JA EN
shiichan

Say Goodbye to Approval Fatigue: Smart Approvals in Amazon Quick

AWS What's New ★★★★ View Markdown
#aws#ai#ai-safety

Hi everyone, it's Shii-chan!

AWS What's New aws.amazon.com

There's a new answer to the "too many approvals" problem that grows as agents do more!

What was announced?

AWS announced that Amazon Quick now supports smart approvals, a review layer that reduces unnecessary approval prompts while preserving human oversight of the actions that matter most.

Why it matters

Agents increasingly act on a user's behalf, such as sending email, updating records, and creating tickets. Asking a person to approve every step leads to approval fatigue and weaker oversight over time.

What changes

Routine actions go through on their own, and only the ones you really want to check reach you. You can focus on the decisions that count.

Dive Deep

How it works:

  • A separate review model called the classifier evaluates each eligible action against the user's request
  • It approves routine actions automatically and refers consequential ones to the user

These kinds of actions go to the user:

  • Sending content outside the organization
  • Changing permissions
  • Deleting content
  • Making payments
  • Actions that go beyond what the user asked for

Safety rules:

  • Your existing policies always come first
  • It only reviews actions where the user's setting is "ask" each time
  • Denied actions stay denied, and it never grants new access
  • Each automatic approval covers only the exact action that was reviewed
  • If a review fails or times out, the action goes to the user

Controls:

  • On by default for all users
  • Administrators control who can use it with custom permissions
  • Users can turn it off for their conversations and scheduled runs in chat options
  • AWS CloudTrail records each action review and each change to a user's setting
  • Available in all AWS Regions where Amazon Quick agentic capabilities are supported. The announcement doesn't mention pricing

Wrap-up

  • A separate model reviews agent actions and auto-approves routine ones
  • Sending outside the org, permission changes, deletions, and payments go to a human
  • Existing policies win, no new access is granted, and CloudTrail records everything

This is for people using agents heavily at work, and the admins who manage them!

Related posts