On the first Monday in November, my morning list had 38 items waiting on me. Working through them took 18 minutes.

I run a design and automation practice for about two dozen client organizations. An AI system I built does most of the day-to-day operations, working from written rules. It reads the inbox, keeps each client’s record current, drafts the replies and checks the websites. What it does not do is decide what lands on my list. Since September 25, it asks first.

There is a lot of advice right now about keeping “a human in the loop” with AI. Most of it stops at the phrase. This is what my loop holds, and what it has cost me.

// WHAT IT ASKS

What it has to ask about

The rule is short. Before the system does any of these, it puts a numbered request on my list and waits:

• Anything new a client will see. A reply, a change to their website, a date we promise them.
• Money. A price, an invoice, any new spending. It can prepare an invoice. It cannot send one.
• A change to what we agreed. Scope, plans, deadlines.
• Taking away anyone’s access. A login, a group, a page someone used to reach on a client’s system.
• New work for me. A task on my calendar that I did not already approve.

Each request arrives with the system’s recommendation and a real example: the actual record it would change, or the actual before and after. That is why a yes takes seconds instead of a meeting.

Why ask at all? I wrote the reason into the rule on the first day: I want to approve what lands on my list, once a day, rather than find it there.

// WHAT IT DECIDES

What it decides by itself

Everything about how it does its own work: how it reads a file, how it checks a website, how it formats a report, what order it works in. When it changes one of those rules, it writes the change down with the date. Every Monday I get the list of what it changed that week, and I can undo any of it that day.

A mistake does not need my approval to get fixed. When something goes wrong, the fix goes into the checklist it belongs to, so the same thing does not happen twice.

// THE MIDDLE

Where the line is not obvious

Not everything a client sees waits for a fresh yes. A weekly status note goes to clients on Fridays, and site health reports go out on their own. I approved the format once. The system fills it in from the record.

The sentences in that weekly note are written by the system, though. So the day before it goes out, I get a list of every sentence it wrote or changed that week. I do not have to act for the note to go. I can strike anything on the list.

That is the real shape of the rule. It is not “the AI never talks to clients.” It is this: anything new waits for me, and anything routine shows me what changed.

// THE RECORD

What the asking has looked like

Between September 25 and the start of November, the system made close to 500 numbered requests. Of the roughly 400 still on file, I said no to 8. It withdrew 21 more itself, mostly because the question was already settled by the time I got to it.

Eight out of 400 looks like an argument for dropping the step. I think it is the opposite.

The no’s are the ones that matter. In October the system asked whether “do what you can” on a publishing task should count as my yes to publish a post. Its own recommendation was no. I agreed. One yes a month is not worth loosening the rule.

Silence is never a yes. An unanswered request stays on the list. After five business days it moves to the top, marked with how long it has waited. Nothing gets decided because I was busy that week.

// YOUR LIST

Write your own list

You do not need my setup to use this. If you or anyone on your team has started letting an AI tool draft emails, answer customers, post for you or touch your books, write down the short list of things it must ask about first. Then answer three questions:

1. What can it send to a customer without a person seeing it first?
2. What can it spend, or promise, on your behalf?
3. When it changes how it works, who finds out?

If you can answer all three in a sentence each, you have a loop. If you cannot, the tool is deciding things you think you are deciding.

If you want help drawing that line for your own organization, get in touch. It usually takes one conversation to write the first list.