Death or serious harm to health
Example: an AI-supported triage system misclassifies an emergency, treatment is delayed and the patient suffers lasting harm. A death shortens the deadline to at most 10 days.
Incident reporting · Article 73 AI Act
Updated 6 October 2026 · AI Act incl. Digital Omnibus
Reporting serious AI incidents: Article 73 AI Act, deadlines and process
When does an incident involving an AI system have to be reported, who reports to whom – and how many days are left? This page explains the definition in Article 3(49), the three reporting deadlines, the duties of deployers and the parallel routes under the GDPR and NIS2. With a deadline checker to click through.
- Definition of “serious incident” with examples
- 15, 10 and 2 days checked against the text
- The deployer route via Article 26(5)
- Deadline checker with date calculator
app.simpleact.de
Incidents
Operations & Monitoring
ENA
- HIGH
False-positive rate in screening increased
AI-based Application Screening
IN_PROGRESSReassessment trigger recommended - MEDIUM
Answer with outdated tariff data
Service Chatbot "Lia"
OPEN - LOW
Text output without AI notice
Marketing Copy Generator
RESOLVED
Go-Live Gateblocked: 2 open findings
The report to the authority itself is filed outside the platform.
Short answer
Providers of a high-risk AI system report a serious incident to the market surveillance authority of the Member State where it occurred – immediately after establishing the causal link, and no later than 15 days after becoming aware. A death allows at most 10 days; critical infrastructure or a widespread infringement 2 days. Deployers inform the provider first, without undue delay.
What this is about
Article 73 is a duty for the worst day – and it starts well before it
The AI Act (Regulation (EU) 2024/1689) requires providers of high-risk AI systems to report serious incidents to market surveillance. That sounds like a rare edge case. In practice, whether a report goes out on time is decided weeks earlier: who spots an incident, who assesses it, who is allowed to report, and where is the data an authority will want to see?
The deadlines are short. The general case allows at most 15 days, a death at most 10 days, a serious disruption of critical infrastructure or a widespread infringement only 2 days. Every deadline runs from awareness of the incident – not from the moment the cause is fully understood. That is why Article 73(5) expressly allows an initial, incomplete report followed by a complete one.
Deployers – organisations using a high-risk AI system under their own authority – do not report under Article 73 themselves as long as they can reach the provider. Under Article 26(5), however, they must immediately inform first the provider, then the importer or distributor and the market surveillance authorities. If the provider cannot be reached, Article 73 applies to them accordingly.
On timing: Article 73 sits in Chapter IX of the Regulation, which has applied since 2 August 2026. The duty is tied to high-risk AI systems, though, and after the Digital Omnibus (Regulation (EU) 2026/1744) the classification rules in Chapter III apply to Annex III systems only from 2 December 2027 and to Annex I products from 2 August 2028. In practice the reporting duty therefore becomes relevant on those dates. Building an incident process now leaves enough time to rehearse it.
- maximum in the general case (Art. 73(2))
- 15 days
- maximum in the general case (Art. 73(2))
- in the event of a death (Art. 73(4))
- 10 days
- in the event of a death (Art. 73(4))
- critical infrastructure or widespread infringement (Art. 73(3))
- 2 days
- critical infrastructure or widespread infringement (Art. 73(3))
Legal basis
The provisions behind incident reporting
Article 73 is the centre, but not the only provision. The definition, deployer duties, corrective action and the quality management system all interlock.
