The Incident Took Ninety Seconds. The Report Is Due in Two Days.

Your incident response process probably ends with a postmortem. A thoughtful document, written by calm people, two or three weeks after the fact.

The EU AI Act's Article 73 would like a word. Specifically, it would like that word within two days.

The reporting regime is no longer hypothetical

As of August 2, 2026, providers of high-risk AI systems in the EU are under a live obligation to report serious incidents to national market surveillance authorities. The European Commission spent late 2025 consulting on draft guidance and a reporting template for exactly how this works. The consultation is over. The clocks are running.

Article 73 defines the deadlines by severity, and they are not generous:

• 15 days after discovery for a standard serious incident
• 10 days if the incident may have caused a death
• 2 days for a widespread infringement of fundamental rights or serious disruption to critical infrastructure

Miss the window, or fumble the report, and the penalty exposure runs up to €15 million or 3% of global annual turnover. For most companies reading this, that's not a fine. That's a funding round.

"Serious incident" is broader than you're hoping

The definition under Article 3(49) covers an AI system malfunction that directly or indirectly causes death or serious harm to health, serious and irreversible disruption of critical infrastructure, infringement of fundamental rights, or substantial damage to property or the environment.

The Commission's draft guidance interprets "indirectly" the way a regulator interprets everything: broadly. Their own example — an AI system produces an incorrect medical analysis, a clinician acts on it, a patient is harmed — is a reportable incident, even though a human made the final call. The human in the loop is not a liability shield. (Regular readers will remember that the human in the loop wasn't a security control either.)

Now put agents in that picture. An agent that takes actions — calls tools, moves data, touches infrastructure — has far more paths to "indirectly caused" than a model that just emits text. Every tool call is a candidate link in a causal chain that ends at a regulator's desk.

The clock starts at discovery. Discovery requires visibility.

Here's the uncomfortable arithmetic. DigiCert's July 2026 survey of 1,001 IT and security leaders found that 50% of enterprises experienced a security incident tied to unauthorized or misconfigured AI agents within six months — and roughly half of respondents reported limited or no visibility into how their AI systems reach decisions.

Read those two findings together. Half of enterprises are having agent incidents. Half can't reconstruct what their systems did. The overlap of those populations is now subject to a two-day reporting deadline for the worst category of incident.

A serious-incident report isn't a tweet. The template asks what happened, when, which system, what harm resulted, and what you've done about it. Answering those questions in 48 hours is not a writing problem. It's an evidence problem. If your agent activity lives in application logs that rotate weekly, scattered across services, mutable by anyone with shell access — you don't have evidence. You have vibes with timestamps.

And there's a detail in the draft guidance that should stop you mid-scroll: once an incident is under investigation, you're expected not to alter the AI system without notifying the authority first. So your evidence trail needs to prove not just what the agent did, but that nobody quietly tidied up afterward. Mutable logs can't prove a negative. Immutable ones can.

Reporting readiness is an infrastructure property

Legal will own the filing. But legal can only file what engineering can produce. Article 73 readiness decomposes into three capabilities, and all three have to exist before the incident:

You have to detect the incident. The 15-day clock — or the 2-day one — starts at discovery, and "we discovered it when a journalist called" is the worst possible opening line. VeriSwarm Gate ingests every agent event against a standardized taxonomy and scores behavior continuously, so anomalies surface as trust-score drops and policy-tier changes in real time — not in next quarter's audit.

You have to reconstruct what happened. Vault writes every agent action into an immutable, hash-chained audit ledger. When the report template asks for the sequence of events, you export the chain: what the agent did, when, in what order, cryptographically linked so that gaps and tampering are detectable. That also answers the don't-alter-the-system problem — chain verification demonstrates your record is intact, which is precisely the posture you want when a market surveillance authority starts asking questions. We've written before about the difference between cryptographic and declarative evidence; Article 73 is where that difference stops being philosophical.

You have to stop the bleeding without destroying the record. Guard's kill switch halts a misbehaving agent immediately — containment that doesn't touch the evidence, because the record of everything up to the halt is already sealed in Vault.

The two-day drill

Here's a test you can run this week, no regulators required. Pick one production agent. Pretend it just caused a reportable incident at 4pm on a Friday. Can you produce, by Sunday, a complete and trustworthy account of everything it did in the past 30 days?

If yes — you're ahead of most of the enterprises in that DigiCert survey, and Article 73 is paperwork.

If no — the gap between your answer and "yes" is exactly the infrastructure Article 73 quietly assumes you already have. Gate is free and always on. Start there, and the next time a clock starts ticking, it'll be ticking for someone else.