Übernimmt die KI alles? Unser CTO über Banken, EU AI Act und deutsche Qualität | Umair Zaffar
Umair Zaffar November 10, 2026 0 Comments

EU AI Act Article 73: Reporting Serious Incidents, Implemented in Practice

EU AI Act Article 73 requires the provider of a high-risk system to report serious incidents to the market surveillance authority within 15, 10 or 2 days depending on severity. We turn it into a runbook in the governance layer: signals from monitoring and review, classification under Art. 3(49), kill switch, a draft report built from the evidence log.

An incident is the moment that shows whether governance exists or is merely described. Banks and insurers have well-rehearsed processes for IT outages, but hardly any has one for “the model has been disadvantaging a group of people for three weeks”. That is exactly what Article 73 demands. It addresses the provider, in paragraphs 2 to 5 also the deployer, and through Art. 26(5) every deployer of a high-risk system; paragraphs 7, 8 and 11 address the authorities.

The EU AI Act has been in force since 1 August 2024. The prohibitions and AI literacy duties have applied since February 2025, the obligations for GPAI models since August 2025, the transparency obligations under Article 50 since August 2026. The obligations for high-risk systems under Annex III apply, after the postponement by amending Regulation (EU) 2026/1744 (Digital Omnibus), from 2 December 2027, and for Annex I products from 2 August 2028. Article 73 attaches to high-risk systems and follows their timeline; the Omnibus moved deadlines, it removed no obligation. The overview is in EU AI Act in practice.

What Article 73 requires

First the definition. A serious incident under Art. 3(49) is an incident or malfunctioning of an AI system that directly or indirectly leads to one of four outcomes: (a) the death of a person or serious harm to a person’s health, (b) a serious and irreversible disruption of critical infrastructure, (c) an infringement of obligations under Union law intended to protect fundamental rights, (d) serious harm to property or the environment. Point (c) bites most often in the financial sector: equal treatment and data protection are such obligations.

  • Art. 73(1): Providers of high-risk systems placed on the Union market report any serious incident to the market surveillance authority of the Member State where it occurred.
  • Art. 73(2): The report is made immediately after the provider has established a causal link, or the reasonable likelihood of one, and no later than 15 days after the provider or, where applicable, the deployer becomes aware. Severity determines the period.
  • Art. 73(3): For a widespread infringement under Art. 3(61) or an incident under point (b), critical infrastructure, immediately and no later than two days after awareness.
  • Art. 73(4): For the death of a person, immediately after a causal relationship is established or as soon as one is suspected, and no later than ten days after awareness.
  • Art. 73(5): To report in time, the provider or deployer may submit an initial incomplete report and follow up with a complete one.
  • Art. 73(6): After reporting, the provider investigates the incident and the system without delay, assesses the risk, takes corrective action and cooperates with the authorities and, where relevant, the notified body. Before informing them it does not alter the system in a way that would impair the evaluation of the causes.
  • Art. 73(7) and (8): The authority informs the fundamental-rights authorities under Art. 77(1) and acts within seven days under Article 19 of Regulation (EU) 2019/1020; the Commission issues guidance.
  • Art. 73(9) and (10): For Annex III systems from providers under Union law with equivalent reporting obligations, and for medical devices under Regulations (EU) 2017/745 and 2017/746, reporting is limited to point (c).
  • Art. 73(11): National authorities notify the Commission of every serious incident immediately.
  • Art. 26(5) (deployers): The deployer reports a serious incident first to the provider, then to the importer or distributor and the authorities. If the provider cannot be reached, Article 73 applies to the deployer mutatis mutandis.

The technical example

An insurer runs a purchased claims-triage model. Every claim receives a priority and a route: reimburse automatically, case handler, claims review. The insurer is the deployer, the manufacturer the provider. The provider ships an update, the quality gates pass, the business-owned evaluation set shows nothing. Three weeks later two reviewers in the blind sample independently report the same pattern: claims from a group whose common feature correlates with the age of the insured are systematically landing in the lowest priority. Monitoring had seen the drift per group, but the threshold was set on a weekly view.

