Runtime governance is the enforcement of AI policy at the moment a call is made, rather than the documentation of policy before deployment or the review of logs afterwards. The decision point sits in the request path: a prompt is evaluated, an action is permitted or refused, and the record of that decision is written as it happens.
How it differs from governance documentation
Most AI governance programs produce artifacts: a model registry, risk assessments, conformity documentation, a control mapping. These answer what the organization intends. Runtime governance answers what the organization did.
The gap between them is where regulatory exposure lives. A program with complete documentation and no enforcement produces an accurate description of controls that may not be operating, and no way to tell the difference until something goes wrong. The distinction is worked through on the Credo AI comparison, where documentation and enforcement are treated as complementary rather than competing.
How it differs from detection
Detection observes an action and raises an alert. Runtime governance intervenes before the action executes. On a model call that difference is the difference between knowing that customer data reached a third-party model and preventing it from doing so.
It also changes what the record is worth. A detection log says something happened. A runtime enforcement record says what was requested, which policy version was in force, what was decided and why — because the decision and the record are produced by the same operation. That is what makes examination evidence generable rather than assembled, and it depends on tamper-evident records underneath.
What it requires architecturally
Three properties, and a product either has them or it does not.
A position in the path. Governance that sits beside the traffic can describe it. Governance that sits in it can stop it. Every call an application or agent makes has to traverse the decision point, which is why runtime governance is a gateway problem — covered in the enterprise AI gateway guide and the AI control plane guide.
Policy as a versioned object. A decision is only defensible if the rule that produced it can be identified later, which means policy as code with a change history rather than a configuration screen.
An identity for every caller. Including the non-human ones. An enforcement record that names the person whose credential an agent borrowed is a record of the wrong actor — see agent authority scope and agent governance.
The fourth property is situational and decides the shortlist in regulated environments: whether the decision point can operate with no external dependency at request time. See air-gapped AI deployment and the Trust Fabric architecture.
Related terms
Policy as code · Agent authority scope · Air-gapped AI deployment · Tamper-evident audit log · Full glossary
Put this in the path of your own agents.
Policy enforced inline between your agents and every model and tool they reach, with a record bound to the human who owns it.
Request a Demo Read the docs