.

Why a line in the prompt is not a control

The Act describes oversight measures that are built into the system or put in place by the deployer. An instruction in a system prompt is neither in any reliable sense. The model can misread it, and a malicious instruction hidden in an email or web page can tell it to ignore it.

Controls you can point to live outside the model: permissions the agent cannot change, approval steps it cannot skip, a stop that works whatever the agent is doing, and a record written by the platform rather than by the agent. That is the layer regulators keep describing, whether or not your agent is ever classed as high-risk.

A checklist to start before December 2027

None of this depends on the final standards. All of it is useful even if no rule ever applies to you.

  • List every agent, its owner, its purpose and the tools it can call
  • Classify each agent's use against Annex III, and record why
  • Add a clear AI notice wherever an agent talks to people
  • Set each agent's permissions outside the model, tool by tool
  • Require a person's approval for consequential or irreversible actions
  • Name who oversees each agent, and make sure they have the authority to overrule it
  • Make sure anyone authorized can stop a run or pause an agent at once
  • Record every run, tool call, approval and rejection with who and when
  • Keep logs for at least six months for anything that may be high-risk
  • Review approval rates regularly to catch reviewers who have stopped reading

Where Quaspar fits

Quaspar is the governance layer for AI agents. Teams set what each agent may do, people approve the actions that matter, and everyone can see what happened. Each agent's permissions are set when it is created and enforced by Quaspar, never by the model, and risky capabilities default to requiring approval. Outgoing messages, submissions and changes to connected systems wait in Workspace for a person. An administrator can cancel any run and pause any agent instantly. Every agent run, tool call, approval and rejection goes into an activity record with who, what, when and the outcome.

Quaspar does not classify your systems, make an AI system compliant, or replace legal advice. It gives your team the controls and the record that the AI Act keeps asking for. Quaspar is in private preview and is not SOC 2 certified.

Next step

Quaspar is the governance layer for AI agents and is in private preview. Request access to set agent permissions, route important actions to people and keep a record of what every agent did.

Request Studio access

Sources

Authoritative references

Rules, standards, and guidance can change. The links below were reviewed on November 18, 2026.

  1. 1European Commission, Shaping Europe's digital future. AI Omnibus enters into force (27 July 2026)
  2. 2European Commission, Shaping Europe's digital future. EU agrees to simplify AI rules (political agreement, 7 May 2026)
  3. 3European Commission, AI Act Service Desk. AI Act, Article 6: Classification rules for high-risk AI systems
  4. 4European Commission, AI Act Service Desk. AI Act, Annex III: High-risk AI systems
  5. 5European Commission, AI Act Service Desk. AI Act, Article 14: Human oversight
  6. 6European Commission, AI Act Service Desk. AI Act, Article 26: Obligations of deployers of high-risk AI systems
  7. 7European Commission, AI Act Service Desk. AI Act, Article 12: Record-keeping
  8. 8European Commission, AI Act Service Desk. AI Act, Article 50: Transparency obligations
  9. 9European Commission, AI Act Service Desk. AI Act, Article 99: Penalties