Where the obligation bites: delayed reimbursements for an age group may be an infringement of equal treatment law, so an incident under point (c). Whether it legally is one is not the engineer’s call. But as soon as the reviewers confirm the pattern, the clock in paragraph 2 is running: 15 days. The insurer has to inform the provider, the provider the authority. And nobody may now quietly roll the model back and let the logs rotate.

The second case is built in-house, so the insurer is provider and deployer at once. An agent prepares correspondence on claims; its allow-list: read file, draft letter, create task. In one run it calls “send letter”, which is not on the list. The gateway blocks, the run stops. No harm, no serious incident, but an incident for post-market monitoring under Art. 72. Had the agent, through an over-privileged tool, actually sent letters to wrong recipients, Art. 33 GDPR with its 72 hours would be in play alongside Article 73.

How we implement it

At SIFAMO we treat Article 73 as a runbook, not a form. The runbook is code in the governance layer; it reads the registry, monitoring and evidence log and writes to the incident file. Simplified:

# incident runbook, simplified (governance layer, module incidents)
SIGNALS = {
  "monitoring": ["drift_per_group", "disagreement_rate", "blocked_tool_call", "budget_stop"],
  "reviewers":  ["overturn_on_protected_ground", "pattern_report"],
  "external":   ["complaint", "deployer_report_art26_5"],
}

def classify(inc):                                   # Art. 3(49)
    if inc.death or inc.serious_health_harm:          return "a"
    if inc.critical_infrastructure_disruption:        return "b"
    if inc.fundamental_rights_infringement:           return "c"   # equal treatment, data protection
    if inc.serious_property_or_environment_harm:      return "d"
    return None            # an incident, but not serious: Art. 72, no report

def deadline(inc, cls):                              # Art. 73(2) to (4), from awareness
    if inc.death:                       return days(10)
    if inc.widespread or cls == "b":    return days(2)
    return days(15)                     # severity may require sooner

def report_chain(rec):                               # who reports to whom
    if rec.role == "deployer":                       # Art. 26(5)
        return ["provider", "importer_or_distributor", "market_surveillance_authority"]
    return ["market_surveillance_authority_of_incident_state"]       # Art. 73(1)
    # provider unreachable: deployer reports itself, Art. 73 mutatis mutandis

def on_incident(inc):
    rec = registry.get(inc.model_id)
    rec.kill_switch.set(by=rec.business_owner, reason=inc.id)   # route removed, manual fallback
    freeze(rec.version, inc.evidence_range)                     # Art. 73(6): change nothing
    cls = classify(inc)
    if cls:
        due = inc.awareness_at + deadline(inc, cls)
        send(initial_report(inc, cls, incomplete=True), to=report_chain(rec), due=due)  # para 5
    corrective_actions_art20(rec, inc)   # rollback, re-assess cases, inform deployers
    if rec.is_financial_entity:
        dora.evaluate(inc)               # separate path, separate deadlines, same file
    return evidence_package(inc, rec)

def evidence_package(inc, rec):
    return {
      "registry_record": rec.snapshot(),
      "model_versions": [rec.version, rec.previous_version],
      "evidence_log": log.query(model=rec.id, range=inc.evidence_range),
      "monitoring": monitoring.export(inc.signals),
      "reviewer_actions": log.reviewer_actions(inc.evidence_range),
      "gate_report": gates.report(rec.version),
      "timeline": inc.timeline(), "reports_sent": inc.reports,
    }

Detection and awareness

Three signal sources feed the same file: monitoring (drift per group, disagreement rate, blocked tool calls, budget stops), reviewers (an overturned case on a protected ground, a pattern report) and external (a complaint, a deployer’s report under Art. 26(5)). Awareness counts from the moment triage confirms the signal as an incident, and at the latest the next working day after the signal; the alert itself stays in the file.

Classification and deadline

Classification follows the four points of Art. 3(49) and is set by the business owner and the risk function. For point (c) a reasoned suspicion is enough for us; legal reviews in parallel, not beforehand. The system computes the deadline from awareness and classification: ten days for a death, two for critical infrastructure or a widespread infringement, otherwise 15. In the example: 15 days, initial report on day three. Paragraph 5 is for me the most important line of the article, because it removes the argument “we do not know enough yet”.

