A company brain is a shared model of what a company knows. It connects the systems where work happens, keeps the sources and access rules behind each answer, and gives people and AI agents a record they can build on.
At a private equity firm, that might mean joining a company's history across banker decks, meetings and the CRM. At a software business, it might mean keeping a sales promise connected to the delivery decision that changed it. In both cases, the answer needs to reflect who is asking.
The phrase has become popular. Garry Tan of Y Combinator gave a talk at AI Engineer World's Fair 2026 called Every company should have a Brain. Gartner put AI context platforms among its technology trends for 2027. Forrester called context engineering and ContextOps the missing foundation for scaling agentic AI.
The label now covers a wiki, a folder of agent notes, enterprise search and several other useful tools. A client gave us a better way to judge them. Before a pilot, its lead wrote down seven requirements: aggregated, consistent, time-aware, permission-aware, self-maintaining, auditable and sovereign. Each word can be tested.
The Definition
A working definition
A company brain is a shared, permissioned and time-aware model of what a company knows, built from the systems where work happens, which people and agents read from and write back to.
Shared means people and agents work from the same underlying record. Their answers can differ because their access differs. The evidence and identity rules remain consistent.
A model of what the company knows joins information around customers, people, projects, deals and decisions. Documents provide evidence about those things. The reader can follow a customer from a CRM record to a meeting, a proposal and the decision made afterwards.
Built from the systems where work happens means that connected sources feed the record. Someone still owns the source selection, access and quality checks. They should be able to spend their time governing the knowledge instead of rewriting it into another handbook.
Permissioned and time-aware means each answer depends on who can see the evidence and when that evidence was true. A current price and a superseded price can coexist without becoming two equally plausible answers.
Read and written means finished work can become evidence for the next task. A researcher contributes findings with their sources. A team approves a decision and keeps its reasons. Write-back needs its own access rules and review.
| Property | What it means in practice |
|---|---|
| Aggregated | The sources a use case needs join around the same customers, people and projects. |
| Consistent | Different names for the same entity resolve to one record, with lookalikes kept apart. |
| Time-aware | Facts retain their dates and replacement history. |
| Permission-aware | Each passage retains its source's access rules, enforced when it is retrieved. |
| Self-maintaining | Connected sources refresh on an agreed schedule, with freshness measured. |
| Auditable | A claim can be traced to the passage, file and date behind it. |
| Sovereign | The company chooses where its context runs and can export it with its relationships and evidence. |
The Building Blocks
What gets called a company brain
Six approaches appear under the name. Products often combine them, and some already provide several of the properties above. Evaluate the implementation against your questions.
A wiki or handbook
A handbook gives people an edited account of how the company works. It can be an excellent source for a brain. Its freshness depends on someone maintaining it, and decisions made in meetings or chat can take time to reach the page.
Enterprise search
Search helps people find material across systems. If the results remain separate documents, the reader still has to join the customer history and establish which version applies. Ask whether the product resolves entities, understands changes over time and enforces the original permissions through every query path.
Markdown memory files and hand-built knowledge bases
Notes that an agent reads at the start of a session work well for a defined job. A growing team needs to decide who maintains them, how changes reach other agents and which notes each reader may see. Files can inherit folder permissions, but passages copied between files need their access and provenance preserved too. Loading the whole collection on every request also becomes expensive as it grows.
A fine-tuned model
Fine-tuning can teach a house format or a specialist task. Changing company facts need a separate record that can be updated, cited and permission-checked. In Oded Ovadia and colleagues' comparison of knowledge injection methods, retrieval outperformed unsupervised fine-tuning on the knowledge-intensive tasks they tested. That is evidence for the retrieval approach, rather than a verdict on every form of fine-tuning.
Information learned into shared weights has no built-in source-level access check for each person asking. Nor do those weights provide a reliable citation trail. We cover that distinction in What you own when you own your AI.
An assistant with connectors
Connecting an assistant to Slack, email and Drive gives it useful reach. The next questions are how it joins records across those systems, selects the current evidence and preserves access. The results depend on the connectors and retrieval design. A shared context layer can give several assistants the same maintained record instead of rebuilding it inside each conversation.
A vector database
Similarity search retrieves passages related to a question. A vector database is a useful part of that system. Entity matching, dates, permissions and evidence-linked write-back still need to be represented and enforced in the surrounding application.
A company brain can use all these components. Its value comes from how they work together: connected sources become a record of the business that people and agents can query and extend.
The Sources
What goes in
A first use case often needs a handful of well-chosen sources.
- Documents: SharePoint, OneDrive, Google Drive and Notion, including PDFs, presentations and spreadsheets.
- Communication: email, calendars, Slack, Teams and WhatsApp where people choose to connect it.
