AI Automation with n8n: Human Approval for Critical Actions

A customer wants to cancel an order. AI reads the message, finds the order and drafts a reply. So far, the job looks straightforward. The consequential decision comes next: should the system cancel the order and initiate a refund on its own, or put the proposed action in front of an employee?

AI automation built with n8n can place that decision inside the workflow. The official human review documentation describes pausing execution before a selected tool runs. Approval lets the tool execute with its proposed inputs; denial prevents that action from running. This provides a concrete starting point when evaluating agents that can access company systems.

Separate the work into three parts

An AI workflow does not need to give every step the same freedom. The first part reads information: checking the right customer’s order status, retrieving product details and classifying the request. The second proposes a response: a draft message, a cancellation reason or the team that should handle the case. The third changes something outside the workflow: sending a message, modifying a record or initiating a refund.

For a first pilot, automating the proposal while requiring approval for the action makes results easier to inspect. Reading should still have limits; one customer request is no reason to expose the entire customer database to a model. Supply only the fields needed for the case.

Not every request needs an agent either. If the order number is in a known field and the decision follows an explicit rule, ordinary nodes and conditions may be easier to understand. Place the model where its contribution is useful, such as interpreting free text or drafting a response.

An example cancellation workflow

The following is a proposed design, not a deployed customer system or a refund policy built into n8n.

1. Validate the request. Check the source, customer match and event identifier of an incoming webhook. Do not modify a record when identity information is missing.
2. Read the current state. Retrieve dispatch status from the authoritative system. The model’s conversation history is not a substitute for current stock or shipment data.
3. Prepare a proposal. Have the model return a summary, proposed action and reason in defined fields. Compare values such as the order identifier and amount with the source record.
4. Apply business rules. Route ineligible orders or actions above the company’s limit to the responsible team. Do not rely solely on instructions written into a prompt.
5. Request human approval. Show the reviewer the customer, order number, action, amount and proposed reply together.
6. Recheck the state before execution. The order might have shipped while approval was pending. If the state or action parameters changed, reassess instead of using the earlier approval.
7. Confirm the result at the source. Check the API acknowledgement and then the actual order state. Do not tell a customer that a refund is complete when the operation is still pending.

Give reviewers enough information to decide

“Should we continue?” is insufficient on its own. A useful approval workflow shows what will change and which record will be affected. Hide unnecessary personal information while keeping the decision fields visible. Define who can approve and where the request goes if nobody responds.

n8n documents passing the tool name and model-supplied parameters to the reviewer through channels such as Slack, Teams or n8n Chat. Choosing a channel does not automatically establish your organisation’s approval authority. Your workflow still needs to verify the reviewer, their permissions and the exact version of the action being approved.

Use access restrictions to prevent an agent from pursuing a denied action through another tool. Explain how denial should be handled in the system instructions, and retain application-level permission checks alongside those instructions.

Keep failures and retries from becoming duplicate actions

A network timeout can leave you unsure whether the first request completed at the destination. Immediately repeating a cancellation or refund request can create a duplicate. Use an idempotency key if the target API supports one. Otherwise, design result lookups and duplicate controls around an operation identifier in the integration.

The n8n error handling documentation describes assigning a separate error workflow that begins with Error Trigger. Stop And Error can deliberately fail an execution when a condition requires it. A green execution status still does not prove the business outcome is correct; validate source data and the resulting state separately.

Measure more than minutes saved

n8n’s evaluation approach runs sample inputs through a workflow and compares the results. Include ordinary requests alongside incorrect order numbers, missing customer matches, repeated events, denied approvals and unavailable APIs in your pilot dataset.

Define these measures before starting:

• Draft correctness: How many reviewed drafts required no meaningful correction?
• Incorrect actions: How many unauthorised, duplicate or wrong-record operations occurred?
• Waiting time: How long did approvals take, and how many remained unresolved?
• Total effort and cost: What did the model, infrastructure, human review and error correction cost together?

An audit log should retain the operation identifier, reviewer and result without passwords, access keys or unnecessary customer data. n8n’s execution data settings control which execution records are saved and how old records are pruned. Plan debugging needs together with your organisation’s retention period.

Start with one request type, a limited user group and an explicit condition for stopping the pilot. As in our article on bringing marketplace questions and reviews into one operations screen, make the work queue understandable first. Expand automatic execution authority when the measurements and observed failures support that decision.

Explore our automation and workflow services and API development and integration, or contact us to define a pilot around your current process.

Sources

Official documentation checked on 5 September 2026. Verify version and channel support in your installation before implementation.

• n8n: Human review for tool calls
• n8n: Error handling
• n8n: Evaluating AI workflows
• n8n: Execution data retention