Scroll the table sideways →
| Provision | Content | Who it affects | What it means in practice |
|---|---|---|---|
| Art. 3(49) | Definition of “serious incident”: an incident or malfunction that directly or indirectly leads to death or serious harm to health, a serious and irreversible disruption of critical infrastructure, an infringement of Union law protecting fundamental rights, or serious harm to property or the environment. | all operators | The test grid for every incident assessment. Without one of these four consequences there is no serious incident under the AI Act. |
| Art. 73(1) | Providers of high-risk AI systems placed on the Union market report serious incidents to the market surveillance authorities of the Member States where the incident occurred. | providers | What counts is where the incident happened, not where the provider is based. In Germany this is in principle the Federal Network Agency (Bundesnetzagentur). |
| Art. 73(2) | Report immediately after establishing the causal link or its reasonable likelihood, and no later than 15 days after becoming aware; the severity is taken into account. | providers (deployers where applicable) | 15 days is the ceiling, not the target. The more serious the incident, the earlier. |
| Art. 73(3) | For a widespread infringement or an incident under point (b) of No. 49 (critical infrastructure): immediately, no later than 2 days after becoming aware. | providers (deployers where applicable) | On-call cover and decision rights must also work at weekends. |
| Art. 73(4) | In the event of a death: immediately after a causal link is established or suspected, no later than 10 days after becoming aware. | providers or deployers | Here, even a suspected link triggers the duty. |
| Art. 73(5) | An initial, incomplete report is allowed where necessary to report on time; a complete report follows. | providers or deployers | Better on time and incomplete than complete and late. Plan the follow-up. |
| Art. 73(6) | Investigate without delay, including a risk assessment and corrective action; cooperate with the authorities; do not alter the system in a way that may affect a later evaluation of the causes before informing the authorities. | providers | No silent hotfix. Secure evidence and inform the authority first, then change. |
| Art. 73(8) | The market surveillance authority takes appropriate measures under Article 19 of Regulation (EU) 2019/1020 within 7 days of receiving the report. | authority | Expect questions, conditions or measures shortly after reporting. |
| Art. 73(9), (10) | Special rules: where Union law imposes equivalent reporting duties, and for medical devices, only incidents under point (c) of No. 49 (fundamental rights) are reported under the AI Act. | providers in certain sectors | Sector reporting routes (e.g. medical devices) stay in the lead; the AI Act adds fundamental-rights cases. |
| Art. 26(5) | Deployers monitor operation on the basis of the instructions for use; where there is a risk under Article 79(1) they inform the provider/distributor and the authority and suspend use; for a serious incident they immediately inform first the provider, then the importer/distributor and authorities. | deployers | Agree the contact route to the provider in advance, contractually and organisationally. |
| Art. 20 | Corrective action in case of non-conformity; where there is a risk, investigate the causes and inform the market surveillance authorities. | providers | Incident assessment and correction belong together; govern both in the QMS. |
| Art. 17(1)(i) | The quality management system includes procedures for reporting a serious incident in accordance with Article 73. | providers | A written reporting procedure is a mandatory part of the QMS – not just good practice. |
| KI-MIG § 2 (DE) | In Germany the Federal Network Agency is the market surveillance authority unless provided otherwise; BaFin for AI directly linked to regulated financial activities; for Annex I products the authorities already responsible. | everyone in Germany | Decide the recipient per system in advance and record it in the incident plan. |
Paraphrased, not the official text. The Official Journal wording prevails (links below). Checked on 6 October 2026.
Definition
What a “serious incident” is – and what it is not
Article 3(49) lists four consequences. If none of them occurs, an error is annoying but not a serious incident under the AI Act.
The starting point is “an incident or malfunctioning of an AI system”. That includes cases where the system works as designed but leads to a harmful outcome in use. What matters is the consequence, not the type of error. The consequence may be direct or indirect – for instance when a person makes a decision based on an AI recommendation that later causes the harm.
The fundamental-rights limb (point c) is the most likely one in everyday business. It applies where obligations under Union law intended to protect fundamental rights are infringed – for example non-discrimination rules or data protection law. An application-screening tool that systematically disadvantages one group can fall under it even though nobody is physically harmed.
An important boundary: an incident with an AI system that is not high-risk does not fall under the Article 73 reporting duty. It may still trigger other duties – a personal data breach under Article 33 GDPR, a reportable security incident under NIS2, or product liability questions. An internal incident register covering all AI systems is therefore sensible, even if only some of them can ever become reportable.
