Common policy enforcement gaps when policy stays a PDF

The gaps that appear when policy stays a PDF

A written policy feels like coverage. In an agent system, it is not. The document describes what should happen; the runtime has no way to enforce it. The failures that follow are predictable, and none of them look like a "hack."

Gap 1: silent overrides

An agent is told to "process the refund" and does — including the one above the manager-approval threshold. Nothing blocked it, because the threshold existed only in prose. The override was silent; the customer got the money back; the audit trail shows a normal transaction.

Gap 2: no evidence trail

EU AI Act Art. 12 requires logging of high-risk system activity. A PDF policy produces no logs. When a regulator or your own DPO asks "did the agent ever violate the data-residency rule?", a document cannot answer. Only an enforced, logged rule can.

Gap 3: drift after a prompt change

Someone edits the agent prompt to "be more helpful with refunds." The policy did not change — but the agent's behavior did. Without a rules file the agent evaluates on every run, the policy and the behavior quietly diverge, and nobody notices until an incident.

Closing the gaps

Encode each policy as a condition + action rule, wire high-risk rules to human approval or a block, and log every decision. Re-run the checklist after any prompt or tool change. The rules are a view of your obligations, not a guarantee — but they turn an invisible policy into an enforceable, auditable control.

Authoritative references

EU AI Act (Reg. 2024/1689), Art. 12 (logging): https://eur-lex.europa.eu/eli/reg/2024/1689/oj
AI Act explorer: https://artificialintelligenceact.eu/
European Commission AI policy: https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai

FAQ

What is the most common policy enforcement gap in agent systems?

Silent overrides — an agent takes a high-risk action (like a refund above threshold) because the policy existed only in prose and nothing in the runtime enforced it.

Does the EU AI Act require logging of agent actions?

For systems in an Annex III high-risk use, Art. 12 requires logging of high-risk system activity. An enforced, logged rule is what lets you demonstrate that requirement was met.

Can a PDF policy alone satisfy EU AI Act obligations?

No. The Act requires actual controls (risk management under Art. 9, human oversight under Art. 14, logging under Art. 12). A document describes intent; only an enforced rule provides the control and the evidence.

How does prompt drift break policy?

Editing an agent prompt changes behavior without changing the written policy. Without a rules file the agent evaluates every run, the two silently diverge until an incident exposes the gap.

Is AgentPolicy a guarantee against policy violations?

No. AgentPolicy helps you extract and enforce rules and see where they map to obligations. It is decision-support, not a legal opinion, a certification, or a guarantee of conformity.

## Disclaimer

Disclaimer (decision-support, not legal advice): AgentPolicy is a tool that helps teams turn policy text into agent-callable checks and a compliance checklist. It is not a legal opinion, a compliance certificate, or a guarantee that an agent is safe or compliant. EU AI Act conformity depends on the full deployed system and is the responsibility of the deploying organization. Verify current obligations and effective dates with qualified counsel before relying on any timeline.

Try AgentPolicy to find and close the gaps between your written policy and what your agent actually enforces — try AgentPolicy.