Somebody is going to pitch your business a brain this year. Y Combinator named "Company Brain" one of the startup categories it wants to fund in 2026.1 Accenture sells consulting under the banner of an "Intelligent Digital Brain" (that page was updated in April).2 A business book called The Enterprise Brain came out June 30.3 Smaller startups now quote the Y Combinator definition on their own product pages to borrow the legitimacy.
To be fair to the phrase, this is something you want. One place where you ask questions about your own business and get straight answers. It's good marketing.
Before you take the pitch meeting, one fact: none of the big names behind any piece of the pitch calls its own product a brain. I checked across everything the word bundles, from notes and wikis (Obsidian, Notion, Confluence, SharePoint) through AI search (Microsoft 365 Copilot, Glean) to AI memory (Mem0) and agents that act on your systems (Salesforce).4 The companies that build the software describe it as layers, connectors, and permissions. "Brain" comes from the consultants, investors, and startups selling the idea.
What you're actually being sold
A knowledge base is a place where your business deliberately writes down, reviews, and maintains information it intends to reuse. Somebody decided the information belongs there, and something has to keep it right, because that never happens on its own.
When a vendor says their AI is "grounded in your data", it means the AI searches your company's information before it answers, then writes its answer from what it found. Useful, and a much smaller promise than it sounds. Finding a document only proves the search worked. Whether the document is current, correct, or meant for that particular employee's eyes is a separate question.
And the "digital brain" itself? Read the architecture documentation instead of the marketing page and that word disappears. Microsoft's version is three separate layers.5 Oracle's is a database product.6 The startup versions connect to 40-odd apps and leave your data where it sits.7 Every one of them is a query layer over many systems, sold with a cool singular noun to make it easier to understand. They're sophisticated, permissioned search engines with an LLM bolted on both ends.
A business needs its CRM to be right about customers, its accounting system to be right about money, its project tool to be right about the work, and a safe way to ask questions across all three. Underneath the brain language, that is the job every one of these products is attempting.
Why your personal system works
Your personal system might be notes on your phone, a folder of Google Docs, or, at the serious end, an app like Obsidian, the one the "second brain" crowd loves. Whichever it is, I have no argument with you. Obsidian is the version worth examining, because it is the strongest case for the personal approach. You get plain text files on your own machine, readable by any tool, that work offline and search instantly. Obsidian's own pitch is that your files should outlast the app, and they built the product that way.8 That is a real virtue, and it explains the loyalty.
It works because one person is the owner, the editor, the security boundary, and the conflict resolver. Every hard question a company system eventually has to answer (who may see this, who may change this, is it still true, which version wins) has the same answer in a personal vault, and the answer is you.
The famous graph view, the one in every screenshot of glowing connected dots, is an optional layer, and Obsidian's own documentation treats it that way. Files, editing, and search are the product. Practitioners argue about whether the graph helps thinking at all, and none of the sources I read treats it as the reason anyone adopts the tool.9
I should be honest about scale, because personal vaults do decay. The tool handles reference upkeep well (rename a file and every link to it updates automatically), but the content is on you.10 Notes stop being true, orphans pile up, tags multiply, and the folder of plugins from one motivated weekend becomes its own little project. Some people now point an LLM at the vault to hunt for the rot, and fair enough. Even when it finds every stale note, deciding what is still true remains your call. At personal scale this is survivable, because it is one person's mess and one person can fix it.
The client question that breaks it
Now make it a company. Say you run a twelve-person services firm, and you ask a reasonable question about one client. What did we promise them, where are things now, what is blocked, and who is allowed to make the next move?
Walk through where the answer lives. The promise is in a signed proposal, plus deal notes in the CRM. Current status is in the project tool. The blockers are sitting in email threads and chat comments. Whether the client has paid is in the accounting system.
A folder holds the promise, and search will find it. Then you hit three questions no storage product and no search product will answer for you.
Which version of the promise is binding? The proposal was revised twice. Search happily returns all three versions, and the AI will summarize whichever it ranks first.