Serious, irreversible disruption of critical infrastructure
Example: an AI control in grid operation causes an outage that cannot be reversed in the short term. Critical infrastructure is defined in Article 3(62) by reference to Directive (EU) 2022/2557. Deadline: at most 2 days.
Infringement of Union law protecting fundamental rights
Example: a credit-scoring model disadvantages applicants because of a protected characteristic, or a system processes sensitive data unlawfully. Usually at most 15 days; 2 days for a widespread infringement.
Serious harm to property or the environment
Example: an AI-supported plant control causes a fire or a release of pollutants. Usually at most 15 days.
Errors without these consequences
A chatbot gives the wrong opening hours, a text generator invents a quote, a forecast is off – without any of the four harms: not a serious incident. Record it internally and analyse it in monitoring.
“Widespread infringement”
An infringement of Union law protecting individuals that harms interests in at least two Member States other than the one of origin, or occurs as the same practice in at least three Member States. The 2-day deadline then applies as well.
Reporting deadlines
Which deadline applies? The three cases at a glance
Every deadline is a maximum and runs from the moment the provider or, where applicable, the deployer becomes aware of the incident. “Immediately” remains the real benchmark.
Scroll the table sideways →
| Case | Maximum | Trigger | Provision | Note |
|---|---|---|---|---|
| General case (No. 49(a) without death, (c), (d)) | 15 days after awareness | causal link established or reasonably likely | Art. 73(2) | severity shortens what is appropriate |
| Death of a person | 10 days after awareness | link established or suspected | Art. 73(4) | a suspicion is already enough |
| Critical infrastructure (No. 49(b)) | 2 days after awareness | serious and irreversible disruption | Art. 73(3) | plan for weekends and public holidays |
| Widespread infringement (No. 61) | 2 days after awareness | infringement affecting several Member States | Art. 73(3) | authority may inform other bodies |
| Information still missing | deadline unchanged | incomplete initial report | Art. 73(5) | complete report follows |
| Deployer cannot reach provider | as above | deployer reports itself | Art. 26(5), Art. 73 | document the contact attempts |
The AI Act does not spell out how days are counted (start of day, weekends). Recommendation: count conservatively and record the time of awareness to the minute.
How it works
From the first signal to the final report
The process is a recommendation for turning the statutory deadlines into an internal procedure. Steps 3 and 5 are where deadlines are won or lost.
- 1
Detect and record
A signal comes from monitoring, customer service, a business unit, a complaint or the deployer. It is logged as an incident straight away – with the time of awareness, the system concerned, a description and first evidence. The time of awareness starts every deadline.
- Who
- whoever notices it, system owner
- Output
- incident record with timestamp
- Basis
- Art. 73(2)
- 2
Secure the evidence
Logs, inputs, outputs, model and configuration state are preserved before anyone changes the system. Article 73(6) prohibits changes that may affect a later evaluation of the causes before the authorities have been informed.
- Who
- IT operations, system owner
- Output
- secured logs and snapshot
- Basis
- Art. 12, Art. 73(6)
- 3
Classify: serious or not?
Deadline-criticalCheck against Article 3(49): is one of the four consequences present? Is the system high-risk? Is a causal link established, likely or suspected? Record the result and the reasoning – even when the answer is “not reportable”.
- Who
- incident team with compliance/legal
- Output
- documented classification
- Basis
- Art. 3(49)
- 4
Clarify the role, choose the route
Providers report themselves. Deployers inform the provider without delay, then the importer or distributor and the authority. If the provider cannot be reached, the deployer reports under Article 73. In parallel: personal data breach under Article 33 GDPR, security incident under NIS2.
- Who
- compliance, data protection, information security
- Output
- reporting plan with recipients
- Basis
- Art. 26(5)
- 5
Initial report within the deadline
Deadline-criticalDepending on the case, no later than 2, 10 or 15 days after awareness – and always as early as possible. If information is missing, an incomplete initial report goes out, marked as such.
