The short answer
An AI agent approval gate pauses a proposed action until the required reviewer approves it. Policy checks determine which actions can proceed automatically, need review, or must be blocked.
Agentic governance
AI agent approval gates
Place approval before a consequential action, with enough context for a person to make the decision.
Cortez Lake
Design approval gates around consequential agent actions: define what needs review, give reviewers the necessary context, and test how approvals and denials are enforced.
What this means in practice
Define the action being authorized, the person allowed to approve it, and what happens if review is denied or unavailable. The workflow must enforce that decision before the action runs.
Bring H2H in when…
- An agent can send messages, change records, or call tools with consequential permissions.
- Sensitive information needs review before access or transfer.
- A team needs to reconstruct who authorized a particular agent action.
Related H2H pages
AI governance and agentic securityDefine approval and policy controls for the agents and systems in scope.AI governance layer for enterprise workflowsPlace approval gates within a broader system of policy, evidence, and ownership.AI workflow automationTest a workflow improvement with agreed human review and pilot boundaries.
Decide which actions need a person
List the actions the agent can take and the consequences of a mistake. Drafting a message for someone to edit differs from sending it to a customer. Reading a record differs from changing it. The approval rule should reflect that difference.
Use policy to identify routine actions that may proceed and actions that require review or denial. Avoid requiring a human click for every step without considering reviewer capacity and the quality of the decision being requested.
Make the approval request specific
The reviewer should see the proposed action, its destination, the information involved, and the reason review is required. The approval must refer to the action the system will execute, rather than a vague instruction to let the agent continue.
Decide when a changed request needs a new review. For example, approval to send a particular draft to one recipient should not silently cover a different message or destination. Keep the authorized action and the executed action connected in the review record.
- Name the eligible reviewer and provide the context needed for a decision.
- Define denial, expiry, escalation, and unavailable-reviewer behavior.
- Record the proposed action, policy decision, approval, and execution result as appropriate.
Test the gate at the point of action
A review screen is not enough if the same action can be executed through an ungoverned path. Identify the tool calls and integrations covered by the control, then test that the action waits for the required decision.
Include denied requests, repeated attempts, expired approvals, changed inputs, and failed dependencies in evaluation. Plan how the workflow behaves when the control or reviewer is unavailable. Treat these as implementation requirements to verify, rather than capabilities to assume from a product label.
Fit the product to the approval design
Tutela supports policy decisions that can allow an action, issue a warning, require review, or block it for supported integrations. H2H helps teams assess where those controls fit and what additional workflow work is needed.
Coverage depends on the connected tools, actions, and policy configuration. Document those boundaries and review them when integrations change. A record of a decision helps investigation; it does not guarantee that every agent action was covered or every human approval was correct.
Next step
Define which agent actions need approval.
H2H helps map the actions, review responsibilities, and enforcement points your workflow requires, then assess the controls that fit.