AI Operational Governance: The Day-Two Control Loop That Runs on Every Request

Google Mandiant's M-Trends 2026 report put the median attacker handoff time at 22 seconds in 2025, down from over eight hours in 2022. A governance program that meets in a quarterly committee cannot operate at that tempo. The decisions that matter for AI risk happen on individual requests, thousands of times a day, and the committee is asleep for almost all of them. Design-time governance writes the policy. Operational governance is what runs it while everyone is looking elsewhere.

I want to walk through what AI operational governance covers, because it is the layer most programs skip on their way from a strategy document to a compliance audit.

Design-time and day-two

A governance program has two lives. The design-time life produces the strategy, the operating model, and the policies. That work is necessary and it is finite. The day-two life is where those artifacts either run continuously on live traffic or sit in a repository while AI requests flow past them unchecked.

Operational governance is the day-two life. It is the control loop that executes on every AI request: evaluate the request against current policy, decide, record the decision, and feed the record back into review. The loop runs whether or not anyone is watching, which is the property that separates governance that operates from governance that merely exists.

What the day-two loop actually manages

Four operational concerns live in this loop, and each one degrades quietly when no runtime layer owns it.

Policy versioning. Policies change as models, routes, and regulations change. Operational governance tracks which policy version was in effect when a given decision was made, so a decision from June can be judged against June's rules rather than today's. A record without a policy version cannot be audited against the policy that governed it.

Exception expiry. Exceptions granted during design-time reviews accumulate. Netwrix found that 97% of organizations breached through AI lacked proper access controls, and expired-but-still-active exceptions are one way that gap opens. Operational governance gives every exception an owner and an expiry and closes it when the clock runs out.

Drift detection. In a federated organization, business units interpret shared policy differently over time. The decision records make drift visible as divergence across units, which turns a governance abstraction into a query you can run.

On-call for AI policy. When a policy blocks legitimate work or misses a real violation, someone has to adjust it in production, not at the next quarterly meeting. Operational governance names that owner and gives them the telemetry to act.

These four are day-two operations, and none of them survive if governance is treated as a document that ships once.

Why the loop lives at the request layer

The control loop can only run where the requests are. A loop that samples logs after the fact reviews decisions it could not influence. A loop that sits inline at the AI request layer evaluates, decides, and records in the same pass, which is the only place the 22-second attack window can be met. This is the operational face of inline enforcement: the same architecture that contains an incident is the one that runs governance day to day.

The loop's output is also the input to everything above it. The records it produces feed the governance metrics and the oversight package, so the board reviews the same telemetry the loop generated. Design-time governance sets the rules. Operational governance is where the rules meet traffic.

DeepInspect

DeepInspect runs the day-two loop. It sits at the AI request boundary as a stateless proxy between authenticated users or agents and any LLM, evaluating each request against identity-bound policy before the model receives it.

The loop executes on every request: evaluate against the current policy version, decide, and commit a signed record carrying identity, role, policy version, data classification, outcome, and timestamp. Policy versions are tracked so a decision is always auditable against the rules that governed it. Exceptions are scoped rules with owners and expiries rather than standing grants. The operational concerns that decay in a document-based program become properties of a control that runs on every call.

If your AI governance produced a strategy and an operating model but has no loop running on live requests, the day-two layer is missing. Let's talk today.