AWS adds Smart Approvals to Amazon Quick for AI agent actions

The feature sits on top of controls Amazon Quick already provides: Custom Permissions, Quick Guardrails, connector action permissions, and user settings. An action becomes eligible for automated review only when the user's setting is "Ask each time" and no other control requires human approval. Denied actions stay denied, and mandatory human approvals still apply, so Smart Approvals is framed as reducing interruptions rather than loosening the permission boundary.

The review itself is handled by a separate classifier model that weighs what an action does, where data goes, and who it affects. It can consult the user's knowledge graph — people and working relationships drawn from connected sources — to judge whether a recipient sits in the user's reporting chain, and that lookup stays inside the Quick account. AWS says the content the classifier reviews is not used to train models or improve the service. Reviewing actions against the user's request also adds a check against indirect prompt injection, where instructions embedded in retrieved content try to redirect the agent; Quick tells both the agent and the classifier to treat such content as information, not instructions. AWS nonetheless cautions that model review can make mistakes, so actions that must always get human approval should be enforced through connector action permissions.

Enforcement is kept separate from judgment: each automatic approval is single-use, applies only to the exact action and parameters reviewed, and grants no ongoing permission. The permission service records associated events in AWS CloudTrail and permission-check verdicts, and chat and agent metadata logs delivered through Amazon CloudWatch Logs let investigators correlate activity by user and time. AWS notes that an approval marker does not prove the action ran — records from the destination system, such as a sent email, are needed to confirm completion.