Is the status current right now? A copy of the project data synced last week produces a confident and wrong answer, which is worse than no answer.
Who is allowed to act? That is a rule about your business. It is written down nowhere, so no retrieval system on earth can fetch it.
Get these wrong and the failures are ordinary business failures. You quote the superseded proposal to the client, or promise a delivery date against stale status. Meanwhile the subcontractor asks the AI a question and gets the client's payment history, because the search tool had no idea contractors were different from staff.
Where the boundary actually sits
The first difference between your notes and a company knowledge base is who else is allowed in, and on what terms.
A shared vault or a shared drive can be a legitimate company knowledge system. Obsidian Sync takes up to 20 collaborators at $4 a head per month, encrypted end to end.11 If everyone in your company is allowed to see everything in the vault, you might be good to go.
But Obsidian's own help pages spell out what you are agreeing to. Fine-grained permissions "are not supported yet." Every collaborator gets the same access as the owner, the only recorded history is the last editor's name, and every edit goes live the moment it syncs. When two people edit the same non-text file, the later save wins and the earlier one disappears without a word.12
The boundary gets crossed the first time two people are allowed different things, or a stale answer costs real money, or a record has to survive a departure or a dispute, or software starts acting on the information without a person checking first. Headcount has nothing to do with it.
Sharing is cheap now. Every "brain" pitch, underneath, is charging you for governance, whether the seller uses the word or not. A better body-part analogy would be "covering your ass."
What the AI is reading when it answers
To search all your systems fast, these products usually copy your content, its details, and its access rules into an index, and the AI reads the copy. The alternative is a live lookup, where the question travels to the source system carrying the asker's own credentials. Live is slower and always current, and the major vendors document both modes, reserving live lookups for sensitive or fast-moving data.13 The copy is where the trouble lives, because of how the two modes fail. A live lookup that breaks tends to fail in plain sight, returning nothing. A copy that breaks keeps answering, and nobody can tell from the answer.
Freshness first. Glean, a leading AI search product, documents refresh rates from under five minutes on some connections to full passes 28 days apart on others, and a deleted file can keep appearing in results until the next full pass.14 Ask the AI about a document you deleted three weeks ago and it may still cite it.
Permissions are the same story. The access check runs against a copied permission list. Microsoft's Azure search service, the one developers use to build custom AI search over SharePoint files, documents that permission changes made at the site or folder level are not picked up automatically, and that until someone forces a resync, the index serves the old access rules.15 Take away access that way and the search tool can keep serving the old results.
And the person asking gets told none of this. Every staleness signal I found lives in an administrator console.16 The employee gets a confident answer in a friendly tone.
In fairness, the vendors publish most of this. The gap sits between the marketing page and the technical documentation, and the sales deck gets built from the first one.
Seven questions that sort the vendors
The following questions come straight out of the documented failures above. Take them into the meeting.
1. Where does our information stay, what copies are you making? A credible vendor names exactly what gets copied, what stays live, and what happens to the copies when the original changes or gets deleted. "Everything becomes a single source of truth," offered as a complete answer, is a warning sign.
2. If two systems disagree, which one wins and who decided? The credible answer is that you pick a winner for each kind of record and it becomes a system rule. "The AI figures it out" means nobody decided.
3. Show me the same question asked by two employees with different permissions. You want a live demo with two accounts, plus honest numbers on how long permission changes take to propagate. Worry when you hear "permissions are fully inherited" and nothing else, because the mechanism underneath is a synced copy with documented lag.
4. When we edit, delete, or let someone go, how fast does your system catch up, and what happens when a connection breaks? Credible vendors give per-connection numbers and can name who watches for failures. "It's all real time" contradicts the published documentation of every major product in this category.
5. Will it ever tell the person asking that the answer might be stale? My favorite question in the set, because the honest industry-wide answer today is no. I read the admin documentation for the major products, and the staleness indicators sit in the administrator console every single time. A vendor who says yes should demonstrate it on screen. If the vendor is surprised by the question, you have learned how often anyone else asks it.
