Regulation (EU) 2024/1689 · Artificial Intelligence Act ● Annex III: — days

AI Incident Reporting Under Article 73: The 15, 10 and 2-Day Deadlines

When a high-risk system goes wrong, Article 73 starts a clock — 15 days in the ordinary case, tighter where people can be hurt. What counts as serious, who reports what, and the runbook that turns panic into a filed report.

Renomira Team · 9 min read

Incident reporting is where paper compliance meets production reality. Article 73 obliges providers of high-risk AI systems to report serious incidents to the relevant market surveillance authority — without undue delay, and in any event within the deadlines — while deployers must not sit on their hands either: they must inform the provider immediately, and report themselves where the provider does not act.

An emergency vehicle responding at night — incident response against the clock
Photo: Tim Evanson / CC BY-SA 2.0 via Wikimedia Commons

The deadlines are the part people remember: 15 days in the ordinary case, with tighter clocks — 10 days for serious disruption of critical infrastructure, and 2 days where there is death or serious harm to health and safety. Miss them and you are in Article 99 territory on top of whatever caused the incident.

What counts as “serious”

The definition is impact-based, not size-based. A serious incident is one that:

  • causes or risks death or serious harm to health;
  • causes serious and irreversible disruption of the operation of critical infrastructure the system supports (or risk thereof);
  • infringes fundamental rights — discriminatory outcomes or denial of access that hit the rights the system was bound to respect, with consequences serious enough to be more than an individual complaint;
  • has equivalent serious effects on public security or the environment, where the system’s failure drives them.

Practical mid-market cases that clear the bar: a triage model systematically misroutes emergency tickets; an access-control face system locks out a protected group; a screening tool’s failure produces discriminatory rejection at scale; a model outage takes a grid-adjacent or payments-adjacent control loop down.

What does not clear it: a bug caught by a user, a model quirk with no affected persons, a vendor incident where your deployment was untouched. Severity of consequence is the test — plus the version-affecting clause: where the provider’s high-risk system or its components cause the incident, scope covers the model version and related systems deployed at the time.

Who reports what

Party Duty Clock
Deployer Inform the provider of an incident caused by the system under your use Immediately
Deployer Report to the market surveillance authority yourself if the provider does not As soon as you know the provider hasn’t
Provider Report the serious incident to the relevant market surveillance authority Without undue delay — 15 days at the outside; 10 days critical-infrastructure; 2 days death or serious harm
Provider Supply intermediate reports, root-cause analysis and corrective actions Per authority request / follow-up cycle

The provider carries the primary filing, which is why procurement must fix incident contact paths before you need them: a deployer whose vendor cannot be reached at hour two has already triggered its own reporting duty.

What goes into the report

Article 73(2) content, compressed:

  1. Nature of the incident — what happened, when, where in the pipeline;
  2. Severity — actual and potential consequences, affected persons and count where known;
  3. System — type, version, how long deployed;
  4. Root cause — initial assessment, “unknown” if genuinely unknown;
  5. Measures — taken or proposed: containment, rollback, user notification;
  6. Distribution — who else (deployers, users) has been informed.

The report goes through the authority’s designated channel (the AI Act reporting system the Commission stood up). A second, fuller submission follows when facts firm up — the first filing’s job is starting the clock honestly, not completing the investigation.

The runbook, sized for a mid-market team

T+0 — detect and classify. One question: does this meet the serious bar? Route through the screening definitions above. Where uncertain, report — Article 73 rewards the early call and punishes the argument.

T+0 — contain. Roll back the model version or switch to manual. Containment never waits for the report; the report never waits for containment to finish.

T+1h — start the clock record. Log the moment awareness began. The legal deadline runs from awareness, which means somebody must own the timestamp — the on-call engineer who first saw it.

T+24h — draft against the six fields. Facts first, causes marked provisional. If the 2-day band plausibly applies (death or serious harm), the filing is the day’s only priority.

After filing — close the loop. Root-cause analysis, corrective actions, deployer notifications, register entry, and a post-incident review that updates the monitoring plan under Article 72. Incidents are also product input: the pattern that fired once fires again.

Test it before you need it

Article 73 has never once been met by a team’s first live attempt. Tabletop it: take your highest-impact system, invent an incident at 16:00 on a Friday, and run the clock. The failures you find — no named owner, no vendor contact, no authority channel — are fixable in an afternoon. The ones you find during a live 2-day clock are not.

How to check your own position in four minutes

The free exposure scan asks twelve questions about your AI usage and governance and returns an exposure score, your ranked exposures with the exact article that fired, and your first three moves — mapped to the dates that apply. Confirm your email and the full report unlocks.

Frequently asked questions

What is a serious incident under the AI Act?

One that causes or risks death or serious harm to health; causes serious and irreversible disruption of critical infrastructure operation; or infringes fundamental rights — assessed by consequence, not by bug size. Provider-side, the scope covers the high-risk system and related systems deployed at the time, including model versions. Anything clearing the bar starts the Article 73 clock.

How long do I have to report an AI incident?

Without undue delay and in any event within 15 days of becoming aware for ordinary serious incidents; 10 days where critical infrastructure suffers serious disruption; 2 days where there is death or serious harm to health and safety. Deployers must inform the provider immediately and step in with an authority report if the provider does not act.

Do deployers report to the regulator, or only the provider?

Both paths exist. Deployers’ primary duty is immediate notification to the provider. If the provider fails to report — unreachable, unresponsive, gone — the deployer must report to the market surveillance authority itself. Procurement should fix the provider contact path in advance, because “we emailed support” at hour twelve is not an immediate notification.

What happens if we miss the Article 73 deadline?

Late or missing incident reporting is an enforcement exposure under Article 99 — up to €15M or 3% of worldwide turnover for breach of most obligations, less for supplying incorrect or misleading information — layered on top of whatever consequence the incident itself carried. Authorities weigh duration, cooperation and self-reporting, which all favour an early honest filing over a late complete one.

How does incident reporting relate to Article 72 post-market monitoring?

Article 72 is the standing system — monitoring plans, thresholds, periodic reporting on high-risk systems’ performance. Article 73 is its trigger path: when monitoring detects a serious event, the reporting clock starts. A monitoring plan without a reporting runbook is a dashboard; a runbook without monitoring waits for a customer to tell you.

Next step

Put this guide to work on your own estate.

Run the free exposure scan to see which of these articles actually apply to your company — 12 questions, four minutes, no card. Confirm your email and the full report unlocks.

  • Free forever plan
  • No credit card
  • EU-hosted · GDPR-native
Run a free scan Sign up free

Build a live AI register in minutes — upgrade only when you need the audit pack.

AI Act Article 73 incident reportingAI serious incident definition EUAI Act 15 day incident reportAI system incident notification authorityAI Act breach reporting deadlinepost-market monitoring AI Actwhat is a serious incident AI Act

Keep reading