The EU AI Act is already law in Ireland, and since 2 August 2026 most of its rules can be enforced. For security and compliance teams, the practical question is one of evidence. Can you show which AI agents exist in your Microsoft 365 tenant, who built them, what they connect to and who can use them? We built Argus to answer that question every day, and the rest of this article explains why the Act makes it worth answering.
TL;DR
- The Act applies in Ireland to providers and deployers of AI systems, and AI agents are covered through the standard definitions of an AI system and a general-purpose AI model. They are not a separate category.
- Prohibitions and AI literacy have applied since 2 February 2025. Transparency duties and the penalty regime apply from 2 August 2026. High-risk duties are deferred to 2 December 2027 (Annex III) or 2 August 2028 (Annex I).
- Ireland enforces through the AI Office of Ireland plus sectoral regulators such as the Data Protection Commission and the Central Bank.
- Fines reach €35 million or 7% of worldwide turnover for prohibited practices, and €15 million or 3% for most other breaches.
- You cannot classify, oversee or log an agent you do not know exists. An inventory with owners and permissions comes first.
Table of Contents
- What the Act covers and who it reaches in Ireland
- How AI agents fit into the risk tiers
- Key dates for 2026 to 2028
- Who enforces the Act in Ireland
- Penalties under Article 99
- Legal duties and good practice are not the same thing
- A practical sequence for security and compliance teams
- How Argus supports this work
- FAQ
- Sources
What the Act covers and who it reaches in Ireland
Regulation (EU) 2024/1689 is directly applicable across the EU, so Ireland does not need to copy it into national law for it to bind you. According to the Commission's guidance, it applies to public and private actors, inside and outside the EU, that place an AI system or general-purpose AI model on the EU market, or put a system into service or use it in the EU. Military, defence and national security systems and pre-market research are exempt.
The Department of Enterprise's SME test adds the roles involved: providers, deployers established or located in the Union, importers, distributors, product manufacturers, authorised representatives, and third-country providers or deployers whose system output is used in the Union.
Most Irish organizations using Microsoft 365 are deployers: they use AI systems built by someone else, or by their own staff, under their own authority. An employee who builds an agent in Copilot Studio and shares it across a department is, in practice, putting a system into service inside your organization. That is why ownership records matter.
Ireland has no separate national definition of an AI system. The Regulation of Artificial Intelligence Act 2026 relies on the EU Act's definition, so there is one rulebook to read.

How AI agents fit into the risk tiers
The Act sorts systems into four tiers, as the Department of Enterprise summary describes them:
| Tier | Examples | What it means |
|---|---|---|
| Unacceptable | Social scoring, workplace emotion recognition, manipulation | Banned |
| High risk | Uses listed in Annex III, such as employment or education decisions | Risk management, logging, human oversight, robustness, cybersecurity |
| Limited risk | Chatbots | Transparency duties |
| Minimal risk | Spam filters, games | Free use |
The EU's AI Act Service Desk confirms that AI agents are not a separate category. Three rules matter most:
- Article 5 prohibitions apply to agents like any other system.
- Article 50 transparency applies from 2 August 2026 to agents that interact with people or generate content.
- High-risk requirements apply if the agent's purpose falls in a high-risk category.
Classification turns on intended purpose, according to the Commission, and must be repeated if the system is substantially modified. That detail is easy to miss with agents. An HR assistant that only answers leave questions is a different thing from one that is later connected to CV files and used to rank applicants. The second use may fall within Annex III's employment category. Under Article 6, profiling of natural persons is always high-risk.
An agent's scope can change with one new connector or one changed sharing setting. Classification is therefore not a one-time exercise. It needs a record of what each agent is connected to, kept current.
Key dates for 2026 to 2028
The AI Act Service Desk timeline sets the sequence:
- 2 February 2025: prohibitions and AI literacy obligations apply.
- 2 August 2025: general-purpose AI rules and governance provisions apply.
- 2 August 2026: prohibited practices, transparency rules and general-purpose AI model rules become enforceable, and the Article 99 penalty regime starts for obligations then in application.
- 2 December 2026: providers of systems placed on the market before 2 August 2026 have until then to meet the Article 50(2) marking and detection duty.
- 2 December 2027: Annex III high-risk obligations, following the Digital Omnibus, which the Commission says moved these dates.
- 2 August 2028: Annex I high-risk obligations, for AI embedded in regulated products.
The deferral gives high-risk preparation more time. It does not delay prohibitions, AI literacy or transparency, which are already live.
Who enforces the Act in Ireland
Ireland chose a distributed model. The Regulation of Artificial Intelligence Act 2026 was signed on 21 July 2026 and establishes the AI Office of Ireland as an independent statutory body, operational since 2 August 2026, according to the Agent Liability EU tracker. It coordinates 15 designated national competent authorities and acts as the single point of contact.
Enforcement sits mainly with sectoral regulators. TechCentral.ie reported that these include the Data Protection Commission, the Central Bank of Ireland, Coimisiún na Meán and the Workplace Relations Commission. The tracker notes the DPC covers AI systems involving personal data, the Central Bank covers financial services, and the Health and Safety Authority, HSE and HPRA cover health.

