AI & AUTOMATION · 3 MIN READ
Where Human Approval Belongs in an AI Workflow
Design useful human approval steps for AI workflows: show the proposed action, bind approval to its details, and handle changes or missing responses.
An approval step can keep a person in control, but only if that person understands the action and has a meaningful opportunity to change it. A notification that says “Approve automation?” is too vague to support a careful decision.
Design the review around the business action, not around the fact that a model was involved.
01
Separate preparation from execution
Consider a fictional sales assistant that researches an enquiry and drafts a reply. Gathering approved information and preparing text are different from sending that text to a customer. The draft stage can proceed without making an external commitment; sending is a useful boundary for review.
Other boundaries might include changing a customer record, approving a discount, or closing a ticket. Decide which consequences matter in the specific process. Requiring approval for every harmless lookup can create a queue that people learn to dismiss mechanically.
02
Show the actual proposal
A reviewer should see the recipient or record, the exact proposed change, the information supporting it, and any unresolved assumptions. For a record update, show the current value and proposed value together. For a message, show the final text and attachments.
Avoid making the reviewer reconstruct the decision from a long execution log. Offer deeper context when needed, but keep the decision surface focused. A useful test is whether a colleague unfamiliar with the implementation can explain what pressing Approve will do.
03
Approve a version, not a general intention
My recommended design binds approval to a particular proposal. If the destination, amount, message, or source data changes materially, return it for review. An approval for one version should not silently authorize another.
Record who approved it and what was approved. Before execution, the application should still check permissions and whether the underlying record has changed. These are application responsibilities; a human's earlier click does not eliminate them.
04
Make rejection and silence normal outcomes
In n8n, selected AI tool calls can pause for human review: approval executes the proposed tool input, while denial cancels that action. The feature provides a useful mechanism, but the surrounding business rules remain part of the implementation.
Define a waiting state, a responsible reviewer, and what happens when the review expires. Silence should leave a consequential action pending or canceled according to the agreed policy, rather than being treated as approval. Rejection should preserve enough context to correct the proposal without repeatedly presenting the same flawed action.
05
Evaluate the review itself
During a pilot, track how often reviewers edit, reject, or misunderstand proposals. Frequent edits may reveal poor source information or an overly broad task. Long delays may mean the review is assigned to someone who cannot assess it.
The objective is a workflow in which people can make informed decisions with less effort. If the review requires as much reconstruction as doing the task manually, simplify the proposal before expanding the automation.
06
Discuss the implementation
See AI workflow automation or how a reliable support workflow is structured.
KEEP EXPLORING
Sources & further reading
Official documentation for the topics discussed. Check implementation details against the versions and services used by your project.
PUT THE IDEAS TO WORK
Need help with
the implementation?
Explore the approach and scope for ai workflow automation.
Explore the service