AI Agent Approval Workflows
Let your team investigate customer issues and request operational actions with a clear decision at each step: allowed, approval required, or blocked. GlowForm checks published tools against workspace rules and the requesting person’s agent policies. The assistant can propose an action; your configured permissions determine whether it can run.
Three outcomes, defined by your policies
Consider a failed setup job for Northline, an illustrative customer. These examples assume an administrator has configured the connection, published the relevant tools, and set the access policies. The retry action depends on your connected system and its recovery procedure.
- Allowed: read the job’s status through a published tool, within the records and fields the teammate is permitted to see.
- Approval required: request a production retry when policy permits the action but requires a designated person or group to approve it.
- Blocked: attempt a tool or environment the teammate has no grant for, access a record outside their scope, or request an action prohibited by workspace rules.
An approval does not override a denial. A connection must also be usable, with the requested capability enabled and the target environment configured.
Review the action before it runs
GlowForm prepares an operation showing the tool, inputs, and environment. When approval is required, the request waits for the configured approver. The requester cannot approve their own request, and being a workspace owner does not automatically make someone its approver.
The decision is bound to the version of the request the approver reviewed. If it changes, the old approval cannot authorize the new action. Pending requests expire after 24 hours; the requester can submit them again for review.
Check permissions again at execution
Approval queues the operation. Before execution, GlowForm checks authorization again, including whether the tool version is still published and the current policy still permits the request. An earlier approval does not preserve access that has since been removed.
Review the operation’s status and execution result alongside the audit history. If an external action has an uncertain outcome, check the target system before retrying. Human approval does not replace a recovery procedure or make every action safe to repeat.
Start with one operational workflow
- Configure a supported connection and enable only the capabilities the workflow needs.
- Publish tools with explicit inputs, target environments, and resource scopes.
- Grant access to the appropriate users or groups and name the approver for actions that need review.
- Use the access simulator to check the expected decisions before trying the workflow in a test environment.
For a walkthrough from investigation to resolution, read How to Reduce Support Escalations to Engineering. For the design principles behind these controls, see human-in-the-loop approvals for AI agents.
Bring a workflow to the demo
Show us an issue your operations team regularly escalates, the systems involved, and who should approve a recovery action. We can discuss the tools and permissions that workflow would need.