What this means for you depends on your sector. A bank answers to the Central Bank. An agent that touches employee or customer personal data brings the DPC into view. Matheson's analysis says the Government intended the Bill as a technical implementing measure that adds no obligations beyond the Act itself.
The Commission's AI Office, separately, holds enforcement powers over general-purpose AI models. That mostly concerns the model providers, not organizations that use their products.
Penalties under Article 99
The Department of Enterprise sets out the maximum fines:
| Breach | Maximum fine |
|---|---|
| Prohibited practices, or non-compliance with data requirements | €35 million or 7% of worldwide annual turnover |
| Other requirements | €15 million or 3% |
| Incorrect, incomplete or misleading information to authorities | €7.5 million or 1.5% |
For SMEs the lower of the two amounts applies. Note the third row. Even a well-run organization can be exposed if it answers a regulator's request with an incomplete picture of its agents. Records that are complete and dated protect you there.
Legal duties and good practice are not the same thing
The Department of Enterprise's guidance for organisations recommends several steps: map all AI tools, determine their risk category, check for prohibited practices, assign oversight responsibility, keep documentation, monitor, disclose where required, train staff and verify supplier compliance. These are recommended steps. The binding duties are the Act's own.
For deployers of high-risk systems, the consolidated text requires:
- Technical and organizational measures to use systems per their instructions.
- Human oversight assigned to competent, trained and authorized people.
- Monitoring of operation, suspension of use, and notification of the provider and authority where risks arise.
- Immediate reporting of serious incidents.
The Commission lists the wider high-risk obligations: risk assessment and mitigation, activity logging for traceability, documentation for authorities, human oversight, and a high level of robustness, cybersecurity and accuracy.
A recent peer-reviewed study maps the Act's cybersecurity provisions onto technical controls for autonomous agents. It organizes them around threats such as prompt injection, tool poisoning, excessive agency and supply-chain compromise, with verification evidence for each control. The theme is consistent: a control counts when you can show it works.
For guidance on how people keep judgment over AI-assisted outcomes, see this piece on AI-assisted decision making. It bears directly on the human oversight duty.
A practical sequence for security and compliance teams
Every step below depends on knowing what is running. This is the order we would work in.
- Inventory every agent. Record what exists in the tenant, not what was approved. Agents built by individual staff are the ones that go unlisted. Argus audits all AI agents in Microsoft 365 daily, so the inventory does not go stale.
- Attach an owner to each agent. Oversight must be assigned to a named, competent person. Argus links each agent to its creator and current owner, which gives compliance officers the accountability record the Act's oversight language assumes.
- Map connections and permissions. Note who can use each agent, what data sources it reaches and whether sharing is unrestricted. Argus records connection points and user permissions, and ranks risks such as unrestricted sharing and sensitive connections.
- Classify by purpose. Decide whether each agent is minimal, limited or potentially high-risk, and check against prohibited practices. Re-run this when an agent's connections or use change. Argus alerts in real time on risky changes, which is the trigger for reclassification.
- Fix and prove the fix. Logging and traceability are central to the high-risk regime. Argus verifies that remediation was actually implemented and keeps the evidence, and it supports integration with tools like Jira or ServiceNow so fixes are tracked as tickets.
- Train your people. AI literacy has applied since February 2025. Knowing which agents each team has built shows where training matters most.
For technology providers working out whether a system is high-risk and documenting oversight and data controls, insurance workflow automation with human review shows one sector's approach to keeping a person in the loop.
How Argus supports this work
Argus is our security platform for Microsoft 365. It audits AI agents daily and records their creators, connection points and permissions. It ranks risks, alerts on changes and verifies that fixes were carried out. It does this without making changes to your tenant and without requiring additional Microsoft licenses.

If your task for the coming quarter is an agent inventory with owners and evidence, see how Argus works. You can also review what Argus checks in Copilot Studio agents or read customer stories.
FAQ
Does the EU AI Act apply to Irish companies that only use AI tools?
Yes. Deployers established or located in the Union are within scope, per the Department of Enterprise. Duties are lighter than for providers, but they include transparency, AI literacy and, for high-risk uses, human oversight and monitoring.
Are AI agents high-risk under the Act?
Not automatically. Agents are classified by intended purpose like any AI system. An agent used in areas listed in Annex III, such as employment, can be high-risk, and profiling of natural persons always is. Those duties apply from 2 December 2027 or 2 August 2028.
Who regulates AI in Ireland?
The AI Office of Ireland coordinates 15 designated authorities. Sectoral regulators enforce in their fields: the Data Protection Commission for personal data, the Central Bank for financial services, and others for media, workplace and health matters.
What are the maximum fines?
Up to €35 million or 7% of worldwide annual turnover for prohibited practices, €15 million or 3% for other requirements, and €7.5 million or 1.5% for misleading information to authorities. SMEs face the lower of the two amounts.
Sources
- Navigating the AI Act, European Commission
- Artificial Intelligence Act SME Test, Department of Enterprise
- Legal 500 AI Guide Ireland Chapter 2026
- How are AI agents addressed within the AI Act?
- Article 6: Classification rules for high-risk AI systems
- Timeline for the Implementation of the EU AI Act
- AI Act Service Desk FAQ
- AI Act, European Commission
- EU AI Act implementation status 2026
- Ireland's AI Office is five weeks away, TechCentral.ie
- Irish Government publishes the Regulation of Artificial Intelligence Bill 2026, Matheson
- EU Artificial Intelligence (AI) Act, gov.ie
- The EU AI Act and my organisation, DETE
- Consolidated text of Regulation 2024/1689, EUR-Lex
- From obligation to enforcement: mapping EU AI Act and CRA, Springer
