Amazon Quick uses smart approvals to provide human oversight where it matters.
Amazon Quick Smart Approvals automate routine AI agent checks while preserving human review for consequential actions under existing permissions.
Key Takeaways
- Smart Approvals reduces routine confirmation prompts while preserving human review for consequential AI agent actions, without adding new permissions or overriding existing policy.
- It only considers an action eligible for automated review when the user has chosen “Ask each time” and no other control requires human approval.
- A separate classifier judges whether a proposed action fits the user’s request and context, considering data destination, affected people, and relationship cues from the knowledge graph.
- High-risk or broad actions, such as external sharing, access changes, deletions, payments, or unexpected recipients, still trigger human review; automatic approvals are one-time and scoped to the exact action.
- Audit events in CloudTrail record Smart Approvals-related permission checks and setting changes, but they show the trail rather than proving that an action was approved or executed.
Meta description: Amazon Quick Smart Approvals reduces routine prompts while keeping human review for consequential AI agent actions under existing permissions.
Tags: Amazon Quick, Smart Approvals, AWS, AI governance, security leaders
Featured image: abstract, friendly illustration of human review checkpoints, no real people, no logos, no text.
Smart Approvals in Amazon Quick gives security teams a way to let AI agents do more without losing human control. The feature screens actions that are already eligible under a company’s controls. It can approve routine steps automatically, or hand them back to the person who must decide. It does not add permissions, and it does not erase approvals that existing policy requires.
More prompts are not always safer oversight
Permissions decide what an agent may attempt, but permission does not decide whether a person should still look at the action. If every step needs confirmation, a multi-step task can produce repeated interruptions even when the steps are routine. Studies of security warnings have shown that people pay less attention and follow instructions less closely when they see repeated alerts. The real problem is keeping human attention available for the choices that deserve it.
Quick checks each proposed action against Custom Permissions, Quick Guardrails, connector action permissions, and the user’s own settings before the automated review begins. Any action denied by policy stays denied. Any action that must be approved by a person still needs that approval. Smart Approvals only considers an action eligible when the user has chosen "Ask each time" and no other control demands human review. When a connector is created, its action permissions can require human approval for particular actions.
Custom Permissions let administrators decide which people may use Smart Approvals. The feature is enabled by default, but profiles using deny-by-default must allow it explicitly. Users can disable it for themselves, and that choice follows them across conversations and scheduled runs. Once a user disables Smart Approvals, any action that would otherwise be reviewed automatically must be approved by that user, and no account-level setting can re-enable it for them.
A classifier judges eligibility, not authority
Smart Approvals relies on a separate model, described as a classifier, to compare the proposed action with the user’s request and a limited view of the conversation. It looks at what the action does, where information goes, and who is affected. It may also consult the user’s knowledge graph, which maps people and working relationships from connected sources. That lookup can reveal, for example, whether a recipient sits in the user’s reporting chain. The knowledge graph query remains inside the Quick account. Quick does not use the content reviewed by the classifier, including knowledge graph data, to train models or improve the service.
Some actions still deserve a person even when they appear to fit the request. Examples include sending information outside the organization, sending it to senior leaders in the user’s reporting chain, changing access, deleting records, or starting payments. That list is not exhaustive, and the classifier also considers context. Comparing the action with the original request also helps guard against indirect prompt injection, a situation where retrieved text contains instructions that try to steer the agent. Quick tells both the agent and the classifier to treat retrieved content as information rather than commands.
Automated review can be wrong. For that reason, any action that should always have human review, especially actions that move sensitive data, should be forced through connector action permissions.
The classifier supplies judgment; the permission service supplies authority. An automatic approval is used once and only for the exact action and parameters that were reviewed. It does not create standing permission. Quick marks automatic approvals in the conversation. If the classifier does not approve, the action returns to the user. Timeouts and other failed reviews also return to the user and never create automatic approval. These limits constrain what the approval authorizes, but they do not prove the classifier made the right decision.
A wider audience triggers human review
Consider an account manager who asks Quick to build a quarterly business review brief for AnyCompany from her team’s account notes and last quarter’s email. She then asks Quick to send it to two named colleagues. If the send is eligible for automated review, Smart Approvals checks the proposed send against her request and the conversation context.
If one account note says the brief should also go to an all-sales distribution list of about 2,000 people, the audience is no longer the two colleagues she specified. Smart Approvals is designed to return that proposed send to her for review, where she can reject it. Approval for the original recipients cannot be reused for the larger list.
Audit events show the trail, not the outcome
The Quick permission service records permission updates tied to Smart Approvals in AWS CloudTrail in the customer’s account. Events include the user, connector, action, and time, and they mark the request as associated with Smart Approvals. That marker alone does not prove the permission service accepted the grant or that the action executed. Quick also records permission-check verdicts, including denials from connector action permissions, and changes to a user’s Smart Approvals setting.
For investigations, these permission events can be matched by user and time with other Quick activity in CloudTrail and with logs sent through Amazon CloudWatch Logs. Chat logs preserve the user’s request, the system response, and the conversation context. Agent metadata logs record changes to agents. Destination-system records, such as a sent email or an updated ticket, show whether the action completed. Used together, these sources let a team reconstruct the sequence without treating an approval as proof of execution.
CloudTrail Event history keeps events for 90 days, so a dedicated trail is needed for longer investigations. Chat logs can contain sensitive material, so access should be restricted and CloudWatch Logs data protection policies can help mask sensitive data.
Pilot with measurable controls
Because Smart Approvals is on by default, teams should first confirm that actions requiring human approval are already enforced through connector action permissions. Then they should configure a CloudTrail trail and send chat logs and agent metadata logs to CloudWatch Logs to support investigation. Chat logs begin only after delivery is configured, so older approvals will lack conversation context. With those controls in place, administrators can use Custom Permissions to limit Smart Approvals to a pilot team handling a defined workflow before rolling it out to more teams.
Success should be defined before the pilot: routine actions generate fewer prompts, while consequential actions still reach a person. During the pilot, teams should sample automatic approvals and test high-impact scenarios, such as sending content outside the organization or adding a recipient introduced by a document. Those scenarios must return to the user. Teams also need to trace each action from approval through execution. Broad expansion should wait until the pilot meets those checks, and the review should happen again as actions and policies change.
The sign of effective oversight is not how often an agent asks for approval. It is whether people still control the decisions that matter. For security leaders, the argument for expanding Smart Approvals should be supported by pilot evidence. Fewer prompts are worthwhile only if they preserve attention without weakening accountability.
Setup details are in the governance section of the Amazon Quick User Guide under Smart Approvals.
Who wrote the guidance
Steve Vinsik is a Principal Product Manager at AWS, responsible for governance, identity, and security for Amazon Quick. He works directly with enterprise chief information officers and chief security officers to shape the controls organizations need when adopting AI agents at scale, including access permissions, guardrails, model management, data loss prevention, and observability. Before joining Quick, he led product for AWS Identity, covering AWS IAM Identity Center and AWS Builder ID, which he scaled from launch to millions of users. He has more than 25 years of experience delivering mission-critical security programs for government customers, with executive roles at General Dynamics, Unisys, and Safran.

Comments (0)