An AI partner and assistant for a company that isn't mine

Why I did this at all

I'll start with why, because in this story it's the only part that really matters. The technical half is the easy one, and the short one.

I have my own system. It works, I use it daily, and I've written more than one article about it. That proves nothing. A method that only works for its author isn't a method. It's a habit with documentation. The person who built the system always knows instinctively where to nudge it, where to look away, and which small undocumented step to take. Someone else doesn't know that. And shouldn't have to.

The second reason is the company's, and it's far more concrete. In a small business almost all the knowledge lives in one person's head and in two years of message threads. Where that invoice is. Why we stopped working with that supplier. What we promised the client in spring. What the last delivery cost and why it came in higher than planned. Every one of those questions runs through one person, and it's usually the same person who writes everything down in the evenings. When they're away the company doesn't stop. It slows down quietly, and nobody can point to where.

The third reason is uncomfortable, which is why I'm saying it out loud. I teach this. If I had only ever built one for myself, I'd be selling a method tested on the friendliest user in the world. Me. That isn't proof. That's self-service.

A company doesn't start trusting AI because it's clever. It starts when the AI knows what happened last Thursday.

A tool waits. A partner speaks first.

AI usually enters a company as a tool. Someone opens a chat, types a question, gets a smooth answer, copies it and closes the window. Next time it all starts from zero. That isn't bad, it's just a calculator. A calculator waits to be opened.

A partner speaks first. It says this client has asked the same thing three months running. That the invoice you're about to send doesn't match what was agreed in March. That the process you're rebuilding from scratch was already written once and lives in a folder nobody opened.

The difference between the two isn't the model and isn't the price. The difference is context. The tool has none. The partner does.

Technically it's a second brain. Inside a company I don't call it that.

What we built is, structurally, exactly what I built for myself and described in my earlier article on second brains. Same layers, same files, same logic.

But in a business conversation I don't use that phrase, and there's a reason. "Second brain" is a personal development term. It sounds like something the person needs to fix about themselves. In a company the first question isn't how do I think more clearly. The first question is who does this when I'm not here.

So inside a company I use two words. An assistant does what you tell it. A partner tells you what you didn't ask. One system holds both. Most people start using it for the assistant and stay for the partner.

The entry point wasn't strategy

I expected we'd start with marketing or content. That would have been my comfortable corner. We started with bookkeeping, and it turned out to be a better choice than the one I would have made.

Three reasons.

One, the material already existed. Invoices, statements, spreadsheets. Nobody had to create anything new just because we were building a system. If step one requires writing a hundred pages of fresh content, step one never happens.

Two, it hurt every week. Not once a quarter. A weekly rhythm is the only thing that can carry a new habit.

Three, the answer could be checked. The number either matches or it doesn't. There is no room there for being impressed by how smoothly the AI writes.

The third one matters most, and I'll state it as a rule. The first job an AI does inside a company must be one where a wrong answer is immediately visible. Start with writing text and you notice the wrong answer a month later, by which point trust has been built on nothing. Start with numbers and you see the error the same day, and the trust that follows is real.

What transfers and what doesn't

Three things transfer. The five-layer structure. The daily loop, where something happened, it got written where the next session sees it, and it changed the next move. And one small file saying what matters right now, alive enough that the system argues back when you propose something new.

My folders don't transfer. And that's the first thing people want to copy. A structure whose names the person wouldn't have chosen stays unopened. I tested this with my folders and my vocabulary. It didn't work. It had to be built from their documents and their words, including in the places where I thought my name for it was better.

Discipline doesn't transfer. This is the hard one. I can build the system, I can teach it, and I can stay close through the first weeks. I cannot hand someone a habit. It's also the only place where this project can genuinely fail, and technology has nothing to do with it.

The five skills it stands on

The whole method is public and free, CC licensed, in our AIOS repository on GitHub. There are twenty-seven skills there, but a company partner stands on exactly these five. Each one below has a direct link, one sentence on what it is, and one do and one don't.

1. constitution7 — the file your AI reads before every conversation

This is the foundation everything else attaches to. Without it the other skills are methods with nowhere to live.