- Who
- authorised reporter (named in advance)
- Output
- initial report with proof of receipt
- Basis
- Art. 73(2)–(5)
- 6
Investigate and correct
Root-cause analysis, risk assessment of the incident, corrective action. Cooperation with the authority and, where relevant, the notified body. Changes to the system only after the authority has been informed.
- Who
- provider, engineering, deployer supporting
- Output
- investigation report, action plan
- Basis
- Art. 73(6), Art. 20
- 7
Complete report and re-assessment
The complete report supplements the initial one. Risk management, technical documentation and, where needed, the instructions for use are then updated, and the lessons feed into the post-market monitoring plan.
- Who
- provider
- Output
- final report, updated documentation
- Basis
- Art. 9, Art. 72
- 8
Close and rehearse
The incident is only closed once actions are implemented and evidenced. Recommendation: rehearse the process at least once a year with a fictitious case, including a weekend scenario for the 2-day deadline.
- Who
- incident team, management informed
- Output
- closed incident, exercise record
- Basis
- Art. 17(1)(i) (recommendation)
Roles are suggestions. Whoever reports should be named in writing in advance – with a deputy.
Deadline check
Is this a serious incident under Article 73 – and by when must it be reported?
Answer the questions for a specific incident. The check shows the classification, the applicable maximum deadline with a date and the next steps for your role. Everything runs in your browser; nothing is stored.
Without a date the check shows the deadline in days.
app.simpleact.de
Result
–
ENA
Answer at least role, the high-risk question and the consequence.
Orientation only, not legal advice. Make the case-by-case classification with your legal or compliance function.
Parallel reporting
AI Act, GDPR and NIS2: one incident, several clocks
An AI incident is often a data breach or a security incident at the same time. The reporting routes then run side by side – with different recipients and deadlines.
An example: a customer-service AI assistant discloses other customers’ contract data because of a configuration error. Whether that is a serious incident under the AI Act depends on whether the system is high-risk and whether Union law protecting fundamental rights was infringed. A personal data breach under Article 33 GDPR, on the other hand, is almost certainly present – with a deadline of, where feasible, 72 hours to the data protection authority.
Entities covered by the NIS2 transposition in Germany also face § 32 of the BSI Act: an early warning within 24 hours, an incident notification within 72 hours and a final report no later than one month afterwards. An internal incident process should therefore not be split by statute; it needs one intake and a switch that checks all regimes at once.
Scroll the table sideways →
| Regime | Trigger | Who reports | To whom | Deadline |
|---|---|---|---|---|
| AI Act Art. 73 | serious incident with a high-risk AI system | provider (deployer if provider unreachable) | market surveillance authority where it occurred (DE: in principle the Federal Network Agency) | immediately, at most 15 / 10 / 2 days |
| AI Act Art. 26(5) | serious incident or risk in operation | deployer | provider first, then importer/distributor and authority | without undue delay |
| GDPR Art. 33 | personal data breach posing a risk | controller | competent data protection authority | without undue delay, where feasible within 72 hours |
| GDPR Art. 34 | likely high risk to data subjects | controller | data subjects | without undue delay |
| BSIG § 32 (NIS2, DE) | significant security incident | essential and important entities | joint reporting office of BSI and BBK | 24 h early warning, 72 h notification, 1 month final report |
Article 34 GDPR for orientation only; the duty depends on a likely high risk. Sector-specific duties (e.g. medical devices, financial supervision) may apply on top.
Deployers
What organisations must do when they only use AI
Most mid-sized companies are deployers, not providers. For them Article 26(5) is the key provision – and it asks for more than many expect.
Deployers monitor the operation of the high-risk AI system on the basis of the instructions for use. That is not a passive duty: if you do not know the instructions, you cannot recognise deviations. Where deployers have reason to consider that use in accordance with the instructions may present a risk to health, safety or fundamental rights, they inform the provider or distributor and the market surveillance authority without undue delay and suspend use.
