Evaluation method
Human Approval for AI Agents: What to Approve Before a Tool Changes Data
Approve a concrete action with a defined target, scope and effect. Then enforce that decision in the system that performs the action. This guide is a method for designing and testing controls; it does not report a completed Braio access-control or Slack demonstration.
Distinguish permission from approval of an action
Permission determines whether an identity may perform an operation on a resource. Approval records a decision about a particular proposed action. An approval message cannot safely expand the underlying account's authority, and a broad account permission does not establish that a person intended a particular change.
Start from the operation, target organization and resource. Give the tool only the access needed for the task, deny actions outside that scope and enforce authorization on each request. These principles follow OWASP's authorization guidance; a natural-language instruction alone is not an enforcement mechanism.
Choose controls by effect and reversibility
Use this proposed matrix to agree a policy. The right boundary depends on your data and workflow; this table does not claim that a particular product supports every control.
| Action | Review before execution | Evidence after execution |
|---|---|---|
| Read a scoped record | Identity, allowed organization and permitted fields | Resource and request outcome without exposing secrets |
| Change a record | Target, exact fields, current value, proposed value and recovery path | Actual diff and operation outcome |
| Send a message | Recipient, channel, final text and attachments | Delivery result and destination |
| Merge code | Current revision, review outcome, checks and authorized merger | Merged revision and subsequent verification |
| Delete or affect many records | Exact set, count, dependencies and explicit authority | Affected set and any partial failure |
Make the proposed change reviewable
Show the final target and effect before seeking approval. For example, a proposed change might specify a test workspace, a draft ticket, its old and new status, and the reason. This is an illustrative decision record, not a recorded Braio transaction. If the target, content or revision changes, determine whether the existing decision still covers the new action.
Define who can approve, how long the decision remains valid, and whether approval can be reused. Handle duplicate requests, expired decisions and partial execution explicitly. Keep approval records linked to execution results so that 'approved' cannot be mistaken for 'completed'.
Test refusal and revoked access as well as success
Use a non-production environment and authorized accounts. Test an allowed operation, an operation outside the role or organization, a denied approval and an action after revoking access. Record the actual result at the system boundary. A reassuring assistant response is insufficient if the downstream operation still succeeds.
Name the owner who can suspend the integration and reconcile any partial changes. Repeat the relevant tests after changes to roles, connectors or approval logic. A passing sample does not constitute a security certification or proof that every authorization path is safe.
Confirm the controls for the proposed Braio workflow
Braio's backend implements authentication, roles and organization boundaries. This is an architecture statement, not a published penetration-test result. Actual deployment, retention, residency and contractual requirements need confirmation for the intended use.
Selected operations show estimates and require consent according to configured cost settings. Do not infer approval before every action. Slack and Teams remain configuration-dependent integration scopes; a real demonstration of acceptance, refusal and revocation requires its own environment and evidence.
Questions and answers
- Does human approval replace access controls?
- No. Approval expresses a decision about an action. The executing system must still check the acting identity, target and allowed operation. Test both the decision and the enforcement.
Sources
Continue with your workflow
- An AI Delivery Workflow for Software Agencies: Client Request to Reviewed Result
- Independent Review for AI-Generated Code: An Evaluation Checklist
- A Seven-Day Pilot Plan for an AI Coworker in a Software Team
- AI agent security - access, approvals and costs | Braio
- AI agents for team workflows - tasks, results and approval | Braio

