What counts as a serious incident
A serious incident is reportable by the provider of a high-risk AI system to the market surveillance authorities of every member state where it occurred, under Article 73(1) of Regulation (EU) 2024/1689. It is due immediately once a causal link is established, or on awareness alone where limb (b) or a widespread infringement is engaged, with an outer limit of 15, 10 or two days from awareness.
The definition is in Article 3, point (49): an incident or malfunctioning of an AI system that directly or indirectly leads to one of four outcomes. A malfunctioning counts on its own, and indirect causation counts, so a chain running through a human decision does not break the link. The Digital Omnibus, Regulation (EU) 2026/1744, amended neither Article 3(49) nor Article 73 nor Article 26, though it did insert an express derogation from Article 73 in the new Article 75(1a).
| Article 3(49) limb | What it covers | Outer limit from awareness |
|---|---|---|
| (a) | Death of a person, or serious harm to a person's health | 10 days on death, otherwise 15 |
| (b) | Serious and irreversible disruption of the management or operation of critical infrastructure | Two days |
| (c) | Infringement of obligations under Union law intended to protect fundamental rights | 15 days |
| (d) | Serious harm to property or the environment | 15 days |
Limb (b) is narrower than it reads: both "serious" and "irreversible" must hold, and it covers disruption of the management or operation of critical infrastructure, not damage to the asset. Article 3(62) takes the term from Article 2, point (4) of Directive (EU) 2022/2557.
Limb (c) carries no qualifier where (b) and (d) require seriousness, and no de minimis appears anywhere in the Regulation. That omission reads as deliberate, so treat any infringement of fundamental-rights protective Union law as engaging the limb rather than filter on gravity. Limb (c) alone survives the narrowing rules in Article 73(9) and (10), and it triggers the Article 73(7) referral to the bodies listed under Article 77(1).
Who reports, and to whom
Article 73(1) binds providers, and its addressee is plural: the market surveillance authorities of the member states where the incident occurred. That need not be the provider's home state, and one incident spanning several markets produces several reports, to the bodies set out in the national authorities guide.
Two rules redirect the filing. Article 73(10) sends reports about medical devices under Regulation (EU) 2017/745 or (EU) 2017/746 to the national competent authority chosen by the state where the incident occurred. Article 75(1a), inserted by the Omnibus and in force since 27 July 2026, sends the report to the AI Office instead where the system falls under its exclusive competence under the widened Article 75(1), with Article 73(2) to (9) applying mutatis mutandis.
Deployers do not report under Article 73, and the summary that they simply inform the provider is half the duty. Article 26(5) requires a deployer that has identified a serious incident to immediately inform first the provider, and then the importer or distributor and the relevant market surveillance authorities. Where it cannot reach the provider, Article 73 applies mutatis mutandis and the deployer inherits the provider's clocks. The Regulation does not say what counts as unable to reach, which puts that escalation route in the contract rather than the statute, and into vendor due diligence.
A lower trigger sits in the same paragraph: a deployer with reason to consider that use in accordance with the instructions may result in the system presenting a risk within Article 79(1) must inform the provider or distributor and the market surveillance authority without undue delay, and suspend use. No harm need have materialised, and the serious-incident limb requires no suspension. Which duty lands on which party turns on role, so settle provider or deployer status first.
The three clocks
No day count runs from establishing the causal link. The link triggers the word "immediately"; every outer limit runs from awareness of the incident by the provider or, where applicable, the deployer.
| Regime | Immediate report triggered by | Outer limit from awareness |
|---|---|---|
| General, Article 73(2) | A causal link between system and incident, or the reasonable likelihood of one | 15 days |
| Widespread infringement, Article 3(61), or critical infrastructure, Article 3(49)(b), Article 73(3) | No causal-link condition: immediate on awareness | Two days |
| Death of a person, Article 73(4) | The provider or deployer having established, or as soon as it suspects, a causal relationship | 10 days |
Three points get reversed in practice. The death case has the longer backstop, 10 days against two, but the lower immediacy trigger, since suspicion is enough. The two-day regime has no causal-link limb, so it cannot be deferred while the investigation runs. And 15 days is a ceiling, not an entitlement: the second subparagraph of Article 73(2) requires the period to take account of severity. Article 73(5) then makes the deadlines absolute, permitting an incomplete initial report followed by a complete one, so an unfinished investigation never excuses a missed clock.
The Regulation never says when a corporate provider becomes aware, or whose knowledge counts. Since Article 73(2) names the deployer's awareness alongside the provider's, a provider cannot postpone its own backstop by pointing to a slow deployer escalation. Nor is Article 3(61) listed in Article 3(49), so a widespread infringement is reportable only where it also engages limb (c), which it ordinarily will.