- Meetings: transcripts and notes from the conversations where decisions happen.
- Business systems: the CRM, ticket tracker and code repository.
- Licensed data: outside company information, stored and reused where the licence permits.
- Finished work: findings, decisions and drafts written back with their evidence.
Agree freshness per source. A priority chat channel might need updates within minutes; an archive may refresh overnight. The connector and deployment determine what is possible. Measure the age of the evidence used in answers, alongside the health of the last sync.
Grant the access the use case needs. Start with the relevant folders, channels and accounts. Preserve source restrictions as the knowledge moves into the graph. A shared connector credential must not turn every retrieved passage into company-wide information.
Leave out the noise. Automated alerts and duplicated exports can dominate a retrieval result without adding much knowledge. Test source selection against real questions before assuming that another connector will improve the answers.
The company brain and the personal brain
Shared company knowledge and personal context need separate ownership rules. In our model, the company brain holds the agreed shared sources. A personal brain holds a person's private email, calendar, notes, direct messages and relationships. Company search does not read that private record.
A source can feed different scopes: shared Slack channels belong in the agreed company context, while direct messages belong to the person. Personal knowledge becomes shared only through the sharing rules the person chooses.
When someone leaves, the company retains the shared work it is entitled to keep. Private personal context stays separate. Agree retention, access removal and export arrangements during setup. Our permissions model explains that boundary in more detail.
The Architecture
How one is built
The order of operations determines which claims the system can support. Follow one document from its source into an answer.
- Fetch permissions and dates with the text. Keep the source's modification time separate from ingestion time. Loading a 2012 deck today must not make it newer than the 2025 revision.
- Preserve the source passages. Split the text into addressable passages, keeping the file, position, version and access rules. Store model findings alongside that evidence so a citation can still lead back to the original text.
- Resolve entities. Match people, companies, projects and deals across systems using evidence beyond a similar name. Our fictional roll-up dataset includes one customer under four names and a separate lookalike. The published retrieval tests check whether a system joins the four and keeps the fifth apart.
- Build and review the ontology. An ontology describes the entities and relationships the use case needs. Generate a starting model from the sources, review it with the people doing the work and revise it as their questions change.
- Link facts and record change. Connect passages to the entities they describe. Extract commitments, decisions, blockers and scope changes. Keep the evidence and dates behind a replacement instead of silently overwriting the earlier account.
- Enrich the record. Attach licensed outside data to the right entity. Reuse it within its licence and refresh schedule, so repeated questions can avoid unnecessary vendor calls.
- Retrieve in stages. Apply access and source filters before returning evidence. Give the agent short excerpts and identifiers first, then fetch the full passages it needs. Graph traversal can connect a customer to its people, projects and decisions.
- Write back with evidence and approval. Store useful finished work with links to its sources. Separate an agent's proposed finding from an approved change to the shared record, and record who made that change.
Fetching, extraction and matching can often run on code or smaller models. Use more capable models where a task needs judgement. Measure quality and cost on your own archive before choosing the processing mix.
The Work
Answer, push, act and remember
Answer a question with evidence
People ask through the assistant or workspace they already use. The same company context can serve Claude, ChatGPT, a coding agent or a team workflow through an integration such as MCP. Each answer needs the sources and dates that support it. If the record cannot support an answer, say what was searched and what remains unknown.
Push the changes people need to see
A daily brief, a drafted stand-up or a meeting note can bring changes into view before two teams repeat the same work. One pilot uses the quality of its drafted stand-up as a practical test of context: how much must a person correct before sending it?
Give agents the context for an action
A proposal can start from a comparable project. An onboarding brief can draw on past configuration decisions. An introduction request can start with the colleague who has a useful relationship to the target. People approve what is sent or changed.
Remember useful finished work
An analyst's research can become the next person's starting point. An approved engineering decision can retain the customer conversation that prompted it. Each contribution needs ownership, sources and a way to correct or replace it later.
The same foundation supports consultancy proposals and precedents, investment research and the handover between sales and delivery. Sold, scoped, shipped follows that last workflow through a project.
Why Now
More AI work makes shared context more valuable
Coordination was expensive before AI. Atlassian's account of its 2025 State of Teams findings says roughly half of knowledge workers report teams unknowingly working on the same things. More parallel AI threads create more places for that work to diverge.
Faros AI's 2025 study of over 10,000 developers found that teams with high AI adoption completed 21% more tasks and merged 98% more pull requests, while review time rose 91%. It found no significant correlation with improvement at company level. These are observational findings. They show why individual output and company outcomes need to be measured separately.
