Skip to content
AX Framework
Trust pattern

Human Approval

Pause consequential actions until a person confirms, with a self-contained summary of what will happen.

AX proposalTrust & Control

Problem

Some actions are too consequential to delegate entirely, but asking for confirmation inside an agent's chat leaves no trustworthy record and is easy to spoof.

Context

Use this pattern when:

  • Actions involve money, legal commitments, data sharing, or irreversible changes.
  • The person may not be watching the agent in real time.

Recommended solution

When an action requires approval, return an approval request instead of executing. The request includes a plain-language summary, the exact parameters, the reason approval is needed, and an expiry. The person approves through a channel the product controls, not through the agent.

After approval, the agent resumes with the approval ID and the product executes exactly what was approved.

  1. 01Declare which actions require approval in the action contract.
  2. 02Return approval_required with a self-contained summary.
  3. 03Collect the decision through a product-controlled channel.
  4. 04Execute only the approved parameters; reject any changes.

Interaction diagram

Interaction diagram for Human Approval. 1. AI agent to Product · AX layer: place_order(cart). 2. Product · AX layer to AI agent: approval_required apr_903. 3. Product · AX layer to Person: Notification: approve 116.10 EUR?. 4. Person to Product · AX layer: Approve. 5. AI agent to Product · AX layer: place_order(approval: apr_903). 6. Product · AX layer to AI agent: completed + receipt.PersonAI agentProduct · AX layerplace_order(cart)1approval_required apr_9032Notification: approve 116.10 EUR?3Approve4place_order(approval: apr_903)5completed + receipt6
request response internal step

Example implementation

Illustrative example

Examples use a fictional store and illustrative field names. Adapt the shape to your stack; the qualities matter more than the exact format.

approval_requiredjson
{
  "status": "approval_required",
  "approval": {
    "id": "apr_903",
    "action": "place_order",
    "summary": "Order 1 × Trail Runner (Black, EU 42), deliver to Home, total 116.10 EUR.",
    "reason": "Purchases above 100 EUR require your confirmation.",
    "parametersHash": "sha256:8e1f…",
    "expiresAt": "2026-10-06T15:00:00Z"
  }
}

Accessibility and security

Accessibility

  • Approval notifications should be readable without context: what, how much, why, and what happens if ignored.
  • Provide approval through accessible channels (email, app, web) rather than a single modality.

Security

  • Hash the approved parameters and verify them at execution to prevent bait-and-switch.
  • Never accept approval asserted by the agent itself.
  • Expire approvals and make them single-use.

Applies the principles Human Oversight and Progressive Autonomy.