Do we need our own AI? An AI chief of staff, in plain English

Do we need our own AI?

No. Training an AI of your own costs millions, takes years, and you would end up with a worse version of something you can already rent by the month. What most businesses actually need is not an AI of their own. It is their own way of running one: set up inside an account they already pay for, with their name on the door, their rules inside, and a named person who answers for it.

That is a design job, not an engineering job. I call it agentic design, and the rest of this page is what it means.

What is an AI agent?

An AI you give a job to rather than a question. Ask the chat window something and it answers. Give an agent a job and it reads, searches, drafts, checks and files, step after step, without you typing each one. “Agentic” just means it acts in steps.

Here is the part the brochures leave out. Everything I have written elsewhere on this site about AI making things up still applies, and now it applies to every step nobody read. An agent that invents something on step three and builds steps four to nine on top of it is not more useful than the chat window. It is faster at being wrong.

So the question is never “which agent?” It is “what is the room it works in, and who reads what comes out?”

What is an AI chief of staff?

Not the crew. The one that holds the question. I call it Agent 0.

I run my own work this way. Mine is called Bob. Bob does not decide what my books or my papers say. Bob runs a small crew of AI helpers, each with one job: define the question, dig, cut what matters from what doesn’t, record it, write, tear the work apart and check it, make it ready. Then Bob hands the result to me, and I decide.

The crew is plumbing. Plenty of people can build a crew. Agent 0 is the thinking end: the one that knows what the job is for, notices when the work has drifted, and knows what it must never do. That is the part I design, and it is the part almost nobody sells.

It is also what I mean when I say AI stands for Additional Intelligence™, a phrase I have used since April 2026. More intelligence, added to yours. Never instead of it.

What is a cutting room?

The room the crew works in, and the order it works in. Nine steps, and no job skips one. Nothing goes out on its first draft.

1. Define. Three questions before anything moves: what will be different, for whom, and how we will know. No answers, no job.
2. Dig. What is actually there, brought back with its address: the source, the figure, the date, who said it.
3. Cut. Keep the few finds that carry weight. The rest is set aside, not lost.
4. Register. Grade each find: evidence, interpretation or guess. All of it is recorded; only the first kind is built on.
5. Write. A draft in your voice, for a person to take over, not to forward.
6. Tear down and check. A second reader that did not write it looks for the weak claim, the missing source and the five ways AI goes wrong: smoothing, hallucination, affectation, drift, sycophancy. That check is SHaDS®, and it never checks its own work.
7. Make ready. Read where it will actually be read: on a phone on the train, on paper in the meeting.
8. The pause. A person reads the whole thing and sits with it until it is theirs. The pause is built in, not a hold-up.
9. A person signs. And decides what happens next.

The last two steps belong to people, every time. AI does the work. People publish. Nothing is sent, posted or published on the agent’s say-so, and it does not speak to your customers.

How do we set up AI in our business?

Inside your own business account with the maker. Mine runs on Claude, made by Anthropic; yours may already be somewhere else. You administer it, you pay the maker direct, and, on a business plan, your data stays yours and out of training by default. Read that line in your own contract before anything else. Nothing of mine is installed, and nothing of yours comes to me.

Then we design the chief of staff and the room together: what it is for, what it may touch, what it must never do, who owns it day to day, and who signs. It is switched on one job at a time, on your own work first. It touches client work only after it has earned its place on yours. I come back from time to time, and what has not earned its place is switched off.

Your people learn to run the room as it is built. That is the point. In a year it should be yours, not mine.

What I don’t do

I don’t write software, host anything, or carry your production systems. If your plan needs code written and servers run, you need an engineer, and I will say so on the first call. What I design is the room, the seats and the rules, and I teach the people who will sit in it. I take on a handful of these at a time, because each one gets read by me.

Who signs it off?

A named person, every time, who has read the whole thing. Not a department, not a policy, not the machine.

This is the step every AI programme I have seen skips, and it is the one that decides whether the thing is safe. The pause is where the judgement lives, and it is the one part of the room I would never let the machine near.

Where the proof is

This website. Bob is its curator, and no page goes on it until I have read it. I have written 14 million words with AI since the week ChatGPT arrived in December 2022 (counted 30 September 2026), I kept the record, and the room is what that record taught me to build. If you want to see the check itself, it is free to use.

Try the SHaDS checker Talk about your own room