Smart Approvals in Amazon Quick: Human oversight where it matters

Preserving human oversight amid AI agents taking on more tasks is a priority for security leaders. While granting permissions allows agents to perform certain actions, granting approval for every step introduces repetitiveness and potential for diminished attention over time. Smart Approvals in Amazon Quick offers a solution by reviewing eligible actions within existing controls, automatically approving them or returning them to the user for review without adding extra permissions or removing required human approvals.

This allows security leaders to retain control over which actions always require human intervention, while reducing unnecessary interruptions in routine work. Actions that must always be reviewed are determined by policy, while Quick evaluates each proposed action against various controls, including Custom Permissions, Quick Guardrails, connector action permissions, and the user's settings before automated review.

Eligible actions are reviewed using a separate model, the classifier, which assesses the action's purpose, data flow, and affected parties, considering the user's knowledge graph, which includes relationships between people and entities in connected sources. This ensures the classifier understands the context, such as whether a recipient is in the user's reporting chain.

Actions that aren't eligible for automated review, such as sending content outside the organization or to senior leaders, require human review. The classifier provides judgment, while the permission service enforces authority. Automatic approvals are single-use and only apply to the exact action and parameters reviewed, while denied actions or review failures return to the user.

Audit evidence and investigation tools, such as Quick permission service records in AWS CloudTrail, chat logs, system response logs, and destination system evidence, help reconstruct events without treating approval as proof of execution.