Building Your AI Organization

The French version of this article is available ➡️ HERE ⬅️

What building with AI taught me about rules and what I should never delegate.

Although I work with a lot of technology tools in my day-to-day, in my own business I wasn’t using AI tools daily, and frankly…I had no plan to start.

I was too busy delivering for clients so I did not feel that I had enough time to tinker with the 10th “new” model or tool in the last 2 weeks. You could say that I had AI fatigue.🥲

Until I decided to apply to myself what I recommend my clients to do: “one day, you’ll have something to do and instead of doing it like you’ve always done it, I’m asking you to use this new tool to do it – Give it a try”.

Fast forward to months later, I run my projects with a dozen AI agents [moving forward, I’ll simply call them “agents” for brevity], various skills and a set of rules.

None of it was designed in advance. I built it gradually whenever there was something I needed to do.

So, if you’re circling AI without quite touching it, this one is for you. 🧭

What did I actually do first?

I built something I could not have coded alone. I mean, I can’t code to save my own life!

So I started from a product idea; a web app that would help me and a couple of friends find clothes that suited our color palette and our silhouette, and that were in stock in our size at specific online retailers. This turned into a working web app with AI doing the coding. 

I was blown away!🫨

In a few hours (ok, a whole weekend), I had a clickable interface I could show people so that they could engage tangible with my vision. All without having to hire a developer. 

There’s nothing wrong about hiring a developer! It’s that my ambition with this project was straightforward: I wanted to get one thing done, and I was willing to be a beginner while doing it.

Before anything else, do this ⬇️

If you’re hesitant, here’s my first advice.

• Do not pick a tool and hunt for a use case. Look at the frictions and pain points you already encounter in your daily life.
• Do not write a strategy to get started; planning is an excellent form of procrastination! Instead, get started by sharing your request to the AI tool of your choice, and enjoy the ride!

All in all: I picked one real task and got started.

One key rule I haven’t changed in over 9 months: I run every terminal command myself. 

• Not because AI is careless. But because, for now, I want my hands on whatever executes on my machine. 
• Also, AI can get weirdly fixated on the wrong things. So it helps in preventing too many tangents.

It’s a small preference, yes, but I wrote it to the (markdown) file of my ways of working – which every agent has access to, and should read at the beginning of each session when it is invoked for the first time. It ensures that every future session will follow these instructions without me having to remind the agents. 

I didn’t know it yet, but providing thorough feedback to agents and writing my comments (corrections, adjustments, reasoning) to memory was the standard operating procedure for everything after.

What changes when you ask AI to disagree with you?

A few weeks in, I asked for a co-founder who would not flatter me.

I created Optimus, first as a revenue-focused co-founder whose brief was to flag gaps in my thinking and structure. Optimus is technically a Chief of Staff, but I wanted to be fancy, okay? 😌

In the same stretch, I wrote my ground truths into memory: who I am, my ventures, my 2026 goals, my financial targets. From then on every session started from facts instead of from zero.

And Optimus helped me a lot! It still does to this day.

• It stress-tests all of my plans against my goals and my capacity,
• It continuously reviews my projects against my priorities, and;
• It NEVER hesitates to call me out when I am trying to work on deprioritized or descoped tasks.

Overall, Optimus caught many errors or assumptions in my planning which forced me to be honest with myself in terms of what I would be able to accomplish.

From one AI agent to an AI organization

At some point I stopped prompting and started naming roles for each repetitive set of prompts.

• Optimus became my Executive Integrator, with a COO, a product lead and a CFO reporting to it
• Then a Chief AI Officer for token governance joined, followed by a six-agent go-to-market team. 

I named a scope for each: what each role owns, and what it explicitly doesn’t. For instance:

• my proposal writer never sets prices
• my pipeline tracker measures but never decides.

One detail I like: my goal wasn’t to create an agent, my goal was to package roles and responsibilities in a way that made them easier to use. 

Indeed, before my outreach writer existed as an agent, I used its role description as a prompt and produced 44 personalized outreach drafts. The job description was the product. The agent was packaging.

Then I defined a standard for every agent because I wanted them to manage themselves systematically with a feedback loop. So each agent has to:

• Keep self-notes. Write down what it learns about how I work.
• Study my edits. If I rewrite its drafts, it has to analyze the edits and find out why.
• Skip details that do not correlate to value. That one sentence prevents a lot of slop!
• Draft only. No agent sends, publishes or signs anything.

That last line is not a safety footnote. It is by design. I own revenue from start to finish. What I handed to the agents was the preparation. The selling/actual outreach stayed with me.

The trap of getting lost in the tinkering

When my Chief AI Officer ran its first review, it highlighted that roughly 90% of that week’s AI usage went into managing the org.

That was a shock, to say the least! 🫠

It was also a sobering reminder that structure can pass for traction. Optimus repeated the same warning when the token audit got to it: “MC, an org chart is not a pipeline”. I know, I know…The endless tinkering is so tempting, with its endless hits of dopamine from making something work.

