NEW

SurrealDB 4 is here, introducing SurrealDB Earth and SurrealDS!

Learn more

How to build a company brain: where to start and what it costs

AI

Nov 21, 202610 min read

Martin Schaer

Martin Schaer

Show all postsHow to build a company brain: where to start and what it costs

To build a company brain, start with one workflow where missing context hurts, such as product and engineering alignment meetings. Capture those conversations automatically, give them permissions that match who was in the room, and serve them to the agents doing the follow-up work. Prove that one use case, then add sources.

This is the practical follow-up to What is a company brain?. That post covers the idea. This one covers the decisions: where to start, who owns it, what to buy or build, and what it costs.

Our newsletter

Get tutorials, AI agent recipes, webinars, and early product updates in your inbox every two weeks

No. A context layer gives agents access to data that is already recorded and structured: the CRM, the data warehouse, policy documents. A company brain adds what was never written down: what was decided on a call, why it was decided in a chat thread, what a client said in passing.

Context layerEnterprise searchCompany brain
What it holdsStructured data you already storeDocuments and messages across your appsDecisions, reasons and commitments, including from calls and private chats
How it learnsWhen someone updates a systemWhen a document changesContinuously, from everyday work
Who uses itAgentsPeoplePeople and agents
Hardest problemData qualityRelevancePermissions on private conversations

Most organisations need both. The context layer tells an agent what the records say. The company brain tells it why.

Start with a quick win that has a clear benefit for the business. The trap is designing an organisation-wide system on day one. You learn more from one working use case than from months of planning, and that learning shapes the next step.

My recommendation is to start as if you were deploying a fleet of notetakers. Picture a product manager and two engineers discussing a feature on a call. Today, the coding agent that later picks up that ticket knows none of it. With a company brain, it automatically gets the decisions, the constraints and the reasons from that meeting. For the people involved, that feels like a superpower.

A sequence that works:

  1. 1.

    Pick one recurring meeting where decisions get made and work follows, such as product and engineering alignment.

  2. 2.

    Capture it automatically, with the people in the room knowing what is captured and who can see it.

  3. 3.

    Scope each memory to its participants and their teams, so access matches who was there.

  4. 4.

    Serve it to the agents doing the follow-up work, starting with coding agents.

  5. 5.

    Add the sources that record the outcome, such as tickets and product requirement documents, so the brain knows what was decided and what was delivered.

  6. 6.

    Move to the next workflow where context nobody could reach before would change how people work.

Engineering is a good first partner. Engineers tend to use AI the most, their agents benefit immediately, and they often arrive with ideas of their own.

A business unit or a dedicated role with a broad view of the business should own it. It is a company-wide asset, and its owner needs to understand where the business value is. Pair that owner closely with engineering, which will build most of the integrations and is usually the first team to benefit.

A company brain has four layers. You can buy, build or assemble each one separately.

  • Capture: connectors that pull in meetings, chats, tickets and documents. This is where most of the custom work goes.

  • Understanding: turning raw transcripts and files into facts, people, decisions and relationships, and reconciling them when they contradict each other.

  • Memory: the database that stores all of it, finds it by meaning, by keyword and by relationship, and enforces who can see what.

  • Serving: the interface agents and people use to ask questions, usually through MCP or a native integration.

The memory layer is the one to choose most carefully. It holds the private data and enforces the permissions, and replacing it later means migrating everything.

MCP (Model Context Protocol) is the plug, not the brain. It is an open standard that lets any AI agent connect to a tool or data source in the same way. That makes it a fast way to connect your company brain to Claude, ChatGPT, coding agents and internal assistants without building each integration by hand.

Storing knowledge and deciding who may see it are jobs for the memory layer underneath. Two more points matter when you plan:

  • MCP is not the only protocol. Agents can also reach data through APIs, command-line tools and native integrations.

  • Native plugs can be cheaper. A purpose-built integration is often faster and uses fewer tokens than a generic MCP server, which adds up when thousands of agent calls run every day.

Use MCP to move fast, and switch to a native integration where volume makes the token bill matter.

Assemble. Off-the-shelf work AI such as Glean or Microsoft Copilot covers search across popular apps well. It reaches its limits where your knowledge lives in internal tools, private conversations and workflows nobody else has.

