← Vetro Attest · Free Art. 12/14 checker · Free EU AI Act tools
EU AI Act Article 14: turning human oversight into evidence
*Engineering guidance, not legal advice.*
Article 14 requires high-risk AI systems to be designed so that natural persons can effectively oversee them while in use: understand the system's capacities and limits, monitor its operation, detect anomalies, decide not to use or to override an output, and interrupt the system through a stop procedure.
For a chatbot that answers one question at a time, a human reads every output. For an agent that runs twenty tool calls unattended, oversight has to be designed in, and then proven.
What oversight looks like for an agent
| Article 14 capability | Agent mechanism | Evidence it leaves |
|---|---|---|
| Monitor operation | Live view of each run's calls and costs | Per-event log with timestamps |
| Detect anomalies | Budgets, rate limits, forbidden-action rules | Policy decision events: warn / block, with reason |
| Override or disregard output | Approval step before sensitive actions | Approval request + human decision recorded in the run |
| Interrupt | Kill switch or hard budget cap | Run end event with the stop reason |
| Avoid automation bias | Show uncertainty, require confirmation | Record of what the human saw when they approved |
The evidence gap most teams have
Guardrails usually exist as code: if cost > limit: raise. When they fire, the exception goes to an error tracker; when they do not fire, nothing is recorded at all. After an incident you can show the rule exists, but not that it ran on this run, or what it decided.
The fix is to make every policy check an event in the same log as the actions, including the checks that passed. Then a reviewer can see, for any run: this action was checked against these rules, this one was blocked, this one waited for a human, this human approved it at this time.
How Vetro records it
Vetro's policy engine runs inline, before the action: budgets, forbidden tools, content rules, in warn or block mode. Every decision is written to the same signed, hash-chained ledger as the model and tool calls, so the oversight evidence carries the same integrity as the action it governed (see audit trail for AI agents).
Vetro Attest shows for free whether a log contains oversight evidence at all; the 29 USD counter-signed report lists it per run, next to the Article 12 record-keeping evidence, with an explicit list of what the log cannot demonstrate (training of overseers, organisational procedures).
See a sample ledger verified in one click, check your own logs free or get the countersigned EU AI Act report.