Article 14 of the EU AI Act requires that a human overseer of a high-risk AI system can do five things: understand the system's capacities and limitations, remain aware of automation bias, correctly interpret its output, decide not to use or disregard output, and intervene or halt the system when required.
Most guardian agent specifications do not cover all five. Many cover none of them explicitly.
A guardian agent enforcing a governance document inherits that document's quality ceiling. If the specification does not tell the guardian what correct output looks like, the guardian cannot correctly interpret output. If it does not name the conditions under which the system should be halted, the guardian cannot enforce a halt. The guardian is not the responsible party. It is the mechanism the responsible party specified. If the specification is incomplete, the mechanism enforces incompleteness.
An empirical study of AI governance documents found that 37% fell below the structural quality threshold: unverified claims, overgeneralised scope, missing assessment criteria. Those documents cannot support Article 14 oversight requirements because the specification is incomplete. A document that passes the threshold is fit to support compliance, but that requires qualified personnel and execution in the field as well. Without the structural foundation, neither of those can stand.
Three conditions have to be true before a guardian agent can work.
The specification must be verifiable. It has to make claims the guardian can actually test. Unverified assertions about system behaviour, overgeneralised scope statements, and missing assessment criteria are not enforceable. A guardian enforcing them is not providing oversight. It is providing confidence in the absence of oversight.
The specification must define the boundary precisely. Not every decision requires a human. Deterministic, rule-clear situations can be handled autonomously. The specification's job is to name which situations are in that category and which require human judgment. The human lives at the boundary, not inside every loop. In high-uncertainty or high-consequence environments that boundary may not be definable, and runtime human presence is the only defensible position.
The specification must carry its own staleness conditions. A governance document that does not specify when it becomes invalid will be used past its validity. Every policy should define conditions under which it requires re-evaluation: a model update above a certain capability threshold, a new regulatory instrument, a change in the production environment.
The guardian agent problem is solvable. The specification is the proof that the Article 14 requirements have been addressed at authorship time. It is also the audit trail when enforcement asks.
Most organisations are not starting there.