Most of the effort goes into integrations that have to be custom-made. This is why forward-deployed engineers, who build inside the customer's organisation, have become so common. AI has also made building much cheaper than it was. The sweet spot is to keep the tools your teams already use and build the bridges between them, on top of a memory layer you control.

Most buyers are still working this out, and they are right to be cautious. These questions separate a demo from something you can run:

  • Permissions: are they enforced inside every search, or applied afterwards by the application?

  • Deployment: does it run where your data rules require?

  • Cost: what does a sample of your own content cost per page and per query?

  • Freshness: can it tell a reversed decision from a current one?

  • Integrations: does it connect to the tools your teams actually use?

  • Exit: how hard is it to move your memory somewhere else later?

It depends on how regulated you are.

ModelBest forTrade-off
SaaS or managed cloudCompanies without strict data rulesEasiest and cheapest to run
Bring your own cloud (BYOC)Companies that need data in their own cloud accountMore control, more operational work
On-premisesBanks, healthcare, governmentFull control, highest cost

A company brain holds your most sensitive conversations, so involve legal and security early. Settle the deployment model before you choose vendors, because it rules some of them out.

The cost depends on how many sources you connect, how complex your permissions are, and how much content you process. Published estimates for custom enterprise search, the closest comparable, range from $40,000 to $200,000 and 12 to 24 months to build (RaftLabs, GoSearch, both vendors, 2026).

The main cost drivers:

  • Connectors. One SharePoint connector with nested folder permissions can take three to five weeks and 30 to 40% of the project timeline, according to GoSearch.

  • Document processing. Turning PDFs, slides and transcripts into usable text, at a price per page.

  • Model and token usage. Every question an agent asks costs tokens, and agents ask a lot.

  • People's time. Someone has to own sources, settle conflicts and review what the brain gets wrong.

The hidden costs are the choices that look small at the start. I saw one company building a context layer spend its whole budget on document ingestion. It picked the most accurate document parser available, which cost at least 30 times more per page than a standard one. Nobody had compared cost against accuracy, or looked at hybrid methods that don't run everything on GPUs.

Other costs that tend to surprise teams:

  • Re-processing. Changing your AI model or parser can mean processing everything again.

  • Permission syncing. Keeping access in step with your identity system as people join, move and leave.

  • Legal and privacy review, especially for recorded calls.

  • Change management. Explaining to employees what is captured, and earning their trust.

Run a small sample through two or three options before committing to any per-page or per-token price at scale.

Most fail on permissions and governance. Recorded conversations mix information with different audiences. When a team can't enforce access per conversation, the safe choice is to include only what everyone may see. The result is a brain built to the lowest common level of access. It misses exactly the private context that made it worth building.

The wider numbers point the same way. MIT's NANDA initiative found that 95% of generative AI pilots delivered no measurable return, and blamed tools that don't learn from real workflows (MIT NANDA, July 2025). Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, citing costs, unclear value and weak risk controls (Gartner, June 2025).

The other common causes:

  • Starting too big, so nothing reaches users before the budget runs out.

  • No clear business benefit, so nobody defends the project when priorities change.

  • Stale answers, when the brain can't tell a reversed decision from a current one.

SurrealDB is the memory layer. It keeps documents, relationships, vector search and full-text search in one database, so an agent can find a memory by meaning, follow it to the people and decisions it relates to, and filter by date in one request. Permissions are checked inside every search, on every record, so a forgotten filter in an application can't leak a private conversation.

SurrealDB Agent Memory builds on that to ingest conversations and serve them to agents. Each memory is scoped to the people it belongs to, and an agent can never see more than the person it works for, and agents can reach it through MCP. SurrealDB itself runs as a managed service, in your own cloud or on-premises.

That is what makes the notetaker approach safe. A product meeting stays with the people who were in it and their team, and their coding agents can still use it.

A single use case is much faster than an organisation-wide brain. Comparable enterprise search projects built from scratch quote 12 to 24 months.

RAG is a technique for feeding retrieved text to a model. A company brain uses retrieval, but also captures knowledge continuously, reconciles contradictions and enforces who can see what.

Meeting transcripts from one recurring decision-making meeting, then the tickets and documents that record what came out of it.

If you're planning a company brain, our solutions engineers can help you pick the first use case and the right deployment model:

Our newsletter

Get tutorials, AI agent recipes, webinars, and early product updates in your inbox every two weeks