What Article 12 Actually Requires From Your Logs
Article 12 of the EU AI Act asks for automatic recording of events over the lifetime of a high risk system. Here is what that means in practice, and why most application logs do not satisfy it.
Most teams read Article 12 of the EU AI Act, see the word "logs", and assume their existing observability stack already covers it. It usually does not. The gap is not volume of data. The gap is evidentiary quality.
What the text asks for
Article 12 requires that high risk AI systems technically allow for the automatic recording of events over the lifetime of the system, to a standard that supports traceability of functioning. Two words in that sentence carry most of the weight.
Automatic. The record cannot depend on a developer remembering to instrument a call path. If an agent can take an action without producing a record, the record set is incomplete by design.
Lifetime. Retention is not a rolling seven day window sized to your logging bill. A regulator asking about an incident from fourteen months ago expects the record to exist.
Why application logs fall short
A typical agent log line records what the application decided to write down after the fact. Three problems follow from that:
- 01The log is written by the same process that took the action, so it can be skipped, suppressed, or rewritten by that process.
- 02The log records an outcome, not the policy that was in force at the time.
- 03The log is mutable. Anyone with write access to the log store can alter history without leaving a trace.
A record that the operator can silently edit is not evidence. It is a claim.
What closes the gap
The record has to be produced at the enforcement point, not by the agent, and it has to be tamper evident. In Gateplex, every intercepted action is written to an append only record containing the input, the matched rule, the verdict, the agent identity, and the timestamp. Each record stores its own SHA-256 hash plus the hash of the record before it. Modify or remove a row and the chain breaks at a verifiable point.
That gives you three things a log file cannot:
- Proof that a policy was in force at the moment of the action
- Proof that the record set is complete and unaltered
- An export a reviewer can verify independently
Where to start
Pick the single agent with the widest blast radius and route it through an enforcement point first. You do not need full coverage on day one. You need one path where the record is produced by something other than the system being audited.
Read the mapping in detail on the Compliance Frameworks page, or see the EU AI Act timeline for what applies when.