The company brain: what your AI needs to know first

Gotcha, the fictional rental company with 50 vehicles from part 1, tried again. An AI assistant wrote an article about renting a van for a house move. It read well. In the third paragraph it said Gotcha delivers the van to your door for free, Sundays included.

Gotcha doesn't deliver. Not for free, not on Sundays, not at all.

The AI hadn't lied in the human sense of the word. The model simply didn't know, and filled the gap with what rental companies sometimes do in general. That's why a harness doesn't start with the agent that writes, but with what that agent reads first: the company brain. AI without a company brain guesses. With a company brain it knows what's true, and what your business stands for. That second part matters just as much: it's the difference between text that fills a page and content that gives customers something.

Why AI makes things up about your business

A language model is trained to give an answer, even when it doesn't know the answer. In September 2025 OpenAI explained why in the research Why language models hallucinate: the way models are trained and evaluated rewards guessing over admitting they don't know. Rare facts, such as what one small rental company does and doesn't offer, can't be inferred from patterns.

For a business owner the consequence is concrete. What your AI says about your business, you say. In February 2024 Air Canada had to compensate a customer because the chatbot on its own website had given a wrong rule about bereavement fares. The argument that the chatbot was responsible for itself didn't convince the Canadian tribunal (summary by the American Bar Association, February 2024). This isn't legal advice, but the lesson is clear: a promise made by your AI is a promise made by you.

A company brain is the remedy. It's one place that says what's true about your business, approved by you. Every AI agent reads it before writing anything, and uses only what's in it.

What goes into a company brain

A company brain isn't a folder full of old documents. It's a short, approved set of answers to seven questions.

• Facts. What you offer and what you don't, your locations, opening hours, terms for deposits and excess.
• Claims with a source. Every number or promise you put out, with where it comes from. No source, no claim.
• Promise and common thread. What you promise customers, in one sentence, and the two or three messages that come back in everything you publish. This is why a customer rents from you and not from the company down the road.
• Tone. How you sound: businesslike or casual, formal or informal, short or long sentences.
• Banned phrases. The wording you never use, and promises you can't keep.
• Audience. Who you write for: the private customer moving house or the contractor who rents five vans a week.
• Voices. How the people who post on behalf of your company sound. The owner writes differently from the planner.

What we built at RentVisie

Our own brain works on one rule: what's in it is approved, and every agent reads it first. The agent that writes an article every week, the agent that proposes FAQs and the agent that rewrites page titles all start from the same knowledge. When something changes, you change it once, and from then on everything uses the new version.

A new item first comes in as a draft. Only after approval does it count. In the two weeks from 25 October to 7 November 2026 the owner approved 21 versions of parts of the brain. On 9 November 2026 it holds twelve phrases we don't use, or only use with a source.

The most important part is the check on numbers. Every claim has a source and a status. An article with a number the owner hasn't confirmed yet can't be published; the button stays grey. The numbers in this article went through that check too.

To be fair: our brain isn't finished. The list of confirmed claims is short, and the description of all our product modules isn't in it yet. That doesn't make it less valuable. What is in it is correct.

Gotcha builds its brain: a template to fill in

Part 1 ended with homework: the 20 questions customers always ask you. That list is the start of your brain. Gotcha fills in the template below, and you can do the same in an ordinary document.

• The 20 customer questions, answered once. How much is the deposit? What's the excess? Can I pick up on Sunday? What happens after damage? Write each answer the way you'd give it at the counter.
• Ten facts. For example: "Gotcha doesn't deliver; pick-up and return at the branch." Short, checkable, no adjectives.
• One promise. Gotcha picks: "With us, you know upfront what you'll pay." The common thread that goes with it: no surprises at the counter, not at pick-up and not at return. Choose a promise you keep every day, not the nicest one you can think of.
• Five banned phrases. For example "free delivery", "always available" and "the cheapest in the region". Anything you can't deliver or can't prove.
• Two voices. Three messages the owner wrote, and three from the planner. That's where AI learns how each of them sounds.

Be honest with yourself: the first time, this takes more than the two hours a week from part 1. Especially if your company's knowledge isn't written down anywhere yet and lives only in your head and your planner's. Count on a few evenings. After that it's maintenance: a new fact in, an old fact out.

The same request, without and with a brain

A brain doesn't only prevent mistakes. It also decides whether AI content makes any sense. Gotcha gave the same AI assistant the same request twice: "Write a short LinkedIn post about renting a van for your house move."

Without a brain:

Moving house soon? Then a van is a must! At Gotcha you rent a van easily and affordably, with flexible pick-up times and service that makes the difference. Your move will be a breeze. Book today!

Nothing in it is provably wrong. But any rental company in the country could have posted it. "Affordably" and "flexible" can't be checked, the exclamation marks and the sales voice aren't how Gotcha sounds, and a reader has no reason at all to rent from Gotcha afterwards.

