eu regulation
EU AI Act
The EU AI Act puts four operational duties on organizations running high-risk AI systems: risk management, automatic event logging, human oversight, and record retention. This page maps each duty, by article, to what a runtime decision record can answer when the system under review is an agent rather than a chat window.
The four duties that matter for agents
The Act's high-risk provisions concentrate on four operational duties. Article 9 requires a risk management system that operates continuously rather than an assessment filed once at deployment. Article 12 requires automatic event logging across the system's life. Article 14 requires human oversight designed into the system, and Article 19 requires the logs to be kept, with a 6-month floor on retention.
For a chat interface those duties are mostly about model output. For an agent they are about actions: which tool calls ran, under what policy, and who was in the loop when it mattered. Whether a given system falls in scope is a classification question for your counsel. The operational duties have the same shape either way.
Logging the system produces by running
Article 12 asks for records generated automatically by the system itself. A runtime control point produces exactly that: every evaluated action emits a decision record carrying the actor, the resource, the matched rule, and the outcome, whether that outcome was permit, deny, redaction, or escalation to a human.
Because the record is written at decision time, there is no gap between what the system did and what the log claims. Each record is cryptographically signed, so an examiner can check that it was not edited or backdated afterward.
Human oversight as an enforced step
Article 14 expects a human who can intervene before an action lands, which a review queue reading yesterday's transcripts cannot do. Approval routing implements the duty literally: a rule can require a named human to release a sensitive action, and the agent waits until that happens.
The approval itself becomes evidence. Each human response is recorded as its own signed event, tied to the action it governed, so the oversight duty and the record-keeping duty answer each other.
Record-keeping starts with knowing what you run
Article 19's retention duty presumes you know which systems exist. Agents arrive the way shadow SaaS did: a framework wired into a service, an assistant with tool access, none of it waiting for review. Discovery on endpoints finds the frameworks, MCP servers, and models actually present and keeps the inventory current.
Retention itself is configurable past the 6-month floor the Act sets. What matters is that the retained records are the runtime originals rather than reconstructions assembled once an inquiry arrives.