The Real Yes Rule: a one-page AI approval checklist for your office

An approval step only protects you if the person approving actually looks. Most approval queues fail quietly: drafts pile up in a shared inbox, someone clears them in a batch, and a wrong amount goes out with your name on it. The Real Yes Rule is three habits that stop that, and the one-page checklist further down puts them on paper.

The Real Yes Rule, in order

1. One named person. Every draft waits in one named person's queue, not a shared inbox.
2. A real check. That person checks the draft against the real record, like the invoice amount in your books, before saying yes.
3. A yes on money and promises. Anything about money or a promise can't go out without that yes. The agent can't send it on its own.

Why clicking approve isn't checking

A confident draft pulls people along, experts included. In a small study reported by the Radiological Society of North America, 27 radiologists read 50 mammograms with an AI suggestion, and the experienced readers were right about 82% of the time when the suggestion was correct and 45.5% of the time when it was wrong. The authors were careful about how small the study was, and the AI was described to readers as AI. But the lesson carries to an office: when a draft looks finished, people tend to agree with it.

Security guidance says the same thing from the other side. OWASP's list of AI risks recommends requiring a human to approve high-impact actions before they're taken, and keeping an agent's permissions to what it needs. The approval only works if the human does more than click.

Who this is for and what you need first

This is for any service business that has an AI agent drafting customer messages, invoices, follow-ups or updates, or is about to. You need:

• A list of every kind of draft the agent makes (reminder texts, invoice nudges, review requests, estimate follow-ups).
• A named approver for each kind. One person, with a backup for days off.
• Read access for that person to the record each draft is based on: the job, the invoice, the estimate.

Habit one: one named person

Shared inboxes are where approvals go to die. Everyone assumes someone else checked.

1. Give every draft type an owner by name. "Kim approves invoice reminders. Marcus approves estimate follow-ups."
2. Put the owner's name on the draft itself, so it's visible who said yes.
3. Name a backup for each owner. When the owner is out, drafts route to the backup, not to "the office".
4. Set a daily time for the queue. Ten minutes at 9 AM beats "whenever I get to it".

Habit two: a real check against the record

The approver compares the draft with the source, not with their memory. For each draft type, write down the two or three things to check. A starting list:

Draft type	Check against	What to compare
Invoice reminder	Your books	Invoice number, amount, due date, whether it's already paid
Estimate follow-up	The estimate	Customer name, what they asked about, the quoted scope
Appointment reminder	The schedule	Date, arrival window, address
Review request	The job ticket	Tech's name, the work done, any callback or complaint
Customer update	The job notes	What got done, what's next, any date promised

If a draft can't be checked because the record isn't there, the answer is no.

Habit three: a yes on money and promises

Some drafts can eventually go out on their own once you've watched them for a while, like a plain appointment reminder with fixed wording. Others never should. Write the never list down:

• Prices, quotes and discounts
• Refunds and credits
• Arrival times and completion dates
• Anything that admits fault or promises a fix
• Anything that mentions a lien, a lawyer or a late fee

For these, the agent drafts and the approver decides. That's the default we build with: on our trust page, sending is off unless you allow it, sensitive actions wait for approval, and every action is recorded.

The one-page checklist

Print this and tape it next to whoever runs the queue.

Before you approve any draft - [ ] Is this draft in my queue, with my name on it? - [ ] Did I open the record it's based on? - [ ] Do the names, numbers and dates match the record exactly? - [ ] Does it mention money, a time or a promise? If yes, am I the person allowed to approve that? - [ ] Would I be fine with the customer reading this out loud to me?

Every week - [ ] How many drafts did I edit? Which kind? - [ ] How many did I reject, and why? - [ ] Is any draft type always approved without edits? Consider fixed wording for it. - [ ] Is any draft type often wrong? Tell whoever set up the agent.

Every month - [ ] Is every draft type still assigned to a named person and a backup? - [ ] Is the never list still right? - [ ] Do we know how to pause the agent if something goes wrong?

How to keep the yes real over time

Approval fatigue is the slow failure. After a few hundred clean drafts, people stop reading. Three things help:

• Keep volume sane. If one person approves more than they can read in ten minutes, split the queue.
• Rotate the check. Once a week, the approver opens the record on every draft, not just the ones that look odd.
• Know where the off switch is. Anthropic's engineering guide to agents describes agents that pause for human feedback at checkpoints or when they hit blockers. Your office should know who can pause an agent and how. The EU's AI Act, which applies to high-risk systems, puts it plainly in Article 14: people overseeing an AI system should understand its limits, watch for over-reliance and be able to stop it.

Our human-in-the-loop page shows how the approval queue looks inside a Cortex build.

How to tell it worked

Track two numbers for a month: drafts edited before approval and drafts that went out with a mistake. A healthy queue has some edits, especially early, and close to zero mistakes reaching customers. If edits are zero from day one, nobody's really reading.

If you want someone to set up the queue with you, your free AI Operations Audit maps which drafts should need a yes in your company.