How to build a company brain: a practical guide for mid-sized organisations
← INSIGHTS

How to build a company brain: a practical guide for mid-sized organisations

Generic AI knows nothing about your business. Here is what a company brain is, six steps to build one, and what goes wrong after the demo.

BY STEFAN MARITZ··9 MIN READ

TL;DR

  • A company brain is a shared, organised record of how your business works, including your products, processes, and the decisions you have made and why, that people and AI tools can both query.
  • Interest has jumped. Google searches for "company brain" rose from about 140 a month in September 2025 to 3,600 in August 2026.
  • Generic AI starts from zero each time. A company brain gives it your context, and it needs to be lean, current, permissioned, and cited.
  • Six steps build one: pick a use case, capture selectively, structure it in layers, enforce permissions and provenance, connect it to your tools, and maintain it.
  • The hardest part comes later. Content drift, deleted sources, and permission rot are what make in-house builds lose trust.
  • Test any version, built or bought, with 20 real questions your people ask, and check each answer for accuracy, citation, and permission.

Many teams using AI today have the same problem, whichever tool they pick. The assistant is capable and it knows nothing about the business. Someone pastes in last year's pricing sheet, or a policy, or a client history, and the assistant forgets it by the next conversation.

A company brain is the fix, and the term is moving fast. This guide explains what it is, how to build one in six steps, where it goes wrong, and how to decide between building and buying.

What a company brain is

We use a plain definition. A company brain is a shared, organised record of how your business works: your products, your processes, and the decisions you have made and why. People add to it over time, and both staff and AI tools can ask it questions.

Vendor definitions are similar. Gyld, which sells a company brain product, describes it as a living knowledge layer that captures what an organisation knows, keeps it current, and makes it queryable by people and AI agents. It cites a guide from Falconer, another vendor, listing four required properties: capture, update, organisation, and monitoring. Both companies sell in this space, so read their framing with that in mind. The properties still make sense as a checklist.

The definition excludes some things that get sold under the label. A wiki stores documents but does not know which are stale. A search tool finds documents without synthesising them. A vector database retrieves similar text and has no sense of permissions or accuracy. A company brain sits above those and adds currency, access control, and sources.

Why generic AI lacks one

In many organisations, one team pays for one assistant and another uses a different one, and nobody in leadership can say which company information has gone into which tool. Our guide to shadow AI covers that pattern.

The obvious answer is to give the model more material, and that has limits. A research paper, Lost in the Middle, found that language models perform best when relevant information sits at the beginning or end of the input, and that performance degrades significantly when it sits in the middle of a long context, even for models built for long inputs. More context is not automatically better context. A company brain works by retrieving the few relevant pieces for each question.

Theory Ventures, which runs an internal company brain, makes a related point about what it calls context bloat. Tool definitions, skills, and long results all consume space, so the team keeps descriptions focused and returns small results first, with the full source available when needed.

Six steps to build one

  1. Start with the questions. Pick one use case and write down the 20 questions people ask most often in it, such as "how do we handle a refund" or "what did we promise this client". A brain that tries to hold everything on day one holds nothing well.
  2. Capture selectively. Audit where knowledge lives: chat threads, email, shared drives, the CRM, policy documents. Gyld's advice is to connect the channels that carry real decisions and skip the rest, because indexing everything adds noise.
  3. Structure it in layers. Keep the stable, always-true context separate from timely raw material that is pulled in on demand. Contengi's guide to setting up a brand knowledge base for AI describes this two-layer model for marketing content: a lean knowledge base loaded every time, and a library of transcripts, past content, and notes drawn on selectively. The same split works for a whole company, with policies and definitions in the first layer and project history in the second.
  4. Enforce permissions and show sources. Not everything is for everyone. Gyld argues that permissions should be checked against the source application at question time, and that every answer should cite where it came from. Theory ties tool calls to each person's verified work identity, and keeps reads and writes in separate tools so write access is granted only when a task needs it.
  5. Connect it to your tools. Agents need a standard way to ask your brain for context. The Model Context Protocol, an open standard introduced by Anthropic in late 2024, is the one both Gyld and Theory use. Gyld notes the protocol is still changing, so choose a setup that lets you adapt.
  6. Maintain it. Give the brain an owner and a review rhythm. This step decides whether people still trust it in a year.

What breaks in year two

Gyld's most useful section covers three failures that appear after the demo.

  • Content drift. A document is edited, and the brain keeps serving the old version, or both versions survive and the model picks one.
  • Deleted sources. Someone removes a file or archives a channel, and the content stays queryable.
  • Permission rot. Access changes at the source, and the brain keeps answering questions the person can no longer ask.

Whatever you build or buy, ask what happens to each of these six months in. Every vendor can demonstrate retrieval on a clean dataset.

Governance and personal information

A company brain will contain personal information unless you decide otherwise. A POPIA executive at South Africa's Information Regulator told ITWeb in September 2026 that AI models often retain personal information beyond its intended purpose, which sits badly with the Act's storage limits. Decide up front what never goes in, using the data rules from our guide to an AI policy, and see our AI governance framework for ownership and review. Ask your legal counsel to check the design.

Build or buy

There are three routes. You can assemble one from a document store and a retrieval layer, which gives you control and makes you responsible for maintenance. You can buy a managed product. Or you can have a partner set one up and train your teams.

Several vendors, including Gyld and Falconer, argue that in-house builds tend to fail at maintenance, because detecting drift without a person in the loop is the hard part. They sell the alternative, so treat the claim as a hypothesis, and test it by asking your own team who will own year two. We are building CompanyLLM for organisations that want a company brain connected to their documents and systems, with access controls leadership sets. It is our own product, and it is on a waitlist while we onboard organisations a few at a time.

A 30-day plan

  • Week one: choose the use case, write the 20 questions, and name an owner.
  • Week two: audit sources, connect the few that matter, and set permissions.
  • Week three: pilot with one team. Ask the 20 questions and score each answer for accuracy, citation, and permission.
  • Week four: fix what failed, set the review rhythm, and agree how you will measure use. Our guide to AI adoption covers that.

If you want a scored view of whether your data, people, and processes are ready for this, the AI Readiness Diagnostic is free to take.

Frequently asked questions

What is a company brain?

A company brain is a shared, organised record of how your business works that people and AI tools can query. It holds products, processes, and decisions, and it stays current, respects permissions, and shows its sources.

What is the difference between a company brain and a knowledge base?

A knowledge base is usually a store of documents that people maintain by hand. A company brain adds automatic updating, access control checked at question time, and sourced answers for AI tools.

Do we need to fine-tune a model to build one?

Usually not. Fine-tuning builds knowledge into a model at one point in time, which makes updates expensive. Retrieving relevant material at question time keeps the model general and the knowledge current.

How long does it take to build a company brain?

Vendors claim hours to connect sources. Scoping a first use case, choosing what to include, and checking the answers takes longer, and our plan allows 30 days for one use case. The slow part is deciding what is true.

Who should own a company brain?

One named person with authority across functions, supported by the teams whose knowledge it holds. Without an owner, content drifts and trust falls.

WORK WITH PRAXES

Ready to move from reading about AI to getting your organisation ready for it? Start with the diagnostic.

Take the AI Readiness Diagnostic