Three EU AI Act milestones are fixed: entry into force on 1 August 2024, prohibitions from 2 February 2025, and general-purpose model obligations from 2 August 2025. AI literacy applies in the first phase too. High-risk dates have been subject to proposed changes, so AI agent deployers should build those controls now rather than wait.
What is the timeline, in plain words?
The EU AI Act (Regulation (EU) 2024/1689) phases in its obligations over several years.
| Milestone | What it brings | Status for deployers |
|---|---|---|
| 1 August 2024 | Entry into force | The Act is law; obligations follow on their own dates |
| 2 February 2025 | Prohibited practices apply; general provisions including AI literacy | Applies now |
| 2 August 2025 | General-purpose AI model obligations | Applies now, to model providers |
| Later phases | Transparency duties to people; high-risk system obligations for providers and deployers | Phasing in; high-risk dates have been subject to proposed changes |
We do not state dates for the later phases here. Proposals to adjust the application of high-risk obligations have been made, and any date we printed could be wrong by the time you read it. Check the current position with counsel before relying on one.
Which obligations already apply to a deployer?
Two sets of duties already bind every business that deploys AI agents in the EU or whose agents' output is used there:
- The prohibitions. No agent may manipulate people to their significant harm, exploit vulnerability, carry out social scoring or engage in the other banned practices. Breach falls in the highest penalty tier, explained in EU AI Act penalties and what they mean for agent deployers.
- AI literacy. Staff who use, configure, approve or oversee agents must have a sufficient level of understanding.
General-purpose model obligations also apply, but to model providers. A deployer benefits from the documentation they must produce but does not carry those duties.
What is still phasing in?
Transparency to people. Telling people they are interacting with an AI system and marking certain generated content. These are deployer and provider duties that will apply to most customer-facing agents.
High-risk obligations. For deployers, the Article 26 set: use as instructed, competent oversight, input data control, monitoring, log retention, notices to workers and affected people, cooperation with authorities. For certain deployers, the Article 27 fundamental rights impact assessment. For providers, the full design and conformity set.
The list of what deployers must do, once all of it applies, is in what the EU AI Act requires deployers of AI agents to do. Whether you are a deployer or have become a provider affects which list you follow; see what a deployer is and what a provider is.
Why should agent deployers not wait for the dates?
Three reasons specific to agents.
The controls take longer to build than for models. Oversight for an agent means hold points before consequential actions, defined per capability. Logging means an action record, not a chat transcript. Both require changes to how agents are wired into systems, which takes months, not weeks.
Other rules already require most of it. Financial supervisors expect model risk management, conduct controls and operational resilience for agents today. GDPR already governs automated decisions and personal data in logs. DORA already applies to ICT risk at financial entities. The AI Act's high-risk duties largely overlap with these.
Agents change faster than the timeline. A minimal-risk agent today can gain a capability next quarter that puts it in a listed high-risk use. A deployer that has built the controls can absorb that change. One that waited cannot.
What should a deployer prepare now?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Prohibitions (applies now) | Review of every customer-facing capability; floor on manipulation and exploitation | Review records; refusal records |
| AI literacy (applies now) | Role-based training for briefers, approvers and overseers | Training and assessment records |
| Transparency (phasing in) | Disclosure wording drafted and tested in every channel | Wording and channel captures |
| Classification (needed before high-risk duties) | Register of agents and capabilities with tier and role | Signed register |
| Oversight (phasing in for high-risk) | Hold points before adverse or irreversible actions | Approval records |
| Logs (phasing in for high-risk) | Append-only action record with retention schedule | Retained records |
| Monitoring (phasing in for high-risk) | Monthly review of refusals, holds and overrides | Monitoring reports |
| Rights impact assessment (where it will apply) | Draft assessment for credit scoring and insurance pricing capabilities | Draft FRIA |
Firms operating in the US face a different set of timelines, including state laws such as Colorado's act aimed at high-risk systems; Colorado AI act duties for deployers of AI agents covers that one in substance.
What will a supervisor or auditor ask?
- What have you done about the prohibitions and AI literacy, which already apply?
- How do you know none of your agents uses a prohibited practice?
- Which agents will be high-risk, and what is your plan?
- Who is tracking changes to the application dates?
What does the timeline not change?
It does not delay GDPR, consumer law, conduct rules or DORA, which apply to agents now. It does not delay the prohibitions or literacy. And it does not change what supervisors in financial services already expect of AI used in regulated activity. For the full map, see AI agent regulation, mapped to controls.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Questions people ask
If high-risk dates move, can we stop work on high-risk controls?
We would not. Most of the controls are already expected under sector rules and GDPR, and agents can become high-risk with one new capability. A later date gives you time to do it properly, not a reason to stop.
Does the transparency duty apply to internal agents?
It applies where an AI system interacts with natural persons, which can include staff. Where it is obvious from context that the person is dealing with AI, the duty may not bite. Many firms disclose anyway.
Who should track changes to the Act's dates?
Your legal or compliance function, with a named owner. Record each change and its effect on your plan, so the decision trail is clear if a supervisor asks.