The architecture question that comes first
Most AI governance conversations begin with the model: which model produced the output, how it was trained, and whether its quality was evaluated. Those questions matter, but they do not reveal where operational risk becomes real.
A recommendation becomes business risk when another system acts on it. A payment is declined. A case is escalated. An agent invokes an enterprise tool. A record is sent to a protected destination. That transition is the governed execution boundary.
What belongs at the execution boundary
At this boundary, the organization can evaluate more than model output. It can check the policy version, caller authority, route, data handling rules, cost ceiling, and downstream destination before allowing execution.
- Policy: Is this action allowed under the current rule set?
- Authority: Is this caller or pipeline permitted to take it?
- Control: Should the action proceed, be modified, be routed for review, or be denied?
- Execution: What actually reached the protected system?
- Evidence: Can the organization later reconstruct the decision and disposition?
The surrounding technology can change
One company may use Kafka between systems. Another may use REST APIs, a payment switch, an MQ broker, an agent framework, or a vendor platform. The transport is an integration choice, not the definition of the control point.
This is why the boundary is a more durable architecture concept than any one transport. Upstream systems can evaluate, score, recommend, or decide. Downstream systems can commit, notify, transfer, update, or invoke. The need to govern the transition remains.
How to find your first governed decision
Choose one decision that is consequential, repeated, and difficult to reconstruct after the fact. Trace it from the system that produces the recommendation to the system that takes action. Then identify the last safe point at which policy and authority can still change the outcome.
That gives a pilot a concrete scope: one decision lane, a small set of enforceable controls, a protected action, and evidence that proves what happened. It avoids turning governance into a broad platform migration.