The AI Office of Ireland becomes operational on 1 August 2026. Here is what it can do, what it cannot do yet, what its first 90 days will probably look like, and what Irish organisations should have in place before then.
In nine weeks, Ireland gets an AI regulator. On 1 August 2026, the AI Office of Ireland becomes operational under the Regulation of Artificial Intelligence Bill 2026, currently moving through the Oireachtas. From that date, an Irish public body holds statutory enforcement powers over AI systems deployed in the State.
Most Irish organisations have not yet absorbed what that means. The conversation in boardrooms over the past twelve months has been about the EU AI Act in the abstract. The conversation from 1 August will be about an actual regulator with actual powers acting on actual organisations. The shift in posture is significant, and the window to prepare is now closing.
What the AI Office is, in practice
The AI Office of Ireland is not a single agency in isolation. The General Scheme of the Bill establishes a coordinated enforcement architecture: a central AI Office responsible for coordination, plus fifteen named competent authorities — sector regulators including the Central Bank, the Data Protection Commission, the Health Products Regulatory Authority, the Financial Services and Pensions Ombudsman, and others — each retaining enforcement responsibility within their existing remit.
This matters because there is no one place to file a notification or check compliance. A financial services firm using high-risk AI will deal with the Central Bank as its competent authority. A health technology provider will deal with the HPRA. A general business deploying high-risk AI outside any specific regulatory regime will deal with the AI Office directly. The mapping is not always obvious — and the Bill is explicit that competent authorities retain primacy in cases of overlap.
The central AI Office will have its own enforcement powers, including compulsory information powers from December 2026. These are significant: the Office can demand documentation, technical files, training data records, and evidence of human oversight without needing a court order. Refusal is itself an offence under the Bill. Fines under the EU AI Act top out at €35 million or 7% of global turnover for prohibited-practice breaches, with proportionate scales below that for other obligations.
What enforcement actually looks like in month one
Predicting the first 90 days requires reading two things: how new regulators have behaved historically, and what the Bill's signals suggest about the AI Office's posture.
Historically, newly established regulators do not lead with prosecutions. They lead with engagement, registration, and information-gathering. The first quarter is typically dominated by guidance publication, sector workshops, and high-level data collection from significant deployers. Enforcement actions are usually held back for clear cases: prohibited practices, AI systems already attracting complaints, and organisations that have visibly failed to engage.
Two signals from the Bill suggest the AI Office will follow this pattern. First, the compulsory information powers do not commence until December 2026 — five months after the Office opens. This is not an oversight. It gives the Office time to engage informally, gather voluntary submissions, and identify the organisations whose AI use looks materially non-compliant before the legal stakes escalate. Second, the Bill includes a defined "good faith" defence for organisations that have engaged with guidance and acted reasonably to comply. That language is consistent with a regulator that wants engagement, not adversarial defence.
The implication is not that the first ninety days will be quiet. The opposite. The AI Office will be very visible — publishing guidance, naming early registrants, identifying sectors of concern, signalling future enforcement priorities. Organisations not visible to the Office in that period will be visible later, when the conversation is no longer informal.
What the AI Office cannot do — yet
Worth being precise about. From 1 August 2026, the AI Office cannot:
- Compel disclosure of information from any organisation that has not voluntarily registered or engaged. The compulsory information powers commence in December 2026.
- Levy administrative penalties for non-compliance with most high-risk AI obligations. The EU AI Act's main high-risk provisions follow a phased timetable. The Omnibus deal in May 2026 moved the substantive deadline for high-risk AI systems to 2 December 2027.
- Prosecute most categories of breach immediately. The criminal offences under the Bill require designated commencement orders and prosecutorial decision-making that will not be in place day one.
What the AI Office can do from day one: enforce the prohibited-practice provisions of the EU AI Act (active since 2 February 2025), enforce Article 4 AI literacy obligations (also active since February 2025), publish guidance, coordinate cross-sector activity, and engage informally with deployers.
This does not mean the August enforcement deadline is soft. It means the architecture is staged. The legal teeth grow sharper over the eighteen months after 1 August. The reputational consequences are immediate.
What Irish organisations should have in place
Three things, in order of priority.
An AI inventory. A documented list of every AI system in use across the organisation, with classification under the EU AI Act risk tiers, ownership assigned, and impact assessed. This is the artefact every regulator will ask for first. Without it, every subsequent conversation starts from behind. With it, the conversation can move to substance.
The audit work to build this inventory is non-trivial. Most organisations underestimate the volume of AI in use. Shadow AI — staff using consumer-grade generative tools without sanction — is rarely captured. AI features embedded in mainstream business software (Microsoft 365 Copilot, Salesforce Einstein, Workday) are often overlooked because they were not procured as "AI systems". A proper inventory surfaces all of it.
Article 4 literacy evidence. Article 4 of the EU AI Act has required since February 2025 that employers ensure staff have sufficient AI literacy for the systems they interact with. The obligation is principles-based and proportionate, which makes it easy to defer. It is also easy to evidence if it has actually been done: training records, policy documents, role-specific guidance, induction materials referencing AI use. Organisations that cannot produce these in the first quarter of 2027 will face questions whose answer determines whether subsequent compliance positions are credible.
Governance accountability mapped to people. Not a policy document. A named accountable executive for AI use, a board-level reporting cadence, an escalation route when an AI system causes harm or behaves unexpectedly, and a documented decision process for new AI deployments. The AI Office will not be impressed by policies. It will be interested in evidence that the policies have produced operational behaviour.
The framing question
A useful diagnostic for board-level conversation between now and 1 August: if an AI Office officer attended our next board meeting as an observer, what would they conclude about our AI governance posture from what they saw on the table?
Most boards we have worked with over the past year would answer that honestly: not much. The board pack would not contain an AI items list. The minutes would not show recent AI-related decisions. The directors would not be able to recite the AI Office's name without checking. The accountable executive for AI would not be identified.
That is the gap August closes. The window to close it on your own terms — without an external prompt, without a regulator visit, without a notifiable incident driving the agenda — is now nine weeks.
The work itself is not difficult. The work is operational, not theoretical. The organisations that complete it now have a governance posture they can defend. The organisations that defer have a governance posture that will be defined for them.
The Bill is moving through the Oireachtas. The AI Office is being staffed. The dates are fixed. The only variable left is whether you are ready when the conversation moves from preparation to enforcement.