AI Knowledge Base: The Twelve Things Every Company Brain Should Hold

An AI knowledge base for a business should hold twelve kinds of canonical knowledge: company facts, product facts with approved claims, objections, customer pain points, use cases, customer voice, the ideal customer, brand voice, employee voice, competitors, brand visuals and the product interface. Start with product facts, objections and the ideal customer.

What should an AI knowledge base hold?

It should hold the facts every piece of work depends on, written as short, sourced statements with an owner. For most businesses that means twelve sets covering the company, the product, the customer, the voice and the market. Each is canonical: one version of the truth that every AI reads instead of the raw pile of documents.

The difference from an ordinary knowledge base is purpose. A help centre or wiki is written for people to browse. A company brain is written for AI to read before every task, so it favours short, specific, sourced statements over long pages.

The gap it fills is clear from how people rate their tools. According to Atlassian's 2026 Pulse survey, only 12% of workers say their AI understands their organisation like a tenured senior colleague. These twelve sets are what that colleague knows.

Which sets describe the company and product?

Three sets cover what is true about the business: company facts, product facts with approved claims, and use cases. Together they stop AI inventing features, quoting wrong details or describing what customers do from guesswork. The product facts set also records which claims are approved for public use.

1. Company facts. Team, history, locations, customers you may name, and how the company describes itself. Source: leadership and existing materials.

2. Product facts and approved claims. What the product does, feature by feature, and which claims may be published. A feature in testing is true but not approved. Source: product team, documentation, pricing.

3. Use cases. What customers use the product for, each marked as told by a customer or told by staff. That label matters: a use case only staff describe should be written as a capability, never as what customers do.

The approved-claims flag is the single most useful control in the brain. It is what stops an AI cheerfully announcing a roadmap item to a prospect.

Which sets describe the customer?

Four sets describe the people you sell to: the ideal customer profile, customer pain points, prospect objections with your best answers, and a customer voice quote bank. These make AI output specific to your buyers, in their language, addressing what actually holds them back.

4. Ideal customer profile. Who buys, who does not, what tips them into buying. The guide to ICP documents for AI covers the format.

5. Customer pain points. Problems customers describe, in their own words.

6. Prospect objections. What buyers push back with, and every answer your people have given, with what happened next. The guide to building an objection library shows how.

7. Customer voice. A verbatim quote bank from calls, reviews and interviews. It supplies headlines, ad copy and FAQ answers in buyers' own words.

Most of this already exists in call recordings and the CRM. It is just not anywhere AI can read it.

Which sets describe how you sound and look?

Four sets cover how the business presents itself: brand voice, employee voice for named people, brand visuals, and the product interface rebuilt in code. They make AI writing sound like you and keep anything designed on brand, without exposing real customer data.

8. Brand voice. How the company sounds when the company is the author: tone, sentence length, phrases used and avoided, punctuation habits.

9. Employee voice. How each named person sounds when something goes out under their name. Only that person can confirm their own file.

10. Brand visuals. Colours, type and logo, read by anything that is designed, such as a lead magnet or a social graphic.

11. Product interface. The product's screens rebuilt in code, so demos and films never show a real customer's account.

Which set describes the market?

One set covers the market: competitors. It records the real competitor set, how each positions itself, and what their own users praise and complain about, from reviews and forums. It lets AI position the business honestly instead of vaguely, and grounds any comparison in evidence rather than assumption.

12. Competitors. For each, what they offer, who they suit and, importantly, what their customers say about them. A comparison built on real user complaints and praise is fairer and more persuasive than one built on a features list.

Knowledge gaps here are common and costly. According to Panopto's 2018 survey, 42% of institutional knowledge is unique to the individual holding it. Competitive insight is often exactly that: in the heads of a few experienced salespeople.

Which three sets should you build first?

Build product facts with approved claims, prospect objections and the ideal customer profile first. Those three improve the largest share of everyday AI work, especially in sales and marketing, and they draw on material most businesses already have. Add customer voice and brand voice next.

A practical order:

1. Product facts and approved claims, because they stop the most damaging errors.
2. Objections with best answers, because they make sales and marketing output useful.
3. The ideal customer profile, because it tells AI who it is writing for.
4. Customer voice and brand voice, because they make output sound right.
5. The rest, as the work that needs them comes up.

Each set needs a source for every statement and a named owner. That is what keeps the brain trustworthy as it grows. According to MIT NANDA's 2025 study, most generative AI failed to return value because systems did not adapt to context. Maintained, owned sets are how they adapt.

Skein builds these twelve sets as the company brain of a Company OS, kept current from calls, tickets and emails. That suits a growing team where several people rely on AI daily. A very small team may only need the first three sets in a shared folder.

Frequently asked questions

What should be in an AI knowledge base?

Twelve kinds of canonical knowledge cover most businesses: company facts, product facts with approved claims, use cases, the ideal customer, pain points, objections with answers, customer quotes, brand voice, employee voice, brand visuals, the product interface and competitors. Each is a short, sourced file with an owner.

What is the difference between an AI knowledge base and a wiki?

A wiki holds documents for people to browse. An AI knowledge base, or company brain, holds distilled, sourced statements that AI reads before every task. Because AI depends on it constantly, errors surface quickly and it stays current, while wikis tend to go stale.

How do I start building an AI knowledge base?

Start with three sets: product facts with approved claims, prospect objections with your best answers, and the ideal customer profile. Draft them from material you already have, add a source for every statement, name an owner for each, and connect one weekly workflow to them.

Should I upload all my documents to AI?

Usually not. Uploading everything gives AI several versions of the same facts and no way to tell which is current. Distil documents into short, canonical statements with sources, and keep raw material separate. Curated knowledge produces more accurate output than a large, unsorted pile.

Who should maintain an AI knowledge base?

Each set needs a named owner, such as the product lead for product facts or the sales lead for objections. Owners settle contradictions and approve claims. New material from calls, tickets and emails can be processed automatically, with anything conflicting sent to the owner to decide.

Related reading