Two people, one question, two right answers
A company brain earns trust when the same question returns different answers to different people, because those people are allowed to know different things.
Take a private equity fund. A partner and an analyst both ask the brain what is happening at a portfolio company. The partner sits on its board and can read the board pack. The analyst cannot. If the brain gives them the same answer, one of two things has gone wrong. Either the analyst has just read the board pack, or the partner got less than they needed.
More of the asking is now done by agents. Nicolas Bustamante described his agent messaging a colleague on Teams, and the colleague's agent answering from that person's own context. That raises a question for any such system: what is the second agent allowed to share, and on whose behalf? The same boundary matters when agents ask who the firm knows.
The question that decides whether a company brain is safe to roll out is the one most designs skip: who is asking, and on whose behalf? That is the line on our website, and the rest of the answer comes down to three things: permissions, setup time and hosting.
Why most company brains cannot tell
The common ways to build a company brain lose the asker's identity somewhere between the source system and the answer.
- One service account. The fastest way to ingest is a single admin credential that reads everything. The brain then knows what the admin can see, and every answer comes from that view.
- Flattened ingest. Folder access on a SharePoint estate varies a lot from one folder to the next. Ingest it as one pile and a locked folder becomes searchable the day it is indexed.
- Rules in the prompt. An instruction such as "do not share HR documents" asks the agent to withhold text it already holds in its context window. The only protection is its judgement on every turn.
- Filters in the query. Every query, every join and every new tool has to carry the right clause. It holds until somebody writes a bad query, and more of those queries are now written by a model.
Buyers notice. In May, nbaschez asked on X for a shared company brain with permissions, versioning and SSO. At the time of our review, the post had drawn 580 likes and 656 bookmarks. In our sample of 336 posts on X this year that used the phrase "company brain", 27 dealt with permissions or governance. People ask for it, and few lead with it.
Every document keeps the access rules of its source
In the 60x Company Brain, every answer respects the permissions of the document it came from. If you could not open the file in SharePoint, the brain will not quote it to you.
That holds down to the individual passage, which is what an agent actually reads and cites. It is not enough to protect the document or the workspace when an agent only ever sees excerpts.
The brain reads each source with the narrowest access that serves the use case, under your own tenant, and keeps each item's permissions and the date it last changed alongside the text. A document that five people can open stays a document that five people can open after it enters the graph.
The evidence an agent cites is the evidence the source produced. No model can alter it or invent a passage, so what your team reads back is what was actually written.
You can check all of this. Whether source permissions are enforced is a setting on your tenant, and the credential health check reports it, so your team verifies it instead of taking our word for it.
Who is asking, by credential
The brain knows who is asking from how they sign in, and each kind of asker can reach a different part of the firm's knowledge.
| Asker | Credential | Reaches | Cannot reach |
|---|---|---|---|
| A person, or an agent working for them | Signs in as that person | Company knowledge their source permissions allow, for the clients they have been granted. Their own personal brain. | Colleagues' personal brains. Clients they have not been granted. |
| An agent working for the firm | A firm-level key for one client | That tenant's company brain. | Anyone's private content, by any path. Any other client. |
Two boundaries carry most of the weight. The first sits between people. Each employee has a personal brain bound to their own credentials, built from their email, calendar and notes, and the company search never reads it. Some sources split across both: Slack channel history goes to the company brain and direct messages go to the person.
The second sits between clients. Each client's data is kept apart in the database itself, so a record in one client's brain cannot point at a record in another, even if the application makes a mistake. One operator can work across several clients with one sign-in and still reaches only the clients they have been granted.
What it looks like in practice
Three situations from recent engagements, with the clients anonymised.
The SharePoint estate
A consultancy came to us with about ten thousand SharePoint files and a CRM it did not trust. We built the brain around the files, because the file estate was accurate and the CRM was not. Folder access on an estate that size varies from folder to folder, and the brain had to respect it. Two people with different folder access now get different answers from the same brain.
Two views of the same company
Our own product shows this one. The company brain holds the firm's position on a company, assembled from CRM, documents, decks and transcripts. Your personal brain holds who you met there, when, and what was said. The two describe overlapping companies and answer to different readers, so they live in separate brains. Merging them would make a colleague's inbox searchable by anyone with a prompt.
