AWS adds Smart Approvals to reduce routine prompts in Amazon Quick

AWS says Smart Approvals in Amazon Quick can automatically approve eligible agent actions while returning consequential decisions to users for review. The feature is designed to reduce repeated prompts without expanding permissions or removing required human oversight.
- Smart Approvals reviews eligible Amazon Quick actions against user requests and existing controls.
- Mandatory human approvals and policy denials remain in place.
- Automatic approvals apply only once, to the exact reviewed action and parameters.
- AWS recommends piloting the feature and auditing approvals through CloudTrail and CloudWatch Logs.
Approval remains bounded by existing controls
Smart Approvals reviews proposed actions against the user’s request and the controls already configured in Amazon Quick. These include Custom Permissions, Quick Guardrails, connector action permissions and user settings. Denied actions remain blocked, while connector permissions can require a person’s approval for specific actions.
An action is eligible for automated review only when the user’s setting is “Ask each time” and no other control requires human approval. Administrators use Custom Permissions to determine who can use Smart Approvals. The feature is on by default, although profiles with deny-by-default enabled must explicitly allow it. Users can also switch it off for themselves across conversations and scheduled runs.
A separate classifier evaluates the request and context
Amazon Quick uses a separate model, called the classifier, to assess eligible actions against the user’s request and a bounded view of the conversation. It considers what an action does, where data goes and who it affects. The classifier can also consult the user’s knowledge graph, including people and working relationships drawn from connected sources. AWS says this lookup stays within the Quick account, and that content reviewed by the classifier is not used to train models or improve the service.
Actions may still be sent back to the user when they involve sending content outside the organization or to senior leaders in the user’s reporting chain, changing access, deleting records or initiating payments. The classifier also checks for situations where retrieved content attempts to redirect the agent through indirect prompt injection. AWS says Quick instructs both the agent and classifier to treat that content as information, not instructions.
Each automatic approval is single-use and applies only to the exact action and parameters reviewed. It grants no ongoing permission, and Quick labels automatic approvals in the conversation. Timeouts and other review failures return the action to the user rather than granting approval. AWS cautions that model review can make mistakes, so actions that must always receive human approval should be controlled through connector action permissions.
AWS recommends piloting the feature with audit logging
AWS gives an example in which an account manager asks Quick to send a brief to two named colleagues. If an account note also instructs Quick to send it to an all-sales distribution list of about 2,000 people, Smart Approvals is designed to return the wider send for review because the audience exceeds the user’s request. Approval for the original recipients cannot be reused for that wider audience.
Permission updates associated with Smart Approvals are recorded in AWS CloudTrail, with the user, connector, action and time identified. Quick also records permission-check verdicts and changes to a user’s Smart Approvals setting. AWS recommends correlating these records with Quick activity, Amazon CloudWatch Logs, chat logs, agent metadata and records from the destination system. CloudTrail Event history retains events for 90 days, so AWS advises configuring a trail for longer investigations.
For deployment, AWS recommends confirming that consequential actions require approval through connector permissions, configuring logging, and using Custom Permissions to limit the feature to a pilot team and defined workflow. Security leaders should test high-impact cases and verify that teams can trace actions from approval through execution. The stated measure of effective oversight is whether people retain control over consequential decisions, rather than how often an agent asks for approval.


