Docs: What is a company brain?
What is a company brain?
On this page
A company brain is the knowledge a business runs on, in people's heads and across its systems, made available to the AI agents the company already uses.
The term is new. Y Combinator listed "Company Brain" among its Requests for Startups for the Summer 2026 batch, the ideas it asks founders to build. Tom Blomfield's entry describes it as "a living map of how a company works". This page says what that means in practice, why agents need one, and how Maplith builds one.
What a company brain is
Every business runs on knowledge that is written down nowhere in full. Which system holds the real customer list. That an invoice reaches its order through a reference someone types by hand, and that the field is often left blank. That a refund above a certain amount needs a manager. That the current price list is a spreadsheet, not the system of record.
Some of it lives in people's heads. The rest is spread across the systems: the web shop, the ledger, the helpdesk, the CRM, a shared drive. People get by because they remember roughly where things are and how they fit together.
A company brain holds that knowledge in a form an agent can use: what the business is made of, where each part lives, how the systems connect and how far each connection can be trusted, what has changed, and how the business handles a given case. An agent consults it before it starts work, not after it has guessed.
Why agents fail without one
The AI assistants companies use today can reason about almost anything. What they do not know is the company they work for.
- Every session starts from zero. An assistant does not carry yesterday's discoveries into today's chat. It reads the same exports again, or someone explains the systems, rules and exceptions again. On an earlier synthetic test company, answering one question from the raw exports took about 26,900 tokens, every session. With the map, it took about 6,200 tokens across three calls.
- Joins are guessed. A question that spans two systems needs to know how they connect, and column names are a poor guide. Asked to guess, a model found 4 of the 10 joins that measurement finds in Maplith's test company's exports.
- The rules live in people's heads. How a return is handled, which amounts need a person's approval, which exception applies to which customer. An agent that cannot see them either asks every time or does the wrong thing.
A better model does not solve this. What is missing is not reasoning but knowledge of one particular business.
The token figures were measured on an earlier synthetic test company, not today's demo company; its exports are 107 KB. A real business exports megabytes; the map's answers are sized by the map, not by the number of rows.
What a company brain is not
- Not enterprise search. Search finds the documents that mention a word. It does not tell an agent that two systems name the same customer differently, or that a link between them holds for most rows and not the rest.
- Not a chatbot over documents. Answering questions from a wiki helps a person read faster. It does not give an agent the structure of the business, what changed since yesterday, or a way to ask before it acts.
- Not a data warehouse. A warehouse copies the rows into one place, and someone has to model and maintain them there. A company brain describes where the data is and how it connects. The data stays where it is.
How Maplith builds one
Maplith builds the brain from the systems a company already uses. The company connects them, or uploads the export files their own export buttons produce. Nothing is migrated, and its tools stay as they are.
- The map. The business objects, such as customers, orders, invoices and returns, with their fields, where each one lives, and the relations between systems. Joins are measured, not guessed: the engine tests pairs of columns across the company's data and publishes each join with its direction, its cardinality and how well each side matches. A join that holds only partly carries its caveat.
- Findings. Where the systems disagree, with the evidence: invoices that never matched their quotes, work handed out in a channel the system of record never sees.
- Playbooks. The steps, checks and exceptions the company's people wrote down. Maplith can draft one from an uploaded procedure; it stays a draft until an admin puts it in use.
- Signals. What changed since the last scan, and the rules the company wrote, as typed events an agent polls instead of re-reading everything.
- Task context. One call that gives an agent what one task touches: the objects, the verified route between them, the open findings, the playbooks, and what was done last time.
- The action gate. Before an agent acts in the company's systems, it asks. The answer is allowed, pending or denied, by the policy the company sets. A person approves what the company says needs approval, and every request, decision and outcome is logged.
The company's own agent does the work in its systems, with its own access. It can be ChatGPT, Claude or any assistant that speaks MCP. Maplith never performs the action and never holds the credentials the agent acts with.
The data stays in the company's systems
Maplith keeps metadata about a company's data, not copies of its records. Uploaded export files are read for a build and then deleted. What remains is the map: objects, fields, relations, processes, metrics and findings. It tells an agent where the data is, by which key, and how much to trust each hop. When the agent needs a row, it reads it in the company's system.
The map is not free of personal data. Field descriptions can quote example values, and findings quote evidence from the files. Before content is sent to the model, Maplith tries to mask email local parts, Danish CPR numbers, IBANs, card numbers, the names of people in the export's own name columns, and the masking terms the company adds. It is a best-effort safeguard, not a guarantee, and the company decides what its exports contain. Hosting, database and file storage are in the EU; a language-model provider in the United States does the model inference. The Security page has the detail.
See one built from your systems
The quickest way to understand a company brain is to look at one built from your own systems.