Where deployers identify a serious incident, they immediately inform first the provider, then the importer or distributor and the relevant market surveillance authorities. If the deployer cannot reach the provider, Article 73 applies accordingly – and the 15-, 10- and 2-day deadlines then run for the deployer itself.
Logging comes on top: deployers keep the automatically generated logs, to the extent they are under their control, for at least six months (Article 26(6)). Without these logs an incident can hardly be reconstructed. Breaches of deployer obligations under Article 26 can be fined under Article 99(4) up to €15 million or 3% of worldwide annual turnover; for SMEs the lower amount applies.
- Record the provider’s incident contact route in writing (address, phone, response time) – recommendation: in the contract.
- Make the instructions for use available to the people exercising human oversight for each system.
- Internal reporting line: anyone in a business unit with a suspicion knows whom to contact.
- Keep logs for at least six months and protect them from being overwritten once an incident is suspected.
- Assign the authority to suspend use in advance – otherwise everyone waits for everyone else.
- Note the authority per system: in Germany in principle the Federal Network Agency, BaFin for financial activities.
Report contents
What belongs in an initial report
The AI Act does not prescribe a form; Article 73(7) provides for Commission guidance. Until a binding template is settled, this outline works well as a recommendation.
- Reporting organisation and role (provider, deployer, authorised representative) with contact person and deputy.
- AI system concerned: name, version, EU database registration number if available, intended purpose.
- Time of the incident and time of awareness – stated separately.
- Location and Member States affected; whether a widespread infringement may be involved.
- Description of what happened and the consequence, with reference to points (a)–(d) of Article 3(49).
- Assessment of the causal link: established, likely or suspected.
- Immediate measures: suspension, restriction, warnings to deployers, recall.
- Whether this is an incomplete initial report and when the complete report will follow.
Recommendation, not an official template. Before reporting, check whether the authority prescribes its own form or portal.
After the report
Investigation, correction and the authority’s seven days
The real work starts with the report. The provider investigates without delay, assesses the risk and takes corrective action, cooperating with the authorities and, where relevant, the notified body. A trap for engineering teams: changes that may affect a later evaluation of the causes are only permitted after the authority has been informed.
The market surveillance authority takes appropriate measures within seven days of receipt. Where the incident concerns fundamental rights (point (c) of No. 49), it also informs the national public authorities protecting fundamental rights under Article 77(1). Questions therefore come quickly – anyone who only then starts gathering documents loses time and credibility.
Once closed, the findings flow back: into risk management under Article 9, into the technical documentation, into the post-market monitoring plan under Article 72 and, where needed, into the instructions for use for deployers. How that plan is built is explained on the post-market monitoring page.
Common mistakes
Where incident reports fail in practice
Deadline counted from root cause
Deadlines run from awareness of the incident. Waiting for the root-cause analysis quickly breaks a 2- or 10-day deadline. The fix: an incomplete initial report.
Nobody is authorised to report
On Friday evening the only person allowed to report is on holiday. Recommendation: name the authorised reporter and a deputy in writing.
Hotfix before evidence
The team fixes the bug immediately and overwrites logs and the model state in the process. That can breach Article 73(6) and hampers the investigation.
Deployer waits for the provider
The deployer informs the provider and considers the matter closed. Article 26(5) also requires informing the distributor or importer and the authority.
Only one regime checked
The AI Act report is fine, but the personal data breach under Article 33 GDPR was forgotten. One intake with a switch for all regimes prevents this.
“Not reportable” without reasons
A decision not to report needs evidence just as much as a report does. If it is not documented, it cannot be explained later.
In SimpleAct
Record incidents where the system file lives
SimpleAct does not replace the report to the authority – that is filed outside the platform. It does keep incidents, re-assessment and evidence on the AI system concerned, so a report does not start with a hunt for documents.
