What Article 12 requires, and who carries it
Article 12 of Regulation (EU) 2024/1689 states that high-risk AI systems shall technically allow for the automatic recording of events (logs) over the lifetime of the system. Two points matter for implementation. The duty is framed as a system capability, so evidence should come from automated capture rather than manual reconstruction. And the level of traceability must be appropriate to the intended purpose of the system, which leaves the level of detail to context.
Article 12(2) names three purposes for the recorded events. They must help identify situations that may result in a risk within the meaning of Article 79(1) or in a substantial modification. They must facilitate post-market monitoring under Article 72. And they must support monitoring of system operation under Article 26(5). Treat these as test criteria: for every event you record at a tool call, you should be able to say which purpose it serves.
Responsibilities differ by role. Article 19 requires providers of high-risk AI systems to keep the logs referred to in Article 12(1) that their systems generate automatically, to the extent those logs are under their control. Article 12(2)(c) ties logging to deployer monitoring. The regulation also refers to mechanisms that allow deployers to collect, store and interpret logs. A deployer running agents in its own environment should therefore establish early who controls the logs and in what form they can be read. As of October 2026, the Commission describes the Act's provisions as applying progressively.
Mapping the Article 12(2) purposes to tool-call events
For risk identification, the events that matter are those that show the agent attempting something outside its intended purpose: calls to tools it should not reach, repeated blocked attempts, and unusual parameter patterns. Blocked and escalated calls are often more informative than successful ones, so a record that captures only completed actions is incomplete.
For substantial modification, the useful signal is change. Recording the policy version, the tool inventory and the agent configuration in force at each call lets a reviewer see when behaviour changed and whether the change coincided with a new tool, a new policy or a new model. Article 12 does not prescribe this, but the traceability purpose is hard to meet without it.
For post-market monitoring and deployer monitoring, records must be queryable by agent, tool, time window and outcome, and exportable in a form the deployer can interpret. Fragmented application logs across agents, models and tools usually fail here, not because data is missing but because it cannot be joined into one reviewable record of a call.
Recording after the fact versus enforcing in the execution path
There are two common architectures. In the first, each agent, model and tool writes its own application log, and a team assembles the story later. In the second, policy is enforced at the point of the tool call, so the decision and the outcome are written as the call happens.
The first approach can work for debugging, but it has three weaknesses as audit evidence. Completeness is hard to demonstrate, because a call that bypassed a logging path leaves no trace. The policy that applied is often inferred rather than recorded. And records from different systems rarely share identifiers or clocks. The second approach records the decision as part of the control itself, so the absence of a record is itself a signal. No source tests this distinction, so treat it as an architectural judgement to evaluate against your own systems.
A useful review question is whether a tool call could succeed without producing a record. If the answer is yes, the logging capability is not automatic in the sense Article 12 describes.
Limits of a checklist, and where enforcement evidence fits
A checklist can show technical alignment with the text of Article 12. It cannot establish legal compliance, which depends on risk classification, intended purpose and the regulation as a whole. Start by determining which of your systems are high-risk and whether any fall under Annex III point 1(a). Then separate provider and deployer duties before assigning tasks. Track the progressive application timeline and any amendment on official pages, because the status of obligations can change.
Trussed enforces policy in the execution path across agents, tools and MCP servers: allow, redact, block, rate limit, or escalate for human approval. Enforcement at each interaction and tool call generates audit-ready evidence mapped to the EU AI Act. Trussed is deployed self-managed in the customer cloud, with configurable retention and data residency. It addresses the tool-call level described above. It does not provide end-to-end tracing of agentic workflows or transaction-level governance, and it does not replace a legal assessment of your obligations.