When you must use it

Call ASG after you have final AI output and immediately before you perform a real-world side effect.

Where it goes in your request path

AI output → validate (ASG) → PASS/WARN/BLOCK → execute or stop

Minimal request shape

You always send action_kind, action_type, and ai_output. context is optional at the schema level, but many ActionKinds will WARN or BLOCK if managed-policy required context evidence is missing.

json

{
  "action_kind": "money.refund",
  "action_type": "refund",
  "ai_output": {}
}

How to enforce the decision

Your integration must branch on the returned status.

Status	What you do
PASS	Execute the action.
WARN	Do not execute. Persist `decision_id`, route to a reviewer, and wait for approval.
BLOCK	Do not execute. Fix payload/context and re-validate.

Legal & Responsibility Notice

Summary

Informational only

Provided for general guidance. Not legal, compliance, security, or professional advice.

You control implementation

You are responsible for policies, prompts, integrations, workflows, and regulatory requirements.

Liability limitation

To the maximum extent permitted by law, the company disclaims liability for losses arising from use of this documentation or implementations based on it.