An AI policy has to run where the model runs
21 November 2026
Almost every large organisation now has an AI policy. It has been through legal, endorsed by the risk committee and published on the intranet. And in most cases it governs nothing that happens while the systems are running.
That sounds harsh, but the test is simple. Take any statement in the policy, for example that personal information must not be sent to an external model provider, and ask what, specifically, stops it happening on the next request. If the answer is staff training and a design review held months ago, the policy is a statement of intent. If the answer is a control that inspects every request and blocks or redirects the ones that breach it, the policy is enforced. Most organisations have the first and believe they have the second. They find out the difference during an incident, or when an auditor asks for the evidence.
Enforcement at runtime is not exotic. It needs four working parts. The rules have to be written in a form a system can apply, not only in prose: which data classes may go where, which actions need a human, which users may invoke which capabilities. Every request has to pass through a control point that applies those rules, typically a gateway in front of the models rather than logic scattered through each application. Every decision that control point makes has to be recorded and kept. And there has to be an exception path, with a named approver, for the cases where the rule should be set aside, so that exceptions are visible decisions rather than quiet workarounds. If any of the four is missing, what you have is an aspiration.
The record matters as much as the block. For each live AI service, the organisation should be able to answer three questions within minutes rather than weeks: what did the system do, under which policy, and who approved any exceptions. That means capturing the request, the response and any approval event, linking each to the rule that governed it, and retaining the record for as long as your obligations require. This is the baseline, not the gold standard. A service that cannot answer those questions is running without a control plane, and the absence of incidents so far is luck rather than design.
None of this makes design-time review redundant. Reviews before deployment catch the obvious problems: the wrong data source, an action that should never be automated, a missing human check. What they cannot catch is what changes afterwards. A prompt that behaved well last month behaves differently after a model update. A retrieval index picks up documents nobody intended it to see. A user discovers that a series of small, innocuous queries adds up to something that should not leave the building. Design review prevents. Runtime governance detects and responds. A program needs both, and most have only the first.
There is a cost argument as well as a risk one. Organisations without runtime enforcement compensate with process: more review gates, longer approval chains, slower releases. That is expensive and it still leaves the gap open. Moving the control into the system usually lets the process get lighter, because the reviewers can rely on the rule being applied every time rather than hoping it is.
For boards, the useful question is whether management can show, for the most material AI system, the record of a request being stopped by the policy, and the record of an exception being approved. If both exist, governance is real. If neither does, the policy is a document.
A practical place to start: choose one live AI service and route it through a single control point. Turn on a small set of rules, the two or three that matter most, and record every decision. Run it for a month. You will learn more about your actual risk in that month than from another round of policy drafting, and you will have the pattern to extend to the next service.