Incident tickets are where governance stops being a slide deck. Over the last year, the EU, California and House oversight letters have pointed at the same operational gap from different angles: when an AI system or frontier model goes sideways, who writes it down, how fast, and to whom. That is the core of AI incident reporting governance. I would treat a repeatable incident record as the practical spine of any shop that expects to answer a regulator, a customer or its own board without inventing the chronology under pressure.
Templates are no longer imaginary homework. The European Commission has already published draft Article 73 guidance and a high-risk incident template, plus a separate template for systemic-risk GPAI serious incidents. The useful work is matching an internal playbook to duties that already differ by regime, not collapsing them into one imaginary U.S. federal mandate.
Three Regimes, Three Clocks
Start with what is enacted, what was requested in writing, and what is not yet due.
EU AI Act Article 55 Versus Article 73
Regulation (EU) 2024/1689, as consolidated through the Digital Omnibus on AI, draws a hard line between two reporting tracks. Article 55 reaches providers of general-purpose AI models with systemic risk. Among other duties, they must track, document and report, without undue delay, relevant serious-incident information and possible corrective measures to the AI Office and, as appropriate, national competent authorities. Chapter V generally applies from Aug. 2, 2025, but Article 111(3) gives providers of GPAI models placed on the market before that date until Aug. 2, 2027 to comply. On Nov. 4, 2025, the Commission published a reporting template for those systemic-risk GPAI serious incidents and pointed providers to related draft Article 73 materials for high-risk systems.
Article 73 is a different duty. It sits in Chapter IX, on post-market monitoring and serious-incident sharing, not in Chapter III. It applies to providers of high-risk AI systems placed on the Union market. The report must be made immediately after the provider has established a causal link between the AI system and the serious incident, or the reasonable likelihood of such a link. Separately, and in any event, the report is due not later than 15 days after the provider or, where applicable, the deployer becomes aware of the serious incident. That 15-day figure is an outer bound from awareness; it does not restart for a perfect root-cause memo. Faster lanes also exist: immediately, and not later than two days, for a widespread infringement or an Article 3(49)(b) serious incident; and, where a death is involved, immediately after establishing or suspecting a causal relationship, but not later than 10 days after awareness. Incomplete initial reports are allowed when speed matters.
Amended Article 113 in that consolidated text keeps the Regulation’s general application date of Aug. 2, 2026. The Omnibus delay is narrower: Chapter III, Sections 1, 2 and 3—with the exception of Article 6(5)—apply from Dec. 2, 2027 for systems classified as high-risk under Article 6(2) and Annex III, and from Aug. 2, 2028 for systems classified as high-risk under Article 6(1) and Annex I. Article 73 is not listed in those delayed Chapter III sections. Applicability also depends on high-risk classification and Article 111 transition rules for existing systems. Have counsel resolve the system-specific position; this is not legal advice.
California’s Frontier Critical Safety Incidents
California’s Transparency in Frontier Artificial Intelligence Act is already on the books. The Business and Professions Code chapter that begins at section 22757.10, added by SB 53 and effective Jan. 1, 2026, defines frontier models, catastrophic risk and critical safety incidents, then sets the clocks in section 22757.13. A frontier developer must report a critical safety incident for its frontier models to the Office of Emergency Services within 15 days of discovery. If that incident poses an imminent risk of death or serious physical injury, the developer must disclose it within 24 hours to an appropriate authority, including a law-enforcement or public-safety agency with jurisdiction, as required by law. Those are not interchangeable clocks. The 15-day report goes to OES. The 24-hour disclosure is the imminent-harm lane.
“Critical safety incident” is defined narrowly: unauthorized access to, modification of, or exfiltration of frontier-model weights that results in death or bodily injury; harm from materialization of a catastrophic risk; loss of control causing death or bodily injury; or deceptive techniques against the developer outside an evaluation context that demonstrate materially increased catastrophic risk. “Frontier model” means a foundation model trained with more than 10^26 integer or floating-point operations. Deploying a copilot does not by itself make an enterprise a frontier developer. Do not invent a California duty the statute does not cover.
October House Letters Were Requests, Not Law
On Oct. 27, 2026, House Energy and Commerce Ranking Member Frank Pallone Jr., Science Ranking Member Zoe Lofgren and Judiciary Ranking Member Jamie Raskin sent multi-committee letters to frontier companies. The OpenAI letter calls itself “an ongoing request” in the absence of mandatory, standardized reporting requirements. It asked for a full accounting by Nov. 11, 2026, 48-hour notice of new or previously unknown incidents, and 24-hour notice if critical infrastructure appears involved. That Nov. 11 date was a request deadline in the letter, not a statute. Ranking-member letters are oversight requests, not an enacted U.S. federal AI incident-reporting mandate.
Building AI Incident Reporting Governance Without Inventing Duties
The following is my opinion as an operator, not legal advice.
- Keep a system inventory with a regime tag. Label each production AI use as a high-risk candidate under EU rules, GPAI systemic-risk provider scope, California frontier developer scope, or none of the above. Wrong tags create false deadlines.
- Reuse your security incident spine, then fork the fields. Detection time, containment, customer notice and evidence holds still matter. Add AI-specific fields: model or system ID, evaluation versus production context, whether weights or prompts were involved, and which external clock might apply.
- Map clocks side by side, never averaged. Hypothetical: your on-call finds a runaway agent that touched a third-party network at 9:10 a.m. Eastern. That fact pattern may trigger an internal severity process today, a courtesy path under a dated congressional request if you are one of the addressed labs, California’s 24-hour lane only if it meets the statute’s critical-safety and imminent-harm tests, and an EU path only if you are in Article 55 or Article 73 scope. Average those clocks into “report in a week” and you will miss the short ones.
- Draft against published templates without pretending they are your only law. The Commission’s high-risk draft template and the GPAI systemic-risk template are free rehearsal scripts. Fill them once in a tabletop so legal, security and product argue while the room is calm.
- Write who owns awareness and who owns causal-link judgment. Article 73’s immediate duty attaches when a causal link or reasonable likelihood is established; the 15-day outer bound still runs from awareness. If nobody owns either determination, the short clocks become political.
Build the record before the next letter, OES notice or EU filing window arrives with your name on it. Regulators are not waiting for a blank page. Neither should your incident queue.

Leave a Reply