Agents also pass work between tools and people. A shared record makes it possible to inspect what one agent told another, which evidence it used and whose access applied. Keeping that record outside model weights also makes it easier to change models while retaining company context.
The analyst research reflects the same concern. Forrester's October 2026 ContextOps article describes the operational work needed to govern, monitor and improve context. Gartner's 2027 trends announcement identifies AI context platforms and predicts accuracy and cost gains from better semantics. Those gains are a forecast, not a measured result from a 60x installation.
Repeated context assembly also has a cost. At $250 a research run, twice a working day for 21 days, the monthly total is $10,500. That is illustrative arithmetic. An audit has to establish how much of a real bill comes from work that could be reused, including the cost of maintaining the shared record. The $10,000 analyst sets out that test.
The Tests
Ten tests for a company brain
Run these on the system you are evaluating, including ours. Use questions whose answers you already know.
- Two people, one question. Sign in as people with different access. Where the evidence differs, the answers should reflect that difference.
- A source the asker cannot see. Ask about restricted material. Test direct search, graph traversal, saved answers and an agent using a shared credential. The restriction must hold through each route.
- The current version. Ask about a price, scope or decision that changed. Check the answer, source date and replacement history.
- An answer as of a date. Ask what was true before that change. Check that the system retrieves the earlier evidence and labels the time being answered.
- One customer, several names. Join the customer's records across systems. Include a lookalike company to test whether identity matching goes too far.
- A question the record cannot answer. Ask about work with no supporting evidence. The system should explain the gap and distinguish a nearby example from an answer.
- The passage behind each claim. Follow the citations. Check that the original passage supports the claim and that you may still open it.
- A source change. Edit a document or post an update. Measure when it appears in answers, and compare that with the agreed freshness for that source.
- A different model. Change the model behind an agent. Check that the context, access rules and history remain available.
- A complete export. Ask for the evidence, entity relationships, dates and permissions in a usable format. Test what you can query and run independently of the supplier.
Ask where each rule is enforced. Access checks must cover the evidence store and every route that retrieves from it. Demonstrate the failure cases with ordinary user accounts as well as an administrator.
The Limits
What it cannot do for you
It cannot recover knowledge that was never recorded. Transcripts and chat can capture useful context, but some knowledge remains in people's heads. Interviews and human judgement still matter.
Recording changes behaviour. People may move a sensitive conversation when they think it will be shared too widely. Explain the ownership and sharing rules before connecting personal sources, and make those rules visible in the product.
Source permissions can already be wrong. A folder shared too widely needs fixing at the source. The brain must then pick up that correction and apply it to later retrieval.
It cannot settle a disagreement for the team. Two partners can read the same evidence and reach different conclusions. The brain helps them establish the facts and keep the decision they make. Leadership and responsibility stay with people.
The first load needs time and a budget. Processing a historical archive can cost more than the daily refreshes that follow. New sources, changed extraction rules and reprocessing can add further work. Include them when evaluating the running cost.
A First Use Case
Start with one expensive question
Choose a question your team spends too long answering. It determines which sources, relationships and quality checks the first version needs.
- Collect 20 to 30 real questions from the past month, with the answers a good colleague would give. Keep them as an evaluation set.
- Map the sources behind those answers, their owners and their access rules. Fix obvious access problems and leave out irrelevant noise.
- Agree a refresh interval for each source and a way to measure it.
- Review the entities and relationships with the people doing the work.
- Connect one team's existing assistant. Keep changes to shared records under review while the team learns to trust the answers.
- Ship a visible output, such as a daily brief or drafted stand-up, and measure its accuracy and the editing it needs.
We install an agreed first use case in two weeks. Choose 60x Cloud for hosting with EU or US data residency, or Your Cloud inside your Azure, AWS or GCP account. Both can run on open-weight models.
Book a demo and bring the question your team spends most time answering. We will work through the sources, permissions and evidence a useful answer needs.
Sources
Sources
- Garry Tan, Every company should have a Brain, AI Engineer World's Fair 2026.
- Gartner, Top Strategic Technology Trends for 2027, 20 October 2026.
- Charlie Dai, Leslie Joseph and Indranil Bandyopadhyay, Context Engineering and ContextOps, Forrester, 19 October 2026.
- Oded Ovadia, Menachem Brief, Moshik Mishaeli and Oren Elisha, Fine-Tuning or Retrieval? Comparing Knowledge Injection in LLMs, revised January 2024.
- Atlassian, How to make work visible and improve alignment, reporting its State of Teams 2025 findings.
- Faros Research, The AI Productivity Paradox Report 2025.
Client examples are anonymised. The roll-up dataset and visual examples are fictional. The research cost calculation is illustrative. Deployment scope, integrations and refresh intervals are agreed for each installation.