AI governance, and the part that has to run in real time

AI governance is the set of policies, processes, and controls an organization uses to make sure its use of artificial intelligence is lawful, safe, aligned with its intent, and accountable to named people. It covers model selection and evaluation, data handling, fairness and bias review, transparency, third-party risk, and, increasingly, the behavior of autonomous agents acting on the organization's behalf.

Frameworks such as the NIST AI Risk Management Framework, ISO/IEC 42001, and the EU AI Act give the program its structure. Most of what they ask for is documentary: inventories, impact assessments, policies, review cycles, evidence that controls exist.

Where documentary governance runs out

Documentary governance was built for a world where AI produced outputs and humans acted on them. A model recommends; a person decides. The governance question was whether the recommendation was fair, explainable, and well-tested.

Agents change the shape of the problem. An agent does not recommend a refund. It issues one. It does not suggest a purchase order. It approves it. The moment of governance moves from the review meeting to the millisecond before the action executes, and a policy document cannot be in that path.

This is the part of AI governance that has to become operational. Our position is that any AI governance program that includes agents needs a runtime component that answers, for each consequential action, whether the action was authorized, within what limits, and on whose accountability.

What the operational layer looks like

Three things have to exist, in our view.

A definition of authority that humans own. For each class of decision an agent may make, a written statement of the decision rights, the constraints, and the accountable person. This is the bridge between the governance committee and the running system: the policy the committee approves is the policy the agent is checked against.

Enforcement in the action path. The check happens before the action, at the tool call, API tier, gateway, or sidecar, and resolves to allow, constrain, escalate to a human, or deny. It applies to every agent that tries to act, including the ones nobody registered.

A per-decision record. Not an application log, but a record of what was decided, under which policy, against which cumulative ceiling, with what outcome and which human escalation if any. This is the evidence an auditor, regulator, or board asks for when they ask whether AI decisions are controlled.

How this maps to the frameworks

The frameworks ask for governance functions like "govern," "map," "measure," and "manage," or for management system controls with evidence of operation. The runtime layer described above is how "manage" becomes true for agents rather than asserted. The Accountability Ledger is the evidence of operation. The Authority Model is the documented control. The enforcement point is the control operating.

That is a mapping, not a certification claim. Frameworks vary in what they require and how they are assessed, and the right mapping for a given organization depends on its regulatory position.

A concrete case

A bank's AI governance committee approves a policy: the collections agent may offer payment plans within a defined range, may not waive fees above a threshold, and must route any hardship claim to a human. The policy is documented, reviewed, and filed. Six months later an auditor asks how many times the agent waived a fee above the threshold. If the policy lived only in the document, nobody can answer, because the agent was checked against its API scopes and the scopes permit fee waivers. If the policy was enforced at runtime, the answer is in the ledger: the count, each case, and the human who approved the exceptions.

Where BotAris fits

BotAris is the runtime authority layer for an AI governance program. It does not replace the program. It gives the program a way to enforce the decisions it has already made about what agents may do, and to produce the evidence that they were enforced.