6. When you say it learns, what does it keep, who can read that list, and how does a wrong memory die? If the product remembers things, somewhere there is a stored list of machine-written notes about your customers and your staff. You want to see it, correct it, and delete what is wrong. If the vendor answers "it gets smarter over time" and cannot show you the list, that is a hard pass.
7. What is the simpler alternative, and which parts of your pitch are product versus consulting hours? The honest answer often includes configuring the suite you already pay for. A vendor with no simpler alternative to offer, or no answer to "what would make you advise us against buying this," should end the meeting.
About memories
These products keep what the industry calls memories: small machine-written notes stored because they might turn out useful. Knowledge is information your business has decided it is prepared to rely on. The difference between them is a review step, and the review step is a person reading the thing and dciding. No product in this article supplies that person. What a product can do is make that job small enough that somebody actually does it.
Memories are also a much younger problem than knowledge bases, and some of the most interesting engineering in this category is happening there. At the scale of one person, they already pay off. In my own work, an assistant that remembers standing preferences saves real time and rarely misfires. The same boundary applies as with your notes, though. A memory that only serves me is low stakes. Once a business acts on one, it needs the review step.
One prominent AI memory product extracts facts from conversations and, by design, never deletes a fact when a new one contradicts it. Old fact and new fact both stay stored, and a ranking algorithm decides which one surfaces when asked. The vendor documents this openly.17 When one of these products remembers something, it has filed a note that no one reviewed. The same team's 2025 research paper did the opposite. When new information contradicted a memory, the system updated or deleted the old one. Within a year the shipped product dropped that approach and now keeps everything. That is how young this corner of the market is.
One engineering team published an audit of their own deployment of that product. Of roughly 10,000 stored memories, they judged 97.8 percent to be junk, including a hallucinated fact that got recalled into later sessions and re-saved as if a human had said it.18 That is a single unverified report, so read it as one team's experience rather than an industry statistic. The mechanism deserves your attention anyway. A memory system without a quality gate fills with noise, and the noise compounds.
Before an AI's unreviewed impressions of your customers start steering your staff, ask question six.
What to do next
For a lot of businesses the honest answer is that the software you already pay for is most of the system. Microsoft 365 and Google Workspace already carry permissions, search, and AI features that claim to inherit those permissions.19 Configuring what you own beats buying a parallel product, and a trustworthy vendor will say so unprompted. These suites run on the same copy-and-lag mechanics described above (the permission-sync gap earlier is a Microsoft product), so the same questions apply, just to software you already pay for.
If you misjudge your own suite, you lose some configuration time. A bad purchase means paying for unused seats until the contract runs out.
Sometimes centralizing is right, too. Deliberately copying a chosen slice of data into one place can beat wiring live connections into everything, and Microsoft's own documentation frames copy-versus-live as a tradeoff decided by how sensitive and how fast-moving the data is.20 Worry when a product centralizes everything by default, because that tradeoff deserves a decision.
If it is just you, local notes and search are enough. You have outgrown them when other people keep asking you to forward things.
A small trusted team gets a long way on shared documents, named owners, and a light review habit. That stops working the first time you need to hide anything from anyone.
Once you run multiple systems with different access levels, keep records in the systems that own them, and add search or AI that provably respects the source permissions. The permission and staleness questions come before any final decision. Be suspicious of any tool that wants to become the new home for data that already has one. Check the pricing model early, too. Glean publishes no rate card, and third-party estimates run $40 to $75 per user per month with minimums around 100 seats.21
Automation that writes to your systems comes last, after the systems of record, the approval steps, and the audit trail exist.
The least glamorous finding in the research is also the oldest. Intranets decayed in the 90s, wikis in the 2000s, portals in the 2010s, and the retrospectives name the same cause, which is that the humans stopped updating them.22 No product in this article removes the need for a person who decides what is true and keeps it that way. Whatever you buy will work about as well as that habit does.
Three questions of ownership
The "brain" wave will keep coming, because investors are now funding startups specifically to sell it. Some of the underlying technology is genuinely good. The label tells you nothing either way.
