What is a company brain?

Most businesses already have the knowledge they need. The problem is that it's scattered. In a 2025 survey of 4,000 employees at large U.S. and U.K. companies, workers reported spending an average of three hours a day searching for information, and said 42% of what they found didn't apply to their job (Coveo). A company brain fixes that by putting it all in one place.

The filing cabinet metaphor

Imagine you need to know how much you paid for electricity a few years ago. Some of your bills are loose in a couple of desk drawers. Some are in a box in the attic. A few are in a locked safe, and the rest are stuffed in a kitchen drawer with the takeout menus. Finding two specific bills means checking every spot, and you still might miss one.

That's what most businesses look like on the inside. Knowledge is spread across shared drives, email, text messages, Slack and Teams threads, databases, old zip files and photos, each in a different system with a different search box. When someone needs an answer, they go digging, or they ask the one person who remembers.

A company brain is the filing cabinet. Every document that comes in gets filed in the right folder, with labeled tabs, in alphabetical order. Nothing is left in the kitchen drawer. And one drawer still has a lock, so the right people can open it and everyone else can't.

The basic idea

A company brain does four things:

• Puts everything in one place. It connects to the tools you already use and pulls their content in, so there's one place to look.
• Keeps it organized. Content is cleaned up and labeled by source, date, author and topic.
• Keeps track of versions. When a document changes, the brain updates, so you get the current policy, not the one from 2022.
• Is built for AI. It's stored in a way AI tools can search quickly and accurately, which an ordinary shared drive isn't.

This idea has gone mainstream. In October 2025, OpenAI launched “company knowledge” in ChatGPT, which pulls from tools like Slack, SharePoint, Google Drive and HubSpot, limits answers to what each user is already allowed to see, and cites its sources (VentureBeat). A company brain built for your business does the same, across all of your systems, with whichever AI you choose.

How data flows into a company brain

Data moves in one direction: from your sources, into the brain, and then out to you or your AI agents. Here are the steps.

1. Connect to your sources

The brain plugs into each system in one of four ways:

• APIs, the official doors software provides for other programs to read its data.
• Webhooks, where a tool sends an instant notice when something changes.
• Incremental pulls, which fetch only what's new or changed since the last check.
• Custom connectors, for older or in-house systems that don't offer any of the above.

2. Bring the data in and keep it current

• Event-driven updates handle changes as they happen, like a new Slack message or an edited document.
• Scheduled pulls check slower systems every hour or so.
• A nightly reconciliation compares everything against the sources at the end of the day to catch anything missed and remove anything deleted.

3. Store it so it can be found

Incoming content is parsed (text pulled out of PDFs, slides and images) and cleaned up. Then an embedding model turns each passage into a list of numbers that represents its meaning. Passages that mean similar things get similar numbers, even if they use different words, so “time off policy” and “vacation rules” end up close together.

Those numbers go into a vector store, a database built to find the passages closest in meaning to a question. The best setups use a store that handles both this meaning-based search and traditional keyword search natively. Keyword search is better at exact matches like invoice numbers, product codes and names; vector search is better at meaning. Using both, called hybrid search, is becoming the standard, and mainstream search engines like Elasticsearch now support it out of the box (Elastic).

Behind the scenes, the data usually lives in a mix of places: object storage for the original files, a relational database for records and metadata, a search index for fast lookups, and a cache for frequent questions.

At this point the data is stored and ready. What makes a company brain different from any other company database is the specialized retrieval on top of it, which we cover below.

4. Set permissions and governance

Before anyone searches anything, the brain has to know who's allowed to see what. That's the locked drawer in the filing cabinet. Permissions can be set at several levels:

• Per object: a single file or message.
• Per row: individual records in a database, like only the deals a salesperson owns.
• Per tag or section: everything labeled “HR” or “Finance,” for example.

The best approach is to copy the permissions that already exist on the source. If someone can't open a folder in Google Drive, they can't see its contents in the brain either. When permissions change at the source, they change in the brain too. That way nobody has to maintain a second set of rules, and the brain never becomes a back door to sensitive information.

5. Get answers out

People usually use a chat portal: type a question in plain English and get an answer with links to the sources.

AI agents usually connect through MCP, the Model Context Protocol, an open standard for connecting AI to business systems (Model Context Protocol). That lets any agent that supports MCP draw on the brain, with the same permissions. We explain it in What is an MCP server?

How a question gets answered

Here's what happens in the few seconds after someone asks a company brain a question:

1. Understand the question. An AI model rewrites vague questions, expands acronyms and pulls out filters. “Last quarter” becomes a date range; “in Jira” becomes a source filter. It also decides whether the question is simple or needs several rounds of searching.
2. Check permissions. The system looks up who's asking and which groups they belong to, from Okta, Microsoft Entra ID or Google Workspace, so the search only covers what they're allowed to see before anything gets ranked.
3. Run hybrid search. Keyword search (usually an algorithm called BM25) and vector search run side by side, each returning its top 50 to 200 candidates.
4. Combine the results. The two ranked lists are merged, usually with a method called Reciprocal Rank Fusion, which rewards documents that rank well in both.
5. Rerank. A specialized reranking model (Cohere Rerank, Voyage, bge-reranker or Jina are common) reads each candidate alongside the question and reorders them by true relevance. This is often the single biggest quality gain.
6. Assemble the context. Duplicates are removed, short snippets are expanded to their surrounding section, recent and authoritative sources are favored, and everything is trimmed to fit what the AI model can read at once.
7. Answer with citations. The AI writes its answer using only what was retrieved and links to each source, so you can check it.
8. Log and evaluate. Each step is recorded in a tracing tool such as LangSmith, Langfuse or Arize Phoenix, and answer quality is measured over time with tools like Ragas, TruLens or a custom set of test questions.

Reciprocal Rank Fusion comes from 2009 research showing that a simple combination of rankings beat more complex methods (Cormack et al.). Reranking models take a shortlist of results and sort them from most to least relevant to the question (Cohere). Together, these steps are what turn a pile of search results into an answer you can trust.