Five steps to make a company AI-first

Making a company "AI-first" doesn't take a platform, a budget, or an engineering team. It takes a structured set of files and a few rules about how AI is allowed to work inside them. We run a small menswear brand this way: AI handles the everyday running of the company while people own the product and every quality call. Our first product is a long-sleeve polo, made in Portugal. Here are the five steps that make it work, in the order we'd do them again, plus the one habit that keeps the whole thing from rotting.

Key takeaways

• "AI-first" is a structure problem, not a tooling budget. You can stand up a first version in an afternoon.
• Put the company's knowledge in one versioned repo, and put the rules right next to the work.
• Log decisions and current state, so neither a person nor an agent has to guess.
• Build your own skills so the AI's output comes back on-brand the first time.
• Keep a person deciding anything that carries weight, and keep the whole thing clean with AI, not by hand.

1. Make one repository: the company's brain

Put everything the company knows into our company OS: a single Git repository that holds it all, organised by what it's about (brand, product, makers, commerce, website, content, processes, decisions, tasks). One folder per domain.

The point isn't tidiness; the structure is how an agent finds its way. It opens a map file at the root, sees where things live, goes to the right folder, and arrives with the right context instead of guessing. The folders do for an agent exactly what they'd do for a new colleague: tell it where to look.

2. Put the rules where the agents work

A brain full of facts still needs to know how to behave, so the rules live right next to the work, not in someone's head. One entry file every agent reads first (what the company is, what it must never do, where to look), then local rules inside each folder, so an agent drafting a note to our makers already knows the constraints.

One rule earns its keep above the rest: agents add to the structure, they don't quietly reorganise it. Add, rename, or remove a file and the map gets updated in the same change. Skip that and the AI slowly drifts the company into a shape no one chose.

3. Write the decisions and the state down

If a decision, a fact, or the current state of things only lives in a chat thread, the AI can't use it, and neither can the next person. So we write it down: decision logs that capture the reasoning, not just the outcome; a single "where things stand" file; open questions kept as open questions instead of quietly guessed at.

The test we use: could someone (or something) read the repo cold and know where the company is? If yes, there's no private knowledge for the AI to be missing.

4. Build your own skills

A skill is just a saved, reusable instruction the AI follows: package a repeatable task once and it runs the same way every time. Ready-made ones are worth using; there's no prize for rebuilding what already works. The real payoff, though, is building your own: "draft a product page the way we do it," "write a reply to our makers," "log a decision." Each one carries your company's context, so the output comes back on-brand the first time instead of after three rounds of edits.

5. Keep a person in the loop

AI drafts; people decide. The agents produce drafts, summaries, and proposed changes; a person approves anything that carries weight: anything public, prices, commitments to our makers, and above all the product itself. It isn't distrust of the tools. It's the line that keeps taste and quality human, and the safety net for the plain fact that AI sometimes gets things wrong in ways that read as right.

The habit that holds it together: keep it clean, with AI

Five good structures still decay if no one tends them. An AI-run company generates files fast, and left alone it sprawls into a mess the agents themselves can't navigate. So we treat cleanup as ongoing work (prune dead files, fix stray structure, keep the map honest), and we do it through AI, not as manual housekeeping. The same agents that build the company keep it tidy. That habit is what keeps everything above actually usable six months in.

Just start

You don't need all five on day one. Make one folder for company knowledge. Add one file of rules for the AI. Write down the next decision instead of leaving it in a chat. Then let it grow.

The version that runs our company started exactly that small.

FAQ

What does "AI-first company" actually mean? That AI runs the everyday operating work (drafting, organising, summarising, tracking) while people own the product and every weighty decision. A way of working, not a product you buy.

Do you need to be technical to do this? No. The core is a structured set of files and a few written rules. We use ordinary tools (mostly Claude Code, with a bit of ChatGPT), and the first version takes an afternoon.

What should you not hand to AI? Anything that carries weight: public decisions, prices, commitments to your makers, and the product itself. Keep a person deciding those.

How do you stop an AI-run company turning into a mess? Treat cleanup as ongoing work and do it through AI too: prune dead files and keep the structure honest as you go, so the agents can always navigate it.

---

Start here: why we run a menswear brand with AI.