Where other Union law takes over
Two provisions narrow the duty to the same limb. Under Article 73(9), where an Annex III high-risk system comes from a provider subject to Union legislative instruments laying down equivalent reporting obligations, notification shrinks to Article 3(49)(c) incidents only. Article 73(10) shrinks the medical device case to the same limb and changes the recipient. Paragraph 9 covers Annex III systems, and Annex I embedded products get no equivalent relief beyond the medical device rule. The Regulation names no instrument as equivalent and nobody to decide equivalence, and as an exception to a reporting duty the burden sits with the provider invoking it, so record which instrument is relied on and map it against all four limbs.
General-purpose AI is a separate regime. Article 55(1)(c) binds providers of general-purpose AI models with systemic risk only, and requires them to report serious incidents and possible corrective measures, without undue delay, to the AI Office and as appropriate to national competent authorities. There is no day count, so the 15, 10 and two-day clocks do not apply, and Article 55(2) allows reliance on a code of practice under Article 56 until a harmonised standard is published. Where one organisation is both model provider and high-risk system provider, nothing says one filing discharges both, and the standards differ, so file both. The addressee will often be the same body, because Article 75(1), point (a) puts a high-risk system built on that provider's own general-purpose AI model under the AI Office's exclusive competence, and Article 75(1a) then routes the Article 73 report to the AI Office as well, unless one of the carve-outs in that point applies. The model-level picture is in the GPAI obligations guide.
What the process has to capture
Each item below is there because a provision turns on it.
- First awareness, timed, in the organisation and at any deployer: every day count runs from it under Article 73(2), (3) and (4).
- Which Article 3(49) limb is engaged, because it selects the clock and decides whether Article 73(9) or (10) shrinks the duty, and whether Article 3(61) also applies.
- When a causal link was established, its likelihood formed, or in a death case suspected, because that starts the immediacy duty, plus the Article 73(6) risk assessment and corrective action.
- Every member state where the incident occurred, because Article 73(1) owes a report to the authorities of each, and whether Article 75(1a) redirects it to the AI Office.
The feedstock is Article 72, whose post-market monitoring system must actively and systematically collect, document and analyse performance data across the system's lifetime, including data from deployers. The Omnibus replaced Article 72(3): the plan stays mandatory and part of the Annex IV technical documentation, but the binding implementing-act template is gone, and non-binding Commission guidance and a template are due by 2 September 2027.
Article 73(6) can make a hotfix unlawful. No investigation may alter the system in a way which may affect a later evaluation of the causes before the competent authorities are told, yet Article 20(1) requires corrective action immediately. The two reconcile: withdrawing, disabling or recalling is corrective action, not investigation, but overwriting the failing weights, configuration or logs is what paragraph 6 catches.
Dates, and what a failure costs
Article 73 sits in Chapter IX, Section 2, not Chapter III. The Omnibus amended the third paragraph of Article 113, and its new point (c) defers Chapter III, Sections 1, 2 and 3, with the exception of Article 6(5), to 2 December 2027 for systems classified as high-risk under Article 6(2) and Annex III, and to 2 August 2028 for those classified under Article 6(1) and Annex I. Chapter IX is in no carve-out, so on the face of the text Article 73 has applied since the general date of 2 August 2026. What is deferred is the deployer duty, because Article 26 is in Chapter III, Section 3.
The Regulation does not resolve the tension that leaves: Article 73 is textually live while the Chapter III classification rules and provider obligations it depends on are not, so it has little to bite on until the high-risk requirements apply. Article 111(2) also excludes systems placed on the market before that date unless significantly changed in design, with a 2 August 2030 backstop for public authorities. The guidance that would settle the mechanics is overdue: Article 73(7) required it by 2 August 2025, and the draft published on 26 September 2025 is still a draft as at 6 August 2026.
The exposure is asymmetric, and the usual 3% line is only half right. A deployer in breach of Article 26(5) is inside Article 99(4)(e): up to €15,000,000 or 3% of total worldwide annual turnover, whichever is higher. A provider that fails to report is not. Article 99(4) is a closed list whose provider hooks are Article 16 and, since the Omnibus inserted point (da), Article 25(2) and (4), and neither reaches Article 73, so the failure falls to Article 99(1), amended by the Omnibus to reach "any infringement of this Regulation", and the ceiling is national law, not a figure in the Regulation. Two things pull it back up: Article 16(j) covers the Article 20 corrective-action duties, which the same facts usually trigger, and under the new Article 75c(4) an infringement of any provision "including those not listed in Article 99(4)" sits in the 3% tier for systems under AI Office competence. Article 99(7) treats cooperation and self-notification, at points (f) and (h), as mitigating factors, as the penalties guide sets out.