Company Brain AI: A Ready Platform or Your Own Knowledge Base? (2026)

When employees search email, drives and an ERP separately, a company brain gives them one cited answer. A packaged enterprise-search platform fits standard sources and tidy permissions; a custom knowledge base fits proprietary systems or strict data-residency requirements. Syntalith custom builds start at €6,000 net.

Quick answer

There are two routes and both can be right.

• A ready platform. You buy a per-seat subscription, connect it to the systems you already run, and go live in weeks. It wins when the company sits inside one ecosystem, question volume is high, and the knowledge lives in standard tools.
• Your own knowledge base. You build the search layer on your sources, in your environment, with permissions enforced on your side. At Syntalith this starts from €6,000 net, typical full implementations fall in the €6,000–35,000 net range, and maintenance is priced individually. It wins with non-standard sources, strict permissions, a data residency requirement, and at the scale where per-seat pricing turns expensive.

The first step is free either way: a process scan (€0) is a 30-minute engineer call plus a written takeaway in two business days. If you need a portable document with architecture and a fixed quote before deciding, the implementation specification is €1,200 net and if you commission the system build from us, we credit the specification fee toward the build. Full rates are on the Syntalith pricing page.

If the category itself is still new to you, start with the piece on what a RAG knowledge assistant is. This article assumes the category is clear and deals only with choosing a route.

What is a "company brain" exactly?

The category is called enterprise search, and two years of marketing renamed it the company brain. It means one layer over every place a company keeps knowledge: drives, mail, chat, ticketing, product documentation, sometimes a database. An employee asks in plain language, the layer searches many sources, the model composes an answer and shows what it was built from.

Three things separate this product from an ordinary search box. First, it returns a sentence instead of a list of ten links. Second, it cites, so the answer can be checked. Third, and this is the hard part, it respects permissions: an employee only gets answers from documents they have a right to. That third property drives the price of an implementation more often than model quality does. We break it down separately in the piece on permission-aware RAG enforced before the model.

The category is real, and the money shows it. Glean announced in 2026 that it had passed $300M in annual recurring revenue (Glean, 2026). Large companies pay for this, and they pay per seat.

What does the platform market look like in 2026?

The map below is descriptive. Each of these tools is good in its own scenario and none is good in all of them. The last column points to the vendor material each description comes from. As of August 2026.

Two notes on that table. Onyx is not free: the licence costs nothing, while hosting, upgrades, connectors and maintenance cost your team's time, and that line belongs in the calculation. We list Hyper because technical teams ask about it, though a product at this maturity is not a candidate for your HR department's knowledge.

Where do "memory layers" for agents sit on this map?

They occupy a separate corner of it, the one Hyper also sits in. Instead of answering a person's question, that layer feeds agents what the company has already established. Cognee announced version 1.0 in June 2026, running graph and vector memory on a single Postgres (Cognee, 2026). Accuracy comparisons in this category are published by the vendors themselves, Zep against Mem0 for instance, so read them as sales material that contains numbers and run your own eval on your own traffic.

For a buyer shopping for a knowledge base for people, the conclusion is simple: this is a different product and it does not replace enterprise search. If your case is memory for coding agents, look into that corner. If it is employees asking about procedures, go back to the rest of the table.

When is a ready platform the better choice?

We will say this plainly even though we sell builds: in many companies a ready platform is cheaper and faster, and we will advise against a custom implementation.

• One ecosystem. All the knowledge in Google Workspace, all of Microsoft 365, or all of Atlassian. The connectors are already written and tested, and permissions inherit from the source system.
• Questions with document answers. People ask about procedures, decisions and documents. Nobody expects this layer to compute something in a production database or perform an action in a system.
• A high seat count with simple needs. When many people need exactly the same thing, rolling out a ready platform is often faster than running a project. There is no headcount threshold that decides this for you.
• No hard residency requirement. The data may sit with a vendor, and the vendor's DPA and security model satisfy your risk function.
• No team to maintain it. Your own knowledge base is a production system that somebody has to monitor and keep current.

If all five points hold for you, buy the subscription and come back to the build conversation when the platform starts hitting a wall.

When does a custom knowledge base win?

Five conditions. No count of satisfied conditions decides the choice: one hard security requirement can settle it in favour of building, while three soft ones may settle nothing.

• Non-standard sources. The knowledge sits in your own ERP, a production database, an archive of scans, or an industry system no vendor has a connector for. Writing that connector can be more work than the rest of the implementation combined.
• Data residency and self-hosting. The data must not leave your environment. An EU region on its own does not remove the risk of access under a third country's law. Assess the provider's jurisdiction, its actual control over the data, the transfer basis, and whether you can keep both data and models entirely in-house. The legal assessment belongs to your counsel; our job is an architecture that gives them a real choice.
• Strict permissions. When different roles must see different documents and a leak is unacceptable, scope has to be enforced before the model, in the query and in the database. Ready platforms inherit permissions from the source systems, which works well as long as those systems are tidy.
• Per-seat cost at scale. A subscription grows linearly with headcount. A one-off build does not. You calculate the crossover yourself, with the formula below.
• Auditability. When you have to show an auditor where a specific answer came from, which documents were in context and who had rights to them, your own layer gives you logs in your format. That matters for AI Act and GDPR obligations.

What does it cost on your numbers?

The formulas below are a substitution exercise and contain no promise of savings. Calculate both sides over the same horizon, chosen explicitly. The example below uses 24 months; if your budget cycle runs differently, substitute your own:

Platform cost (24 months) =
  seats x monthly rate per seat x 24

Custom knowledge base cost (24 months) =
  one-off implementation (from €6,000 net)
  + 24 x monthly maintenance and hosting
  + 24 x model cost (questions x average token use x provider rates)

A modelled scenario (arithmetic on your assumptions, no measurement at a client). At 120 seats and a monthly rate of X, the platform costs 120 × X × 24 over that horizon. At 25 seats the same formula gives 25 × X × 24, under a fifth of the first figure, and at that size a one-off build usually loses the comparison. Take X from the rate in your vendor quote, because only that version of the calculation means anything. On the build side add implementation, hosting, maintenance and model costs.

Two things are easy to forget on the build side: tidying the documents before launch, and your team's time to maintain the system. Both are real cost lines and we put both into the quote explicitly. One thing is easy to forget on the platform side: seat counts usually grow, and rates rarely fall.

What to demand from a vendor before you sign

Whichever route you take, these questions save the most money:

1. What exactly happens to a document the asker has no right to, before the model starts generating?
2. Where does the index physically live and who has administrative access to it?
3. How does the system behave when two versions of the same document contradict each other?
4. What happens when the answer is not in the sources at all?
5. What does exporting your data and index look like if you switch vendors in a year?

The first two answers usually settle the whole choice. Why questions three and four matter so much is the subject of the piece on why your AI knowledge base cannot find answers.

How to start

1. Book a free process scan: 30 minutes with an engineer, a written takeaway in two business days.
2. Prepare a list of the systems your knowledge lives in, the number of people who would ask questions, and a map of who may see what.
3. After the call you get a recommendation for one of three routes: a ready platform (and which one), a custom knowledge base, or tidying the documents first. There is a fourth answer too: your corpus may be small enough that you need neither. That case is the subject of the piece on when you do not need RAG.

Book a free process scan | See pricing

Related articles

• AI knowledge assistant for business: a knowledge base built on RAG
• Permissions before the model: RAG that betrays no closed documents
• Why your AI knowledge base cannot find answers
• AI knowledge base without RAG: when you really do not need it
• Custom AI apps