Tutorial
n8n Human Review: Approve AI Tool Calls Before They Run
The takeaway
Configure native n8n tool review, show the proposed action, and inspect rejection and timeout paths. Includes a free acceptance worksheet; documentation-based, not runtime-tested.
In this article
Documentation-based walkthrough. Sources checked November 12, 2026. We have not executed this n8n workflow or verified a specific n8n version. The worksheet contains acceptance cases to run in your own test environment, not measured results.
Start with the tool that changes something: issuing a refund, sending a message or updating a record. Review belongs before that tool runs. Checking the agent's final answer cannot undo an earlier action.
n8n documents native review for selected AI Agent tools. Approval permits the proposed call; denial cancels it and tells the agent. This walkthrough applies that feature to a synthetic refund, then gives you a way to inspect the decision and what actually happened. See n8n's human-review guide for the native behavior.
1. Prepare a test action and a reviewer
Use an existing test n8n workflow with an AI Agent, a configured model and a tool that records a mock refund in a test ledger. Keep real payment credentials out of this exercise. Model setup and the mock ledger are prerequisites; this article does not supply an importable workflow or a payment integration.
Save the free acceptance worksheet. Record your n8n and node versions, workflow revision, review channel and storage configuration. Choose who is allowed to decide and where that authority will be checked.
Use this fixture as the proposed tool input. The amount is 4,800 cents, or $48 USD. The identifiers refer only to synthetic records:
{
"operation_id": "refund-demo-001",
"order_id": "order-demo-041",
"amount_cents": 4800,
"currency": "USD",
"destination": "customer-demo-017",
"reason": "Duplicate order"
}These are example application fields. A label such as designated_reviewer in a message does not authenticate that person or enforce their authority. The worksheet separates displayed fields from the component responsible for enforcing each rule.
2. Put the selected tool behind native review
- Open the AI Agent's Tools connector.
- In Human review, select Chat for this walkthrough.
- Configure the channel, then attach the mock refund tool to the review step's tool connector.
These labels follow the documented setup. Check the actual connection used by the agent, not just whether a review node appears somewhere on the canvas.
- Chat Trigger → AI Agent: the conversation supplies a request.
- Agent tool connector → Chat review → mock refund tool: review sits on this tool path.
- Test ledger: inspect the effect separately from the review response.
3. Configure the top-level Chat path
The Chat node documentation requires a Chat Trigger with response mode Using Response Nodes. Use a top-level workflow: Chat is not supported in Embedded mode, as a sub-agent tool, or inside a sub-workflow. Broader support described for other review channels does not establish support for those Chat arrangements.
Use Send and Wait for Response, response type Approval, and Approve and Disapprove. For the initial test, enable Block User Inputso the reviewer uses the buttons. With it disabled, typed replies are documented as disapproval with a message.
Limit Wait Time can resume execution at its deadline. Capture the actual output and downstream behavior before relying on a timeout path. This walkthrough does not assert an unobserved timeout payload or treat a resumed workflow as an approved action.
4. Show the action the reviewer is deciding on
n8n exposes $tool.name and $tool.parameters in the review message, including schema fields supplied through $fromAI(). Check the expression reference.
Review the proposed test action: {{ $tool.name }}
Proposed input:
{{ JSON.stringify($tool.parameters, null, 2) }}
Approve only after checking the amount and destination.
Disapprove if the proposal needs a change.This is a suggested message to adapt, not a tested node export. Inspect its rendered values beside the resolved tool inputs. Fixed destinations, credentials, defaults and environment choices can change an action without appearing in the AI-generated fields. Add the relevant context deliberately and keep secrets out.
Explain the review setup and denial handling in the agent's system prompt. For this example, instruct it to report a denial, ask for clarification when necessary and propose a changed action for fresh review. A prompt helps the conversation; the workflow must enforce the decision at the tool boundary.
5. Exercise approval and rejection separately
Ask the test agent to propose the fixture. Before deciding, capture the pending request and the ledger's effect count. After approval, record the tool invocation and compare its input and effect with the proposal. Do not use the agent's success message as the only evidence.
Use a new operation ID for the rejection case. Disapprove the request, then inspect the agent's response and whether that call produced an effect. Track any subsequent proposal separately. Changing an amount after rejection should not silently reuse the earlier decision.
The worksheet also covers changed actions, expiry, repeated decisions, process restart, concurrent reviewers and a wrong reviewer. Every case starts as Not run. Record unavailable cases honestly and retain failed attempts. An empty ledger can mean rejection, a failed tool, or an incomplete execution; the decision and invocation records distinguish those outcomes.
Troubleshoot the decision and the effect
| Symptom | What to inspect |
|---|---|
| The tool runs without review | Trace the actual tool call. Check whether it is attached through the review step or still reachable directly. A review after the effect is too late. |
| The reviewer cannot see the destination or amount | Compare the message with the resolved tool input, including fixed settings and defaults. Stop the test if a consequential field is missing. |
| A typed “yes” does not approve | With custom Chat input allowed, the documented behavior is disapproval with a message. Test the buttons separately from typed responses. |
| The workflow resumes after waiting | Capture the timeout payload and selected branch. Resumption alone does not establish an approval. Do not infer success from a completed workflow badge. |
| A second click or retry gives an unexpected result | Keep the operation ID constant. Count review decisions, tool invocations and mock effects separately; inspect which component prevents reuse. |
| An execution ends without a clear outcome | Inspect the execution and mock ledger before retrying. Assign an operator to reconcile an uncertain effect rather than automatically starting the action again. |
A standalone Wait node is a separate pause primitive. Building an approval flow around its resume webhook also requires a decision model and an authority check. A webhook authentication setting alone does not establish permission for a specific action.
Apply the findings to one workflow
Keep four records distinct: what was proposed, who could decide, the decision received, and the action's outcome. Name which controls come from native n8n behavior, your workflow, or the external action provider. An execution log is not automatically an independently issued signature.
Use the approval design checklist to turn missing controls into concrete requirements. The authorization foundation explains the difference between delegated access and approval of an exact action.
This tutorial is educational material, not a Permission Protocol n8n integration or a security certification. If native review meets your requirements, use it. If your test exposes an unresolved approval or evidence gap, bring the workflow and the observation to a conversation with us.
For executed examples in other frameworks, see the LangGraph tutorial and OpenAI Agents SDK tutorial. Start with the human-approval guide for framework-neutral review design.
Put it into practice
Plan your approval boundary
Use your worksheet findings to define who decides, what they approve, and how your workflow handles an uncertain outcome.