Where Human Approval Belongs in AI Automation

· by Andrei

A customer asks whether a replacement buckle fits an older hiking pack. The AI finds the right part and drafts a clear reply. It also picks a different customer with the same surname. If the message goes out, the person who asked is still waiting and someone else has their order details.

Put approval just before sending. Show the recipient and the full reply on the same screen, and give the reviewer enough context to spot what’s wrong.

Slack’s November 18 Workflow Builder announcement says an AI or agent step automatically gets an editable review before its output moves on to other tools. That gives teams a place to pause. If you’re designing your own AI integration, you also need to decide exactly what that approval permits.


Put the order and the reply on the same screen

Take that hypothetical buckle request. The support agent opens a review and sees the original ticket, the pack model, the order number and the outgoing email. The recipient’s full address is right above the message. They can compare it with the address on the verified ticket without opening another tab or trusting the draft’s greeting.

I’d give the reviewer three choices.

  • Approve this email to this recipient.
  • Edit the reply and review the updated version.
  • Reject it and return the ticket to the support queue.

Use the same support permissions you already rely on. A colleague who can read a ticket may have different permissions from the person who can reply to it. If the order is ambiguous or the contact can’t be verified, route the ticket to someone who can resolve it. An approval screen can show the mismatch. It can’t tell the reviewer which order is correct.


Approve the exact email that will go out

The workflow should save which action was approved, who approved it, and the recipient and version of the message they saw. If an edit changes the body or address after approval, pause for a fresh review. Do the same if a relevant fact changes, such as the order being reassigned or the replacement going out of stock before the reply is sent.

A rejection ends that send attempt. So does a review that times out. The ticket can go back to the support queue with a clear reason. If the email service is slow and the workflow retries, make sure it checks whether the approved send already went through. The customer should get one reply even if the send step runs twice. If delivery is uncertain, bring that uncertainty to the support queue before anyone sends it again.

Keep enough of a record to answer a later question about what happened. The ticket ID, reviewer, time, approved message version and delivery result are useful. Store them where authorized staff can see them and avoid copying order details into broad team logs. These are design choices for your own workflow. OWASP’s guidance on excessive agency explains why limiting what the AI can do also matters. Give a support assistant access to the order information it needs to draft a reply, and put the send permission in the part of the workflow that checks approval. The same idea applies when granting an AI agent access to WordPress actions.


Pick checks that match the consequences

At the outdoor shop, the AI could tag a ticket “replacement parts” automatically if that label is easy to undo. A support lead can spot-check those tags later. Sending order information outside the team deserves a check in advance. If the team gives people the same approval prompt for every tiny change, it gets harder to tell which click warrants a closer look.

Before launch, try four variations of the buckle ticket with dummy contacts. Give the workflow the wrong recipient and see whether the reviewer can catch it. Change the message after approval and watch it pause again. Reject a draft and let another time out, then check that neither was sent. Finally, retry the send step on an approved reply and check the test inbox for exactly one message. Start with whichever case your current workflow handles least clearly.

← Back to blog