AI governance & policyAgenticJuly 25, 2026

The Human-in-the-Loop Audit

Suggested tool: Claude (strongest at structured, nuanced classification), ChatGPT, or Gemini.

The scenario

Your team has started using AI across your comms work, but no one has drawn a clear line between the steps AI can own and the steps that need a human's judgment, relationships, or accountability. You want that line drawn, documented, and turned into a starter AI-use policy you can share this week.

The prompt

Act as a communications risk and AI governance advisor. Below is a workflow my team runs, broken into steps. For each step, classify it as one of three: AI-Owned (the machine can do this with light review), Human-in-the-Loop (AI drafts or assists, but a person must review and approve before it moves), or Human-Only (judgment, relationships, ethics, or accountability make this a person's job). For each Human-in-the-Loop and Human-Only step, name the specific risk that justifies keeping a person in it: accuracy, tone, legal exposure, relationship, confidentiality, or reputational judgment. Then flag any step where we are currently letting AI do more than it should, and recommend the guardrail that fixes it. Output a table, then write a five-line AI-use summary my team could follow tomorrow.

How to use it

Pick one real workflow your team runs often. Media pitching, executive messaging, internal announcements, monitoring, or reporting. Choose the one where AI has quietly crept in the most.

Write it out as the actual steps you take, start to finish, in plain language. One line per step is enough.

Paste your step list into Claude, ChatGPT, or Gemini along with this prompt:

Review the classification and push back. Move a step you disagree with, tighten a guardrail, or add a risk the model missed. You know your exposure better than it does.

Turn the output into a one-page guardrail you actually publish. Where AI can own the step, where a human must sign off, and the three rules everyone on the team follows.

Why it works

It turns a vague value ("humans still matter") into a concrete map of who owns what, which is the thing your team can actually follow on a Tuesday.
It hands you the first draft of an AI-use policy without a committee, a consultant, or a three-week working group.
It works at any altitude. Run it on one pitch workflow today, or run it across every workflow your team owns.

Cowork version

For Claude Cowork, where the prompt can read the files in a folder.

Run a Human-in-the-Loop Audit for my comms team across these workflows: [list your 2 to 4 workflows]. Do the following:

1. For each workflow, break it into its real steps and classify every step as AI-Owned, Human-in-the-Loop, or Human-Only.

2. For each Human-in-the-Loop and Human-Only step, name the specific risk (accuracy, tone, legal exposure, relationship, confidentiality, or reputational judgment) that justifies keeping a person in it.

3. Flag every step where we are currently letting AI do more than it should, and recommend the guardrail that fixes it.

4. If I shared existing guidelines, check my workflows against them and note where we are out of policy.

5. Write a one-page team AI-use policy (what AI can own, what always needs human review, and the three rules everyone follows). Save it as Comms-AI-Use-Policy.md in [target folder], plus a Human-in-the-Loop-Audit.md with the full step-by-step tables.

This prompt first ran in TIC. Read the original issue →