- Create an incident per AI system: title, description, severity LOW, MEDIUM, HIGH or CRITICAL
- Status workflow OPEN → IN_PROGRESS → RESOLVED → CLOSED
- For HIGH or CRITICAL the system recommends a reassessment trigger with a deadline
- Open incidents and triggers block the Go-Live Gate
- Closing creates an entry in the audit log
- Take over monitoring signals from Grafana, Prometheus or Datadog as evidence
- Sync incidents as ServiceNow records
- Classification, technical documentation and logs on the same system
Feature description according to the product documentation. Availability of individual features depends on the plan; details on the pricing page or in a demo.
Open-Source Framework
simpleact-ai-act-templates
Open-source templates for EU AI Act incident reporting, CAPA processes, and regulatory communication.
FAQ
Frequently asked questions about reporting serious AI incidents
Under Article 3(49) AI Act it is an incident or malfunctioning of an AI system that directly or indirectly leads to death or serious harm to health, a serious and irreversible disruption of critical infrastructure, an infringement of Union law protecting fundamental rights, or serious harm to property or the environment.
In the general case immediately after establishing the link, no later than 15 days after awareness. For the death of a person at most 10 days, for a serious disruption of critical infrastructure or a widespread infringement at most 2 days. All deadlines run from awareness of the incident.
Under Article 73 the provider of the high-risk AI system reports. Under Article 26(5) deployers immediately inform first the provider, then the importer or distributor and the authority. If they cannot reach the provider, Article 73 applies to them accordingly.
The market surveillance authority of the Member State where the incident occurred. In Germany this is, under § 2 KI-MIG, in principle the Federal Network Agency; BaFin for AI directly linked to regulated financial activities, and for Annex I products the market surveillance authority responsible for them.
Article 73 only concerns high-risk AI systems. An incident with a system that is not high-risk is not reportable under it, but may trigger other duties, such as a personal data breach under Article 33 GDPR or a security incident under NIS2.
Yes. Article 73(5) allows an initial, incomplete report where necessary to report on time. The complete report follows. That is usually better than missing the deadline.
Careful: under Article 73(6) the provider must not carry out an investigation that alters the system in a way that may affect a later evaluation of the causes before informing the authorities. Secure the evidence, inform the authority, then change.
Article 73 is part of Chapter IX and has applied since 2 August 2026. It is tied to high-risk AI systems, however, whose classification rules apply after the Digital Omnibus from 2 December 2027 for Annex III and from 2 August 2028 for Annex I. From then on the duty becomes relevant in practice.
They are separate duties with different recipients and deadlines. One incident can trigger both: the report under Article 73 AI Act to market surveillance and the notification under Article 33 GDPR to the data protection authority, where feasible within 72 hours.
The provider investigates, assesses the risk and takes corrective action. The market surveillance authority takes appropriate measures within seven days of receipt. In fundamental-rights cases it also informs the competent national authorities protecting fundamental rights.
The AI Act does not say so literally, but the quality management system must contain a reporting procedure (Article 17(1)(i)). A documented reason why a case was not reportable is the only way to show that later. We expressly recommend it.
More questions? We're happy to help. Send email · Get started
Sources and status
Updated: · SimpleAct editorial team
- Regulation (EU) 2024/1689 (AI Act), EUR-Lex
- Regulation (EU) 2026/1744 (Digital Omnibus), EUR-Lex
- Article 73 AI Act – reporting of serious incidents (ai-act-law.eu)
- Article 3 AI Act – definitions (ai-act-law.eu)
- Article 26 AI Act – obligations of deployers (ai-act-law.eu)
- KI-MIG § 2 – competent market surveillance authorities (German)
- Article 33 GDPR – notification to the supervisory authority
- § 32 BSIG – reporting obligations (German)
This page is general information and not legal advice. Whether a specific incident is reportable depends on the individual case; agree the assessment with your legal counsel. Provisions are paraphrased.
Changes
- 2026-10-06 – Page rebuilt: definition with examples, deadline table, deployer duties, parallel reporting routes, deadline check; competent authority under KI-MIG added.
- 2026-04-03 – First version as the incident management product page.
Record incidents before the clock runs
See in a demo how incidents, re-assessment and evidence come together on the AI system in SimpleAct – or try it free for seven days.