Five Things a Real AI Board Has That a Chat Window Doesn't

If your AI only works on the days you sit down in front of it, you do not have a system. You have a very good conversation partner with amnesia.

I run a board of narrow AI seats on one of my own businesses. It has run every week since June. Below is the sixty-seven second version, and then the five things underneath it.

Watch

1. A charter

A written job. What the seat owns, what it decides on its own, and what it has to bring back to you before acting. Not a prompt you retype and slightly reword every time.

The line that does the most work is the one people skip: how it decides when two reasonable options conflict. Without it you get an assistant that agrees with whatever you said most recently.

2. Floors, not targets

Numbers it has to clear, written down before the week starts, and scored honestly after.

A target is something you look at afterwards and reinterpret. A floor is a number you either cleared or did not. The difference sounds like semantics until the first week you miss one, and the board says so in writing instead of finding a reason the week was actually fine.

3. A schedule

It runs weekly whether or not you open anything. No tab, no prompt, no you.

This is the step that turns AI from a tool into a system, and it is also where the quietest failure lives. A scheduled job that stops does not error. It produces nothing, and nothing looks exactly like a quiet week. I wrote about that specific failure and the check that catches it separately, because it cost me forty days.

4. An archive

Every meeting it has ever run, dated, on disk.

Not a chat history you scroll. Files you can open, diff, and search. The archive is what makes the fourth month different from the first: each run can read what the last one said, score the floors it set, and notice that a number has been sliding for three weeks. A chat window starts from nothing every time, which is why it can never tell you that.

5. A meeting you can act on

Real numbers from the real business, one headline, and the calls it is making this week.

The most common way this whole setup fails is not a bad model or a broken script. It is that the board produced something reasonable and nobody did anything with it. So the output is shaped like a meeting: here is what happened, here is the one thing that matters, here is what I would do. Then a human decides, and logs what would reverse the decision.

Five files, one board

That is the entire trick. A charter, a goals file with floors in it, a schedule, an archive folder, and a meeting format. None of it requires you to write code.

The board does not act. You do. It just makes sure that on the Tuesday you are buried in something else, the work of noticing still happened.

If you want to build one, the course walks through it from an empty folder. If you would rather build yours in a room with me and a small group working on your own numbers, that is the Build Lab.