This is a ready-to-use SOP for recognizing and reporting a serious incident under Article 73 of the EU AI Act (Regulation (EU) 2024/1689) for a high-risk AI system. It is the reactive counterpart to the post-market monitoring plan, which is where most serious-incident candidates first surface. Replace every <<FILL: ...>> placeholder with your own specifics, set your document numbers and dates, and route it through your normal document control, review, and approval. A worked filled specimen follows the template. Verify the current deadlines, definitions, and market surveillance authority contact points against the Commission’s finalised Article 73 guidance and the current Official Journal text before you rely on this procedure; this is general guidance to adapt, not legal or regulatory advice.
Document control header
| Field | Entry |
|---|---|
| Document title | EU AI Act Serious Incident Reporting |
| Document number | <<FILL: SOP-ID>> |
| Version | <<FILL>> |
| Effective date | <<FILL>> |
| Supersedes | <<FILL: prior version or "New">> |
| Document owner | <<FILL: AI Governance Lead / Regulatory Affairs>> |
| Applies to | <<FILL: all high-risk AI systems in the inventory>> |
1. Purpose
This procedure defines how <<FILL: COMPANY NAME>> recognizes a serious incident involving a high-risk AI system, determines the applicable reporting deadline, and reports it to the correct market surveillance authority within that deadline, so a qualifying event is never worked at ordinary deviation-investigation speed and the clock is never missed for lack of a named owner.
2. Scope
This procedure applies to every AI system classified high-risk under the EU AI Act, whether the provider role is held in-house or the company acts as deployer for a system provided by a third party. It covers detection, the serious-incident determination, the report to the market surveillance authority, and coordination with any parallel reporting obligation (device vigilance under MDR/IVDR, pharmacovigilance, a GxP deviation or CAPA). It does not replace those parallel obligations; it adds the AI-Act-specific determination and clock on top of them.
3. Responsibilities
| Role | Responsibility |
|---|---|
| AI Governance / incident owner | Makes the serious-incident determination, starts the clock, and owns the report to the market surveillance authority |
| System owner | Confirms technical facts about the system’s behaviour in the event and supports the investigation |
| Quality | Coordinates this report with any parallel GxP deviation, device vigilance, or pharmacovigilance reporting for the same event |
| Regulatory Affairs | Confirms the correct market surveillance authority and submission channel, and reviews the report before submission |
| Legal | Reviews framing and confirms no conflicting statement is made across parallel reports |
4. Definitions
- Serious incident: an incident or malfunctioning of an AI system that directly or indirectly leads to the death of a person or serious harm to a person’s health; a serious and irreversible disruption of the management or operation of critical infrastructure; the infringement of obligations under Union law intended to protect fundamental rights; or serious harm to property or the environment. (Regulation (EU) 2024/1689, Article 3, point 49; confirm the current definition against the Official Journal text.)
- Market surveillance authority: the national authority in the member state where the incident occurred, responsible for enforcement and to which Article 73 reports are made.
- Causal link established: the point at which the provider (or, where applicable, the deployer) determines, or has reasonable grounds to believe, that the AI system caused or contributed to the outcome. This, not the date the incident physically occurred, is what starts the clock.
5. Procedure
5.1 Detect and triage
- A candidate event can arrive from post-market monitoring (see the linked plan), a complaint, a deployer report, an internal deviation, or a direct observation.
- Within
<<FILL: e.g. one working day>>of any candidate event touching a high-risk AI system, the AI Governance / incident owner runs the serious-incident determination in section 5.2. - Do not wait for a completed root-cause investigation before making the determination; the determination is about whether the outcome meets the Article 3(49) definition and whether a causal link can reasonably be believed to exist, not about having every fact.
5.2 Make the serious-incident determination
| Question | Answer |
|---|---|
| Did the event lead, directly or indirectly, to a person’s death or serious harm to health? | <<FILL: Yes / No>> |
| Did it lead to a serious and irreversible disruption of critical infrastructure? | <<FILL>> |
| Did it infringe a fundamental-rights obligation under Union law? | <<FILL>> |
| Did it cause serious harm to property or the environment? | <<FILL>> |
| Is there a reasonable basis to believe the AI system caused or contributed to the outcome? | <<FILL>> |
| Determination | Serious incident / Not a serious incident, document the reasoning either way |
A “not a serious incident” conclusion is itself a controlled record; write down the reasoning so it can survive a later question.
5.3 Start the clock and classify the tier
Once the determination is “serious incident” and a causal link is reasonably believed to exist, the clock starts the same day. Classify the tier:
| Tier | Criteria | Deadline from causal link established |
|---|---|---|
| Most severe / widespread | Widespread infringement, or the event meets the Act’s more severe criteria | Immediately, and not later than 2 days |
| Death involved | The event caused a person’s death | Not later than 10 days |
| Default | Any other serious incident | Not later than 15 days |
Confirm the exact tier criteria and deadlines against the current Article 73 text and the Commission’s finalised Article 73 guidance before relying on this table for a live event; the tiers above reflect the Regulation’s text as verified at the time this procedure was issued.
5.4 Report to the market surveillance authority
- Identify the correct national market surveillance authority (the state where the incident occurred) using
<<FILL: your maintained list of national contact points, or the Commission's current guidance>>. - Complete the report using
<<FILL: the Commission's current Article 73 reporting template>>, including what is known at the time; an incomplete initial report is acceptable if the full deadline cannot otherwise be met. - Submit within the deadline from section 5.3. Log the submission time and channel in the record (section 8).
- Submit a follow-up complete report as soon as the remaining information is available, referencing the initial submission.
