Add human approval to n8n AI Agent tool calls and gate irreversible actions first
Key Takeaway
On the n8n AI Agent node, a Human review step lets a person approve a tool before it runs. Gate tools that delete data or send messages first, and tell the agent in the system prompt what to do after a denial.

Which tools should get human approval?
Start with actions that cannot be undone or that leave your system. The official documentation gives examples such as deleting data, external communications and purchases, plus work where compliance requires a person's approval and high-impact decisions.
Approval can cover every tool attached to the AI Agent node or only the tools you pick. Begin with the one risky tool, receive a few approval requests, then widen or narrow the scope. The documentation also suggests starting with review turned on and reducing oversight as trust builds. Leaving reversible tools such as lookups without approval keeps the number of requests reviewers receive down.
How do you connect the approval step?
The flow is agent picks a gated tool → workflow pauses → request goes to the approval channel → reviewer responds → tool runs or is canceled. All of the setup happens inside the AI Agent node.
- Click the Tools connector on the AI Agent node to open the tools panel.
- In the Human review section of the panel, choose an approval channel and fill in its credentials and settings.
- Connect each tool that needs approval to the Human review step's tool connector, and configure every tool as usual.
Tools that do not need approval stay unconnected. For how to attach an existing workflow as a tool, see /workflow/n8n-agents-workflow-tools-preview/.
What should the approval request show?
Reviewers need to know what they are approving, so send the tool name and its parameters together. The documentation provides two expressions.
$tool.nameis the tool node's name as shown on the canvas.$tool.parametersholds the parameters the AI is trying to pass to the tool, including fields filled by$fromAI().
The documentation's sample message starts with "The AI wants to use" plus $tool.name, and the next line adds the output of JSON.stringify($tool.parameters, null, 2). Because $fromAI() works on gated tools, the reviewer sees the values the AI chose before approving. This message format is only the documentation's example, so open each channel to see how it actually looks there.
Which approval channels can you choose?
The documentation lists n8n's built-in Chat, Slack, Discord, Telegram, Microsoft Teams, Gmail, WhatsApp Business Cloud, Google Chat and Microsoft Outlook. The review channel can differ from the channel you talk to the agent in: users can chat in n8n Chat while approval requests go to a named person in Slack.
Decide first which channel your approver checks fastest. If you send requests only to a messenger the approver rarely opens, the workflow may stay paused for a long time.
What happens when the reviewer approves, denies or does not answer?
| Reviewer response | Tool behavior | Agent side |
|---|---|---|
| Approve | Tool runs with the input the AI specified | Receives the tool result and continues |
| Deny | Execution is canceled | Is told the request was denied |
| No response | Not stated in the documentation | Not stated in the documentation |
Checked: 2026-10-27 KST (source: n8n documentation, Human-in-the-loop for tools)
The documentation tells you to describe the gated tools and the Human review setup in the system prompt as well. The agent has to know about denials to tell the user what happened, suggest another approach or ask a follow-up question. List which tools go through approval and what the agent should do after a denial instead of repeating the same call.
The documentation does not state version or plan requirements either. Open the tools panel in your n8n environment and confirm the Human review section appears.
What should you test before using it in production?
- Check in the tools panel that each gated tool is really connected to the Human review step.
- Look at whether the approval message shows the tool name and parameters in a readable form.
- Deny a request and confirm the tool does not run and the agent answers the way the system prompt says.
- Leave a request unanswered on purpose and write down what state the workflow stays in.
- If you use a sub-agent as a tool, test its approval step separately. The documentation says human review inside a sub-agent works correctly in that setup.
Pick one test target, such as a deletion or a message where nothing real is at stake, and alternate between approving and denying it.
Frequently asked questions
Do you have to add the approval step to every tool on an AI Agent?
No. The official documentation says human review can apply to all tools on the AI Agent node or only to the tools you select.
Does the approval channel have to match the channel you chat with the agent in?
No. The review channel can differ from the main interaction channel. For example, you can chat with the agent in n8n Chat and receive approval requests in Slack.
What happens if the reviewer never responds to an approval request?
The official documentation does not say what happens without a response. Test it yourself before moving the workflow into production.
--- Official source: n8n Docs (checked 2026-10-27 KST) - Human-in-the-loop for tools: https://docs.n8n.io/advanced-ai/human-in-the-loop-tools/
Frequently Asked Questions
Do you have to add the approval step to every tool on an AI Agent?
No. The official documentation says human review can apply to all tools on the AI Agent node or only to the tools you select.
Does the approval channel have to match the channel you chat with the agent in?
No. The review channel can differ from the main interaction channel. For example, you can chat with the agent in n8n Chat and receive approval requests in Slack.
What happens if the reviewer never responds to an approval request?
The official documentation does not say what happens without a response. Test it yourself before moving the workflow into production.
Sources & Verification Date
All materials have been verified from official announcements and may have changed since verification.