← Back to Learn
approval-workflowsexplaineragent-safety

What are approval workflows for AI agents?

Authensor

An approval workflow is a mechanism that pauses an AI agent when it tries to take a high-risk action and routes the decision to a human reviewer. The action does not execute until the reviewer approves it. If denied, the action is blocked and the agent is notified.

Why approval workflows exist

Some actions are too consequential for full automation. Sending money to a vendor, deleting production data, deploying code, or contacting a customer are all actions where a wrong decision has real costs. Approval workflows let you automate the routine while keeping humans in the loop for the decisions that matter.

How they work

  1. The agent requests an action (e.g., "send $5,000 to vendor")
  2. The policy engine evaluates the action and returns escalate
  3. The system creates an approval request with the action details
  4. The request is routed to the designated reviewer (Slack, email, API)
  5. The reviewer approves or denies the request
  6. If approved, the action executes
  7. If denied, the agent is told the action was rejected
  8. A receipt records the entire flow

Defining escalation rules

Escalation is triggered by policy rules:

rules:
  - tool: "payment.send"
    action: escalate
    when:
      args.amount:
        gt: 100
    reason: "Payments over $100 require approval"

The reason field is shown to the reviewer so they understand why the action was flagged.

Single vs multi-party approval

For most actions, one reviewer is enough. For high-stakes actions, require multiple independent approvals:

  • Single approval: One person reviews and decides
  • Multi-party approval: Two or more people must independently approve
  • Role-based approval: Only reviewers with specific roles can approve

Multi-party approval prevents a single compromised or inattentive reviewer from greenlighting a dangerous action.

Timeouts

Every approval request needs a timeout. If no reviewer responds within the configured window, the action should be denied by default. An action that nobody reviews should not proceed. This is the fail-closed principle applied to approval workflows.

Approval channels

Approval requests can be delivered through:

  • Slack messages with approve/deny buttons
  • Email with one-click approval links
  • API endpoints for custom dashboards
  • SMS for critical, time-sensitive approvals

The channel should match the urgency and the reviewer's workflow.

The agent's perspective

From the agent's perspective, an escalated action simply takes longer. The agent receives a response saying the action is pending. Depending on your configuration, the agent can continue with other tasks while waiting, or it can pause until the decision arrives.

Well-designed agents handle denials gracefully, informing the user that the requested action was not approved and offering alternatives.

Keep learning

Explore more guides on AI agent safety, prompt injection, and building secure systems.

View All Guides