The research that compounds
At another fund, analysts write their findings back into the brain, and the next person picks up the analysis instead of paying to regenerate it. That only works if the people reading it trust it. Every finding traces to the source passages it was built from, so a reader can check the evidence before relying on it.
Live in two weeks
Your first use case is live in about two weeks, and that changes little with how complex your firm turns out to be. Here is what you can expect along the way.
- Contracts first. An NDA, then a subscription agreement and a data processing agreement naming 60x as processor, the hosting region and every third-party provider in the chain. No real data moves before this is signed. If your security review takes longer, the build waits for it.
- One use case. We always recommend starting with one, and it is the recommendation clients argue with most. A brain shaped around one job does that job well.
- Sources, chosen on purpose. Some get held back. One pilot left out the CRM because the client was rebuilding it five weeks later, and bringing in a system about to be replaced means cleaning up after it.
- Narrow access. The brain reads each source with the least access the use case needs, under your own tenant.
- A brain shaped to your firm. We draft its structure from what your sources contain and review it with you. It can change while live. On one client call, we reorganised it around projects during the call itself.
- Your history from day one. The brain starts with six months of history in most installs, then keeps up with each day's work as it happens.
- Staged rollout. Automated checks first, then the operations team, then the wider company a week later. Write-back stays off until after the proof of concept, so nothing changes the brain while you are judging it.
- Measured accuracy. We build a reference set of questions and answers with you during setup and test the workflows against it, so accuracy claims have a number behind them. We also encourage you to run a daily health check that tests every capability and reports what broke.
Two weeks gets your first use case live. Each further use case adds to what the brain covers and the sources it reads, and every source you add keeps its permissions.
In your cloud or ours
Choose between two deployment options before any data moves.
| Option | Where it runs | What it means for you |
|---|---|---|
| 60x Cloud | Hosted by 60x, in the EU or the US, your choice. | We host it, you use it. Live in two weeks with nothing to manage. |
| Your Cloud | Inside your own Azure, AWS or GCP cloud environment, behind your existing private network. | Your data never leaves your network. |
Both options can run on open-weight models, so your data and the model answering questions about it can stay in the same place.
We price each option on its own. With 60x Cloud, 60x acts as processor under a data processing agreement, and every third-party provider is named at onboarding. Adding one later needs your acceptance as a subprocessor. Where a client requires EU-only residency, it is available with no exceptions.
Before real data moves, we provide a written security position covering the deployment and applicable SOC 2, ISO 27001, GDPR and EU AI Act requirements. Read how we handle data, sources and model-generated answers.
Some clients want the brain to hold the records themselves, instead of mirroring a system that lives elsewhere. For them, every read and write is logged, approved records cannot be edited, and a save either completes in full or not at all. That configuration runs in financial services today.
What it does not do
A permission model is only as good as what it inherits, so here is where ours stops.
- It does not fix your sources. Permissions follow the source, so a folder shared with too many people stays shared with them in the graph. Clean up access in the source system and the brain follows.
- It reads where a passage came from, and leaves what it says alone. A sensitive remark in a public Slack channel stays readable to everyone in that channel. Rules based on content are a separate mechanism, and we do not claim them.
- It does not join brains for you. A question that spans the company brain and your personal brain is answered by an agent that holds access to both.
- It does not onboard everyone at once. We take on about one client a week. The constraint is shaping each brain and getting source access.
- It is not instant after upload. New material becomes searchable once processing finishes, and the brain reports when it is ready, so nothing relies on it too early. A new personal brain takes hours to fill in, because the data providers behind it limit how often they can be queried.
How to test a company brain before you buy one
Five checks will tell you whether a company brain knows who is asking, and you can run them on any vendor, including us.
- Sign in as two people with different folder access and ask the same question. The answers should differ.
- Ask about something the asker is not allowed to read. Nothing from that source should come back.
- Run an agent on a company-level key and ask it for someone's private email. There should be no path to it.
- Ask for the source passage behind each claim. Every claim should trace to one.
- Ask the vendor where enforcement lives: in the prompt, in the query, or on the data. Only the last holds when the asker is a model.
Run these checks against our company brain. Book a demo. Bring a question that only some of your team should be able to answer.
See what is included in a two-week company brain installation.
Sources
- nbaschez, post on X, May 2026. Engagement counts as of our review.
- Nicolas Bustamante, post on X, September 2026.
- 60x analysis of 336 public posts on X in 2026 that used the phrase "company brain".