Do: write who the company is, what it sells, how it speaks to a client and what it never does. Every line has to change the AI's behaviour. A line that changes nothing is decoration.

Don't: put this month's projects in it. Those change constantly, and a core file that has gone stale is worse than not having one. Projects live in the focus file.

2. memory7 — the folder shape, built from the company's own material

This decides what folders exist at all and what belongs in each one.

Do: start from a pile of real documents, not a blank page. The structure has to be found in the material, not invented ahead of it.

Don't: copy someone else's folder structure. Including mine. It's the fastest route to a system nobody opens.

3. meeting7 — a conversation that becomes a record instead of a memory

In a company, knowledge is created in conversations and lost in the same conversations. This skill pulls out of one meeting what was decided and what is now owed by whom, and out of many conversations with the same person, a profile of them.

Do: run it right after the call, while the context is still in your head. Two minutes now beats half an hour a week later.

Don't: turn it into minutes. Nobody reads minutes. Only two things are needed: what was decided and what is owed by whom.

4. focus7 — one small file saying what matters right now

Once a week it gets rewritten from real data, and the owner approves the changes one at a time.

Do: keep it short enough to be cheap to load in every session. This is the file the AI always sees.

Don't: let it become a project archive. Once it grows it stops being focus and becomes a list nobody reads.

5. exit7 — closing a session so nothing survives only in the chat

At the end of the day the system finds everything that exists only in the conversation and hands it back as ready-to-approve changes. Not as questions.

Do: end the working day with it. Two minutes.

Don't: skip it. This is exactly where a system stops being a folder of notes. Without it the best thinking of the session leaves with the window.

And one thing that isn't on this list: a suggestion to install them all. The repository carries a rule that you keep no more than three standing skills until the loop has closed at least once. The first two here are one-time builds; they lay the foundation and then go quiet. The standing ones are the last three. Exactly three. In the personal version the capture slot is held by in7, a daily log in one messy sentence. In a company that slot belongs to conversations, which is why meeting7 stands there instead.

The fourth layer: what it is connected to

The five skills above are the method. But a partner that can only read its own files hits a ceiling fast inside a company. Asked how we are doing, it answers with an opinion, because the numbers live somewhere else.

So running in the background there now are connections to the social advertising account, the short-video platform, the website analytics and several research tools. The practical difference is one thing, and it is a big one: the answer about a result arrives as a number from the account rather than a feeling about how it seems to be going.

Those skills are not public yet, which is why this is a description and not a link. They contain account numbers, access paths and things that belong to this company alone. Once that has been stripped out, they go into the same repository as the rest.

And one note that matters more here than enthusiasm does. The more tools are connected, the more places there are where a mistake costs money. In an ad account a wrong move isn't a badly written sentence, it's a spent budget. So the order of connection is the same as everything else: reading first, writing only after that, and writing only with a human's explicit approval. For the first month the system was allowed to look and nothing more.

Where it stands today, honestly

The technical layer is further along than I expected. The human loop is not. I say this because showing a finished system is easy, and only an honest status is useful to anyone.

Working: it answers from its own memory rather than from general knowledge. It has its own core file written in the company's words. It has its own channel on the phone, so it can be messaged from the car. It is connected to advertising and analytics data. And it has already surfaced things the owner had forgotten.

Not yet working: the daily ritual hasn't held. It gets opened when there is a specific need, not every day.

That's the same wall almost everyone hits, including very capable people who built their system themselves. Plenty of people manage to build one. Nobody has a reason to come back to it without a daily ritual. And that isn't a technical problem. The system is ready. The habit isn't.

If you want the same in your company

1. Pick one process that hurts every week and where a wrong answer is immediately visible.

2. Gather the real material that already exists about it. Write nothing new.

3. Write the core file and the folder shape from that material, in the company's own words.

4. Use it for two weeks and add nothing.

5. Only then add the second process.

Step four is the one everybody skips. It's also the only thing separating a system from one more folder full of good intentions.

Egils Boitmanis is a business consultant for founders and experts with 20+ years of experience in business, digital marketing, training and practical AI use. A Google Certified Trainer, he runs hands-on AI workshops for teams and is the author of the AIOS method.