Approval Gates: Nothing Goes Out Under Your Name Without a Person

An AI approval workflow puts a named person between an AI draft and anything that leaves the business: an email sent, a post published, money spent. The person approves a specific version, in a recorded act. "Looks good" in a chat is not approval. The gate never moves, even when the writer is also the approver.

This is general information, not legal advice.

What is an AI approval gate?

An AI approval gate is a point in a workflow where AI work stops and waits for a named person to approve it. AI can research, draft and prepare freely up to that point. Beyond it, the work takes effect: a message reaches a customer, a page goes live, a budget changes. Nothing crosses without the person.

The idea is simple, but most teams run it informally. Someone glances at a draft, says it is fine, and it goes. That works until the day it does not, when nobody can say who checked the figure or which version they saw.

A gate makes three things explicit. Firstly, where it sits in the workflow. Secondly, who may approve. Thirdly, what an approval looks like, so it can be told apart from a passing comment. The broader guide to human in the loop AI covers where gates belong across workflows.

Why does the pause matter more than the second person?

Because in a growing business the writer, the approver and the publisher are often the same person. The gate still works, because its job is to force a stop: a moment where someone deliberately reads the finished thing before it goes out. Without that stop, approval becomes a feeling of being finished rather than a decision.

That point surprises people. They assume an approval gate needs a manager or a second reviewer, and conclude a small team cannot have one. However, the main failure the gate prevents is not a lack of oversight. It is momentum: the draft looked right in the editor, so it went.

The pause also changes how the reader reads. Approving means taking responsibility for every claim, so people check the figure, the name and the recipient in a way they do not when skimming.

The law does not care whether a mistake was AI's. According to the ACCC, "It makes no difference whether a business intends to mislead or not". A claim published by AI is a claim made by the business.

What counts as approval, and what does not?

Approval is a recorded act by a named person that identifies exactly what they approved: this draft, this version, this plan. A thumbs-up in chat, "looks good" or "carry on" is not approval. Approving one thing never licenses another, so approving the ad copy does not approve the budget that runs it.

Watch for drift between surfaces. A draft approved in a review tool, then edited in the publishing tool, is no longer the approved draft. Approval attaches to the words as they were, so any change after approval needs approving again.

Some gates involve a different party. A customer story needs the customer's consent before drafting and their sign-off on the finished words. The client approving the copy is not the customer's sign-off, however well they know each other.

Where should approval gates sit?

Put a gate wherever work becomes hard to undo: publishing, sending, spending and changing records other people rely on. Separate gates for separate acts. Approving the words and putting them live are two decisions, often by two people, and outbound loaded into a sending tool is different from outbound launched.

Splitting gates this way lets AI do more, not less. Once copy is approved, AI can load an outbound campaign into the sending tool, paused, because a paused campaign is visible and easy to fix. The launch, when real messages reach real people, stays with a person. The guide to load, don't launch explains the line.

How do you stop AI crossing the gate?

Make it impossible in the system, not just against the rules. Give AI access that lets it prepare but not execute: drafts in the publishing tool, campaigns created paused, read-only access to records. Never let AI set its own work to approved, fill in an approval field, or treat a chat message as permission.

Rules alone fail under pressure. Someone in a hurry says "just send it", and an AI with send rights obliges. So the access level should hold the line even when instructions do not. That means connections start read only, with write access granted one tool and one action at a time.

Official guidance points the same way. According to the National AI Centre's Foundations guidance, businesses should "Build in human override points" where people can pause, override, roll back or shut down AI systems. The guide to AI agent permissions covers how to set access.

There is one thing AI may do after approval: mirror it. Once a person approves in the review tool, AI can update the status elsewhere, using exactly what the person recorded.

How should approvals be recorded?

Record four things for every approval: who approved, exactly what they approved (the version or plan), what the approval covered, and when. Keep the record next to the work, not in someone's inbox. Then anyone can see later what was agreed, and AI never needs to guess or fill in a date.

A simple record per piece is enough:

1. Who. The named approver, not "the team".
2. What. The version, file or plan, specific enough to compare against what went out.
3. Scope. The wording only, the claims too, or the publish as well.
4. When. The date, as the person recorded it.

These records pay for themselves the first time something is questioned. The guide to AI audit trails covers what else to keep.

Skein builds every workflow so nothing is published, sent or spent without a named person approving it, and every connection starts read only; the shared knowledge layer page shows how those permissions are set. That suits a business putting AI to work on anything customers see. A team using AI only for private notes and research needs nothing this formal.

Frequently asked questions

What is an AI approval workflow?

An AI approval workflow is a process where AI drafts and prepares work, then stops at a fixed point until a named person approves it. The approval is recorded and names the specific version approved. Only then is the work published, sent or acted on, by a person or by a system the person controls.

Does a small business need a second person to approve AI work?

No. In small teams the writer and approver are often the same person, and the gate still works. Its main value is the deliberate pause before anything goes out, which forces a careful read of the finished piece. What matters is that approval is a recorded act, not a feeling.

Is saying "looks good" approval?

No. Approval should be a recorded act by a named person that identifies exactly what they approved, such as setting a status on a specific version. A casual comment in chat is easy to misread, attaches to no particular version and leaves no record of what was agreed.

Can AI approve its own work?

No. AI may draft, check against rules and prepare work for review, but it should never set its own work to approved or fill in approval fields. After a person approves, AI can mirror that decision elsewhere, using exactly what the person recorded, without inventing dates or details.

Where should approval gates go in an AI workflow?

Place them wherever work becomes hard to undo: publishing, sending messages, spending money and changing shared records. Use separate gates for separate acts, such as approving words and then publishing them. Earlier steps like research and drafting can run freely, because they are easy to correct.

Related reading