Reporting chain

The registry knows the role per system. For the purchased model the insurer, as deployer, reports to the provider’s incident contact from the registry record and to the authority; the provider reports under Art. 73(1). If the provider stays silent beyond the contractual period, the insurer reports itself (Art. 26(5)). For the agent the insurer is the provider and reports directly. The competent authority is stored in the registry record; for financial institutions that is, under Art. 74(6), the financial supervisor insofar as the system is connected with the financial service.

Kill switch and corrective action under Art. 20

The first action is the kill switch: a registry flag the business owner sets without a deployment; the gateway removes the model’s route, claims continue manually. The second action is freezing the model version, prompt version, index and evidence log for the period. That is paragraph 6 in code. Only then does Art. 20 begin: rollback through the quality gates, re-assessment of all affected cases from the evidence log, information of the people concerned and, via the provider, of the other deployers.

Evidence package and DORA

The evidence package is assembled from what exists anyway: the registry record with versions and signatures, the evidence log for the period, the monitoring export, the gate report of the release, the timeline and every report sent. The insurer is also a financial entity under DORA. We run both regimes from the same file but with separate checks: a model that disadvantages a group is usually not an ICT-related incident in DORA’s sense, and an outage of the model is usually not a serious incident under Art. 3(49). We do not assume that DORA is an equivalent reporting obligation under paragraph 9.

The auditor sees the file: signal, awareness, classification with two names, deadline, initial report, kill switch with timestamp, frozen versions, corrective actions and the complete report. I am an engineer, not a lawyer. But whoever can produce that within an hour has met Article 73 before the deadline gets tight.

Typical mistakes

  • Incident means outage. The incident process knows availability and security vulnerabilities, but has no category for “the model decides wrongly about people”.
  • Root cause first, report second. The clock runs from awareness, not from root cause. Paragraph 5 allows the incomplete initial report for exactly this.
  • Fix first, ask later. The hotfix overwrites the version, the logs rotate, and the evaluation of causes under paragraph 6 becomes impossible. Freezing comes before repairing.
  • “The provider handles that.” Art. 26(5) obliges the deployer to inform the provider and, if the provider stays silent, to report itself.
  • A kill switch with a change request. If stopping needs an IT approval, it is not a kill switch, and the deadline keeps running.

EU AI Act Article 73 checklist

  • The four points of Art. 3(49) set up as incident categories in the incident process.
  • Signals from monitoring, review and external sources flow into one file; the moment of awareness is defined and logged.
  • Classification by two named roles; the 15, 10 and 2 day deadlines are computed, not looked up.
  • Role per system in the registry (provider or deployer), the provider’s incident contact and the competent authority stored.
  • Kill switch as a registry flag for the business owner; freezing version and logs as the first step after the stop.
  • Evidence package exportable at the push of a button from registry, evidence log, monitoring and gate report.
  • Corrective actions under Art. 20, re-assessment of affected cases and the DORA and GDPR paths in the runbook; rehearsed once a year with a synthetic incident.

Open questions

  • When “awareness” begins is not defined: the alert, triage confirmation or the finding of a causal link. We take triage and log the alert until the Commission guidance settles it.
  • Nobody has set the threshold at which a statistically measurable disadvantage becomes an infringement of obligations protecting fundamental rights under point (c). When in doubt we report.
  • Whether DORA counts as equivalent under paragraph 9, and which authority is competent in a given case, still has to sort itself out between Art. 74(6), national law and supervisory practice.

Frequently asked questions

Does this apply to us as a deployer if we only bought the model?

Yes, through Art. 26(5). You report a serious incident first to the provider, then to the importer or distributor and the authorities. If the provider cannot be reached, Article 73 applies to you mutatis mutandis, with the same deadlines running from your awareness. More in AI Governance.

Do we have to report before we know the cause?

Yes. The period starts with awareness, and paragraph 2 requires the report as soon as a causal link is established or reasonably likely. Paragraph 5 allows an incomplete initial report with a follow-up for exactly that.

Does our DORA report replace the report under Article 73?

I would not rely on it. Paragraph 9 limits the reporting duty under equivalent Union regimes to incidents under point (c), but does not remove it. And disadvantaging a group is precisely a point (c) case, not an ICT incident. The other articles in the series: all articles at a glance.

Related: EU AI Act · All articles · AI Governance · AI Consultant

Leave Comment