Approval workflow
Approval workflows are a practical way to keep AI actions accountable without blocking safe automation. A policy defines what can run, what must be reviewed, and what must be blocked. When risk is detected, the request pauses until a human decision is recorded.
Why approvals exist
B2B environments require clear ownership for high-risk decisions. Approval workflows keep that ownership explicit by making a human decision part of the control path. This reduces the chance that a risky action is executed automatically and provides evidence that a decision was reviewed.
Approvals are especially useful during pilots, when teams are still defining policy boundaries and data handling rules. The workflow creates a safe feedback loop without halting all automation.
Decision outcomes
- allow - the action can proceed.
- deny - the action is blocked.
- approval_required - the action waits for a human decision.
These outcomes are intentionally simple. They are easy to audit and easy to explain during security reviews.
What an approver should see
Approval decisions should be based on a short, review-friendly summary. That summary typically includes the request id, policy reference, risk trigger, and a sanitized description of the intended action. The goal is to support a defensible decision without exposing unnecessary raw content.
The exact summary format is defined in the pilot scope. Some teams prefer a short text summary, others prefer a structured checklist. Both are valid as long as the decision context is preserved.
Lifecycle of an approval
- Request enters the gateway and is evaluated against policy.
- Policy marks the request as approval_required.
- Operator receives a review summary with decision context.
- Operator approves or denies the action.
- Decision is recorded with timestamps and policy reference.
Some pilots also define time limits for approvals or escalation rules. Those controls should be documented in the pilot scope if they are required.
Approval evidence fields
A minimal approval record should include the request id, policy reference, decision outcome, operator identity, and a timestamp. This keeps the trail clear for internal reviews and external audit conversations without exposing unnecessary raw content.
If a pilot needs stronger traceability, add a short decision note or a structured reason code. Keep it short and focused on the policy trigger rather than the full prompt or output.
Escalation and expiry
Some teams require approvals to expire after a defined time window. Others need an escalation path if no decision is recorded. These controls are optional and should be defined in pilot scope so operator load stays predictable.
Operational notes
Approvals work best when the policy rules are clear and the operator workload is predictable. If too many requests are marked as high risk, the system will slow down and approvals will become a bottleneck. The right balance is to keep approval for true risk, not for routine operations.
This is why pilots are scope-disciplined. A single use case with defined risk triggers provides enough signal to refine the policy without overwhelming the operator.
FAQ
What triggers approval?
Approval triggers are defined by policy. Examples include sensitive data, external actions, or exceptions to normal scope. The exact triggers are defined in the pilot scope.
Who can approve?
Approval authority is defined by the operator team. The system records which operator approved or denied the request, along with timestamps.
Do approvals slow everything down?
Approvals only apply to requests marked as high risk. Safe requests continue to run automatically. The intent is to keep safe paths fast while providing control for risky actions.
Is raw content required for approval?
Not necessarily. A sanitized summary is often sufficient for a decision. Raw content handling depends on pilot scope and data handling requirements.
Can approvals expire?
Expiry windows can be defined if needed, but they are not mandatory. If expiry is required, document it in the pilot scope so operators know how long a decision remains valid.