To fight this, I did what I’d ask of any client: I put a budget and a boss on it. In my case, it became a weekly review of consumption by agent and by use case, saved as a reusable workflow, with Optimus signing off on the recommendations, and Jarvis (the COO) updating agents’ work approaches accordingly.

Sexy? No. Useful? Very.

Golden Rules: What are They?

This is the section to screenshot.

A golden rule is a rule I write down after a specific incident or a very detailed feedback, with the reason it exists and when it applies.

It’s not a vague principle, but a rule with a backstory.

My belief is simple: when an AI agent gets something wrong, it’s just as important to fix the output as it is to teach the AI agent why it was a mistake in the first place.

Because you want to ensure that the AI agent learns from the experience and applies your feedback moving forward.

My Golden Rules

I keep mine in a table with three columns: the rule, what caused it, and since when. Every row has a backstory. Here are five.

1️⃣ Read Drive links with Drive tools, never by fetching the web page. Backstory: an agent reported “no access” when it was just using the wrong tool. Wrong tool, not locked door.

2️⃣ Update a project’s source-of-truth file in the same session as the decision. Backstory: a stale source of truth misled every session after it and it took me close to an hour to figure it out.

3️⃣ No contract clauses for problems four years out. Backstory: a proposed “stop-loss” clause for a problem years away kept getting the agent stuck in its work, and had me repeating the same thing over and over again. 

4️⃣ Verify every citation. A confident answer is not a sourced one.

5️⃣ Spreadsheets carry live formulas, never pasted values. Backstory: a values-only compilation of numbers without formulas is a status artefact I can’t update myself.

6️⃣ My voice is the default. I never have to ask for it. Backstory: agents wrote external-facing text as if my voice were optional.

Notice the pattern: 

• Few of these are really lessons about AI ➡️ Each is a lesson about my own operation that the AI happened to expose.
• When I started doing this, my fixes were chat messages. ➡️ By the time I took stock, they were memory files: 28 of them, each rule carrying its Why and its How to apply, plus a source-of-truth file at the root of each project and a working-approach file for each agent. The mistake still happens once. It just stops happening twice.

Who does what, and which acts will I never delegate?

Let me answer the question you may already be asking: Yes, agents draft content in my voice – I built several writing skills for different types of writing. 

What the agents mainly do is help me to NOT start from a blank page.

• I re-read and edit everything myself. 
• An Editor in Chief agent reviews every piece of external content before it reaches me, and I am the one who publishes every article. Including this one.

That’s the pattern across the whole system ➡️ AI takes on more and more of the preparation, I remain responsible for turning the prepared assets into income-generated assets.

The reason is plain:

I, the human, delegate the preparation but I keep the accountability.

(And no, even with the most specific Skill, I believe that a draft in my voice is not the same as a text written by me. That’s why the “MC edit” step exists.)

The 3 Steps of AI Adoption

I’ve named three stages in AI adoption, and I’m only fully at the first two:

• Delegate: AI does the task.
• Codify: AI does it the way you want, every time. But you have to prompt it.
• Compound: AI does it on a schedule, from live data.

I’m proficient at delegating and good at codifying. I have not compounded yet. 

If you hear someone describe their AI setup, ask them which of those three they mean.

My next steps are modest and specific. 

• Give each agent’s output a place in the pipeline (sent, reply, call, proposal, won) so “useful” means something countable. 
• Schedule weekly reviews for some agents. 
• Connect another system of record. 
• Move AI from drafting into delivery. 
• And once a quarter, review every agent as I would a team: keep, merge or retire.
  • Some of my agents haven’t added anything in their logs in months.😅

So where do you start?

Not with “adopting AI.”

Start with my AI Fluency Quiz to understand where you stand: https://tally.so/r/KYVYrM 🤓

Then, follow this plan:

1. Pick one task that is blocking your productivity right now. 
2. Get it done, even if you feel like a beginner while you do. 
3. Give that bottleneck a name, a scope and a boundary. 
4. When the AI gets something wrong, fix the output, then write the golden rule with its context. 
5. Repeat.

That’s it. 

That is MY entire method. 

The only thing I’d add is that the hesitation was never the real obstacle. Finding a first meaningful task to outsource was.

TL;DR – Building Your AI Organization

• How it went: one app first, then a review, then an org of AI agents, then governance (written process and the usage of my own voice). Each step answered a problem that occurred in the previous step.
• The method: pick a job that blocks revenue, name the roles, keep agents on drafts only, and apply the Golden Rule: every correction becomes a written rule with its reason, because you’ll have to repeat a correction that you made only in the chat.
• What’s still open: more automation, add live data to my AI workflows, more AI in delivery/execution and outcome measurement. I’ve delegated and codified. I haven’t compounded.

So, which bottleneck will you tackle first?

If this was useful, subscribe to get the next one. 

And if you know someone still circling AI without touching it, send it their way 😊