Amazon Quick introduces Smart Approvals for automatic approval of AI agent actions

Amazon has introduced the Smart Approvals feature in the Amazon Quick product, designed to reduce fatigue from repeated approval prompts during AI agent operation. Instead of requesting approval for every step, the feature evaluates the proposed action against the user's request, configured permissions (Custom Permissions, Quick Guardrails, connector permissions), and either approves it automatically or returns it to the user for review. According to the company, the feature does not add any new permissions and does not bypass mandatory human approvals defined by the administrator.

The evaluation is performed by a separate model referred to as a classifier, which assesses what the action does, where the data is headed, and who is affected, and may use a knowledge graph of the user's relationships from connected sources for this purpose. According to the company, the data processed by the classifier is not used to train models. Some types of actions – sending content outside the organization or leadership, changing access rights, deleting records, or payments – are, according to the article, always supposed to require human approval, though the list is not exhaustive and administrators can designate additional actions that will never be subject to automatic approval.

Each automatic approval is one-time and applies only to the specific action and parameters assessed by the classifier; it is not a permanent permission. The feature is enabled by default, profiles with default-deny settings must explicitly enable it, and individual users can turn it off themselves for all conversations and scheduled runs, without the ability to re-enable it at the account level. Amazon emphasizes that comparing the action to the user's request also functions as a check against indirect prompt injection attacks, where an agent attempts to follow instructions hidden in retrieved content.

Permission updates associated with Smart Approvals are recorded in AWS CloudTrail within the customer's account, including identification of the user, connector, action, and time. The company notes that a CloudTrail record alone does not prove that the action actually took place – chat logs and records from the target system also serve for verification. CloudTrail event history is retained for 90 days; for longer investigations, a custom trail needs to be set up.