With a brain:

Long moving day and no energy left for the petrol station? Just bring the van back.

Gotcha refuels it for you. You pay the price per litre that's already in your confirmation, with no service charge on top. Rather fill it up yourself? Of course you can.

What do you most like to skip after a moving day?

Same AI, same request. The difference comes from three parts of the brain. The facts: Gotcha refuels at a fixed price per litre, with no surcharge. The promise: knowing upfront what you'll pay, so the post says where that price is. And the tone: informal, short sentences, one question at the end.

The AI chose the angle itself. It doesn't shout that Gotcha is transparent, but starts from a moment every mover knows: tired, van full, tank empty. The promise comes after that, as the solution. Creativity without direction produces text that fits anyone. With a brain, the AI has something to build on.

The same goes for ideas. Ask for five blog topics without a brain and you get "5 tips for a stress-free move", just like every competitor. With a brain, Gotcha got ideas such as "Refuel yourself or have it refuelled: which is cheaper?" and "Five questions about costs to ask any rental company upfront". Both useful for the customer, and both proof of the promise. That's how the common thread becomes visible without you having to explain it every time.

What a company brain earns a rental business

The return sits in four places.

Content that makes sense. Posts, articles and newsletters that show why a customer rents from you, instead of text that fits any rental company. Sharper for you, more useful for the reader.

No wrong promises. An AI text that promises free delivery costs you an argument at the counter or a customer who never comes back. With a brain, it only says what you actually do.

Not explaining the same facts again every week. This is a worked example. Say someone in the office spends fifteen minutes every working day rewording the same answers and quote texts, or correcting what AI made of them. Over 48 working weeks that's 60 hours a year. At €60 an hour that's €3,600. Using €100 per rented day as a hypothetical example, as in the series More revenue from the same fleet, that equals 36 rented days.

Faster answers, as a side effect. A quote or an answer to a customer question is written faster when the facts are ready, by AI or by a new employee. Whoever answers faster wins the booking more often.

How to start this month

You don't need special software for it.

1. Take your 20 customer questions from part 1 and write an answer of at most three sentences for each.
2. Write down ten facts, your promise and five banned phrases in the same document.
3. Put the document into your AI assistant as fixed knowledge. In Claude you can do this in a project with its own knowledge and instructions, on the free plan too (Anthropic help page, checked on 9 November 2026). Other assistants offer something similar.
4. Test it with three questions you know the answer to, and one where the answer should be "we don't do that".
5. Agree who maintains the document, and when: for example every first Monday of the month.

What you don't leave to AI: deciding what's true

AI can help fill your brain. It can read through your customer emails and suggest which questions keep coming back. It can make your answers shorter and clearer.

But what's true about your business, you decide. Whether you're open on Sunday, how much the deposit is, which number you put out: no model can decide that for you. The same goes for your promise. AI can come up with ten ways to word it, but which promise you keep every day is your choice. A brain that AI filled by itself without anyone checking it is exactly the guessing you wanted to avoid, just written down more neatly.

What you have now

After this part you have, like Gotcha, a company brain: a document with your answers, your facts, your promise, your banned phrases and your voices. It isn't a system yet, but it is the base every next part reads from.

In part 3 we build the first agent that does a task on its own every week, with this brain as its first reading and one fixed moment when you approve. All parts are on the page of the series Making money with AI.

Don't want to set this up yourself, but do want the work done every week? That's what RentVisie's digital marketing for rental businesses is for.

Frequently asked questions

What is a company brain?

A company brain is one approved place for what's true about your business: your facts, your claims with their source, your promise and common thread, your tone, the words you never use, your audience and the voices of the people who write on behalf of your company. Every AI agent reads it before writing anything, so it doesn't have to guess.

Does a company brain also help with better marketing content?

Yes. Without a brain, AI writes text that fits any business. With your facts, your promise and your common thread, AI can come up with creative posts, articles and topics that show why a customer buys from you. The AI stays creative, but builds on what you actually do and promise.

Why does AI make things up about my business?

Because a language model is trained to give an answer even when it doesn't know. In September 2025 OpenAI described how training and evaluation reward guessing over admitting something is unknown. A model can't infer facts about one small business, so it fills the gaps with what's true in general.

How long does it take to build a company brain?

A few evenings the first time, especially if your company's knowledge isn't written down anywhere yet. Start with the 20 questions customers always ask you, ten facts and five banned phrases. After that it's maintenance: a new fact in, an old fact out.

Do I need special software for a company brain?

No. An ordinary document is enough to start. Most AI assistants let you add fixed knowledge and instructions, in Claude for example in a project, on the free plan too (checked on 9 November 2026). What counts is that someone approves and maintains the content.