As shorthand for wanting straight answers about your own business, "company brain" is fine. In this research, though, the people who build the software talked about layers, connectors, and permissions, and the word "brain" came from the people selling the idea.
So the three questions to carry into any pitch are who owns the truth, who is allowed to see it, and who is allowed to change it. A vendor with good answers is worth your meeting, whatever they decide to call the product. And if nobody at your company can answer those questions today, no product will answer them for you. Deciding what is true is a job. Name the person who owns it before you spend anything.
Footnotes
-
Y Combinator, Requests for Startups, Summer 2026 batch, verified July 14, 2026. The Company Brain category, authored by YC's Tom Blomfield, describes a system that pulls knowledge out of fragmented sources, structures it, keeps it current, and turns it into "an executable skills file for AI." ↩
-
Accenture, From digital to intelligent enterprise systems, updated April 14, 2026. ↩
-
Ragy Thomas, Sravan Vadigepalli, and Chandhu Nair, The Enterprise Brain (Fast Company Press, June 30, 2026), per the publication announcement. ↩
-
Marketing and product pages checked July 14, 2026: Obsidian, Notion, Confluence, SharePoint, Microsoft 365 Copilot, Glean, Mem0, and Salesforce Agentforce. ↩
-
Microsoft, Fabric IQ: the semantic layer powering trusted AI agents, June 2, 2026. The layers are Work IQ, Fabric IQ, and Foundry IQ. ↩
-
Oracle, Introducing Oracle AI Agent Memory, part of Oracle AI Database. ↩
-
For example, Coworker.ai markets 40 to 50 app integrations for its organizational-memory product. ↩
-
Obsidian, About Obsidian; Steph Ango (Obsidian CEO), File over app. ↩
-
Obsidian, About Obsidian, which presents files, editing, and search as the base product. For the practitioner debate on graph views and recall, see the Zettelkasten forum and Kristoffer Balintona on backlinks. ↩
-
Obsidian, Internal links. ↩
-
Obsidian, Sync for teams and pricing, accessed July 14, 2026. Every collaborator needs their own Sync subscription. ↩
-
Obsidian, Collaborate on a shared vault, Version history, and Sync troubleshooting, accessed July 14, 2026. Markdown files auto-merge; other file types resolve by last save. ↩
-
Microsoft documents synced versus federated connectors, recommending federation for sensitive or dynamic sources, in the Microsoft 365 Copilot connectors overview. Glean documents indexed, live, and hybrid connector modes in How connectors power Glean. ↩
-
Glean connector documentation: crawling refresh rates and deletion handling, accessed July 14, 2026. ↩
-
Microsoft, Use a SharePoint indexer to ingest permission metadata, Azure AI Search documentation (2026-05-01-preview API), updated July 2, 2026: parent-scope permission changes require an explicit resync, and without one "the index serves stale ACL data for previously ingested files." ↩
-
Glean's connector health and change-rate metrics live in the admin console per its crawling FAQ. No product documentation reviewed for this article describes a staleness indicator shown to the person searching. ↩
-
Mem0, OSS v2 to v3 migration guide: memories accumulate, new facts are stored alongside old ones, and retrieval-time ranking decides which one surfaces. The earlier design, with model-decided updates and deletions on contradiction, is described in the team's 2025 research paper. ↩
-
Issue #4573 on the mem0 GitHub repository, a third-party production report, unverified by the vendor at the time of writing. ↩
-
Microsoft markets Microsoft 365 Copilot as inheriting Microsoft 365 permissions, sensitivity labels, and retention policies. Vendor claim; the mechanics behind it are the synced-copy model described above. ↩
-
Microsoft, Copilot connectors overview: synced connectors index broadly, federated connectors fetch live for sensitive, dynamic, or live data sources. ↩
-
Third-party pricing estimates: GoSearch and CheckThat, accessed July 14, 2026. Glean publishes no list pricing. ↩
-
For one such retrospective, see ELO Digital Office, Why enterprise search is replacing traditional knowledge management, itself a vendor with a commercial stake in the framing. ↩