AWS has published a description of Smart Approvals in Amazon Quick, a feature that decides whether an agent's proposed action needs a person to confirm it. The source page does not give a release date, so this item covers what the feature does and how it is controlled.
What it does
Smart Approvals compares each proposed action with the user's request. It then approves the action automatically or returns it to the user. It does not add permissions and does not remove human approvals that are already required.
Policy sets the limits
Before any automated review, Quick checks the action against Custom Permissions, Quick Guardrails, connector action permissions and the user's own settings. Denied actions stay denied. An action is eligible for automated review only when the user's setting is "Ask each time" and nothing else demands human approval.
Administrators use Custom Permissions to decide who can use the feature. It is on by default, but profiles with deny-by-default enabled have to allow it explicitly. Users can switch it off for themselves, and that choice covers conversations and scheduled runs. When a user has switched it off, no account-level setting turns it back on for them.
How the review works
The classifier is a separate model. It looks at what the action does, where data goes and who is affected, using the request and a limited view of the conversation. It can also consult the user's knowledge graph, built from connected sources, to see how a recipient relates to the user, for instance whether that person sits in the user's reporting chain. AWS says this content is not used to train models or improve the service.
Examples that should go back to a person include sending content outside the organisation, sending to senior leaders in the reporting chain, changing access, deleting records and starting payments. AWS says the list is not exhaustive. The request-matching step also acts as a check against indirect prompt injection, where text in retrieved content tries to steer the agent.
AWS is open about the limits. Model review can make mistakes, so actions that must always get human approval should be enforced through connector action permissions. Each automatic approval is single-use and covers only the exact action and parameters reviewed. Timeouts and review failures never produce an approval.
AWS gives an example. An account manager asks for a quarterly review brief to go to two named colleagues. A note in the source material says it should also go to an all-sales list of about 2,000 people. The send is designed to come back to her, because the audience has changed.
Audit trail
Permission updates linked to Smart Approvals are recorded in AWS CloudTrail, with the user, connector, action and time. The marker does not prove the grant was accepted or the action ran, so the destination system, such as the sent email or updated ticket, is the proof of execution. CloudTrail Event history keeps events for 90 days, so longer investigations need a configured trail. Chat logs capture only activity after delivery to Amazon CloudWatch Logs is set up, and they can hold sensitive content, so access should be restricted and data protection policies considered.
Suggested rollout
AWS recommends confirming connector action permissions first, setting up logging, and limiting the feature to a pilot team and a defined workflow. Define success beforehand as fewer prompts for routine actions with consequential ones still reaching a person, and expand only when pilot evidence supports it.