5.5 Coordinate parallel reporting
- Check whether the same event also triggers MDR/IVDR vigilance reporting (device-constituent systems), a pharmacovigilance report, or a GxP deviation and CAPA.
- Where it does, route the AI-Act report and the parallel report through Quality and Regulatory Affairs together so the facts stated are consistent across both, and reference each report in the other’s record.
- The AI-Act report does not substitute for a parallel obligation, and a parallel report does not substitute for the AI-Act report; both run, coordinated, not duplicated in isolation.
5.6 Close out and feed back
- Once the authority has acknowledged the report (typically within about a week per the Act’s own expectation of the authority), continue the investigation to root cause per your normal deviation or CAPA process.
- Feed the confirmed root cause and any corrective action into the system’s technical documentation lifecycle-changes log and the post-market monitoring plan’s threshold review.
- Close the record in section 8 with the final disposition.
6. Acceptance criteria
- Every candidate event touching a high-risk AI system receives a documented serious-incident determination, including a “not a serious incident” conclusion with its reasoning.
- Where the determination is “serious incident,” the clock starts the day the causal link is reasonably believed established, not the day of the incident and not the day a full investigation concludes.
- The correct tier (2, 10, or 15 days) is applied and the report is submitted within it, with an incomplete initial report used where necessary rather than waiting past the deadline.
- Parallel reporting obligations for the same event are coordinated, not siloed.
- The record is complete, signed, and retained.
7. References
Regulation (EU) 2024/1689 (EU AI Act): Article 73 (reporting of serious incidents); Article 3, point 49 (definition of serious incident); Article 72 (post-market monitoring, the usual detection source). Confirm the Commission’s finalised Article 73 guidance and reporting template, and the current list of national market surveillance authorities, against the Commission’s AI Act service desk before relying on this procedure for a live event. Applicable parallel obligations: MDR/IVDR vigilance reporting, pharmacovigilance reporting, and your internal deviation/CAPA procedure, referenced by number in section 5.5.
8. Record generated: serious incident determination and report record
| Field | Entry |
|---|---|
| System name and ID | <<FILL>> |
| Event description | <<FILL>> |
| Date event detected | <<FILL>> |
| Date causal link established (clock start) | <<FILL>> |
| Determination | Serious incident / Not a serious incident, with reasoning |
| Tier and deadline | <<FILL: 2 / 10 / 15 days>> |
| Market surveillance authority notified | <<FILL>> |
| Report submission date and channel | <<FILL>> |
| Follow-up complete report date (if initial was incomplete) | <<FILL>> |
| Parallel reports coordinated (MDR/IVDR, pharmacovigilance, deviation/CAPA reference) | <<FILL>> |
| Root cause and corrective action reference | <<FILL>> |
| Owner (name, signature, date) | <<FILL>> |
| QA / Regulatory approval (name, signature, date) | <<FILL>> |
9. Revision history
| Version | Date | Author | Summary of change |
|---|---|---|---|
<<FILL: 1.0>> | <<FILL>> | <<FILL>> | Initial issue. |
10. Approvals
| Role | Name | Signature | Date |
|---|---|---|---|
| Author | <<FILL>> | ||
| Reviewer (Quality) | <<FILL>> | ||
| Approver (Regulatory) | <<FILL>> |
Filled specimen
The connected-device dosing-guidance feature (the worked example used in the linked post-market monitoring plan and the main article), completed for an example confirmed event.
| Field | Entry |
|---|---|
| System | Dosing-guidance feature embedded in a connected delivery device, combination product |
| Event | A patient at the extreme upper end of the training population’s weight range received a dosing suggestion later confirmed to have contributed to a serious adverse health outcome |
| Date event detected | Flagged by the post-market monitoring drift signal, confirmed by clinical review three days later |
| Date causal link established | The day clinical review confirmed the suggestion contributed to the outcome, which is the clock start, not the earlier detection date |
| Determination | Serious incident: serious harm to a person’s health |
| Tier and deadline | Death-involved tier does not apply (no death); the event otherwise involves serious harm to health, assessed by Regulatory as meeting the 10-day tier given the severity criteria in the current guidance |
| Market surveillance authority | <<FILL: national authority for the member state where the incident occurred>>, notified within the 10-day window |
| Parallel reports coordinated | MDR vigilance report for the same event filed the same week, coordinated through Quality and Regulatory so both reports state the same facts |
| Root cause and corrective action | Training-data representativeness gap at the extreme weight range; model retrained on an expanded population; technical documentation lifecycle-changes log and post-market monitoring thresholds both updated |
Common inspection findings this SOP prevents
- A qualifying event worked at ordinary deviation-investigation speed, missing the 2, 10, or 15-day clock because nobody had the explicit job of starting it.
- The clock measured from the wrong date, the incident date instead of the date the causal link was established, producing either a late report or an unnecessarily rushed one.
- No documented “not a serious incident” reasoning, so a borderline call cannot be defended later.
- The AI-Act report and a parallel MDR/IVDR or pharmacovigilance report filed independently, with inconsistent facts because no one coordinated them.
- A confirmed root cause never fed back into the technical documentation or the post-market monitoring thresholds, so the same failure mode can recur undetected.
How to adapt this SOP
- Set your document number, owner, and effective date in the header.
- Populate the national market surveillance authority contact list your company actually needs, by member state where you place high-risk systems on the market.
- Point section 5.5 to your real MDR/IVDR vigilance, pharmacovigilance, and deviation/CAPA procedures.
- Confirm the current Article 73 deadlines, the definition in Article 3(49), and the Commission’s finalised guidance and reporting template before issue.
- Rehearse the determination step (5.2) with a realistic drill at least once before relying on it for a live event.