What a company second brain is
A company second brain is one written source for how your company works: your products, policies, decisions and the exceptions everyone remembers differently. Each note holds one idea, has a named owner and says whether it is still current. A new hire can read it, and an AI model can search it and answer from it. Both get the same answer.
This guide is for the owner or operations lead of a firm where answers live in inboxes, chat threads and the heads of the longest-serving staff. You have been told to do something with AI. Here is the part the demos skip: connecting ChatGPT to your shared drive doesn't give you a second brain. It gives a confident voice to every out-of-date file in the drive.
We build company second brain systems for clients, and we keep one ourselves. The job, as our service page puts it, is turning scattered knowledge into something a new hire and a model can both query without getting a different answer each time. Our AI engagements start with sitting beside the team to see how the work actually runs. One of them included a company second brain.
How it compares with what you probably have now:
| Shared drive | Wiki | Company second brain | |
|---|---|---|---|
| One answer per question | Several versions of most files | Sometimes, if someone tidied it | Yes, one note per idea |
| A named owner per page | Rarely | Sometimes | Always |
| Shows what is current | File dates only | Rarely | A status and a review date on every note |
| Written for | Whoever saved the file | People | People and AI models |
The term comes from personal note-taking, where a second brain is a system outside your head that remembers for you. The company version adds what a personal one can skip: an owner for every page, and rules about who may read what.
Why it matters now that models can read it
Looking for answers has always cost time. McKinsey estimated in 2012 that knowledge workers spent nearly 20 percent of their working week looking for internal information or for a colleague who could help. Microsoft's 2023 survey of 31,000 people in 31 countries found that 62 percent struggle with too much time spent searching for information.
In an owner-run firm, that search usually ends at the same desk. The longest-serving person answers the same question again, and loses the afternoon they meant to spend on work only they can do.
What changed is that software can now do the looking. ChatGPT's company knowledge launched in October 2025 for its Business and Enterprise plans. It searches connected apps such as SharePoint and Google Drive, and cites its sources. Microsoft 365 Copilot does the same inside Microsoft 365.
A search tool can only find what is there. If your drive holds three versions of the returns policy, the model picks one and cites it, and the citation makes the answer look checked. OpenAI says the tool runs several searches to resolve conflicting details. It can still only choose between the versions you kept.
Meta published the clearest example of doing this well on 2 September 2026: an organizational second brain for compliance work. Its agent answers from more than 200 structured knowledge files, kept apart from the instructions on how to reason. Human experts review every change to those files, and each correction joins a test suite so it stays fixed. Meta reports that assessments which took days now take minutes.
Meta has more engineers than most firms have staff. What made its system work still scales down: knowledge written in small files, and named people who review every change. The model is the easy part now. Deciding what is true, and keeping it true, is the work.
What goes in, and what stays out
Put in the answers people keep asking your longest-serving staff for. Keep out personal data, secrets, figures that change daily and raw transcripts. For anything in between, ask one question: would you be comfortable with any employee getting this answer from a model today?
| Put it in | Keep it out |
|---|---|
| What each product or service includes, and what it doesn't | Personal data about staff and customers, such as HR files |
| Policies such as returns, refunds and expenses | Passwords, API keys and bank details |
| The exceptions: the customer who needs a call before delivery, the supplier who only answers on WhatsApp | Figures that change daily, such as stock and open orders |
| Decisions, with the reason and the date | Raw meeting transcripts and chat logs |
| How-tos, such as month-end or onboarding a client | Drafts nobody approved, unless they are marked as drafts |
| Who to ask about what | Anything you wouldn't want quoted to the wrong person |
Personal data has a legal limit. GDPR, in its UK and EU versions alike, requires it to be "adequate, relevant and limited to what is necessary" for the purpose you hold it for. A knowledge base that answers everyday questions has no purpose that needs salaries in it. Leave HR files in the HR system, where access is already limited.
Secrets belong in a password manager. A model that can read a password can repeat it.
Figures that change daily belong in the system that owns them. The note should say where to look, for example "stock levels: the warehouse system", rather than copy a number that will be wrong by lunch. Meta draws the same line: dense, frequently used knowledge lives in its structured files, and sparser material is searched where it sits.
Transcripts get distilled, never dumped. The first rule in our own second brain: never save a conversation as a note; write down what it decided.
How to structure it so people and models can both use it
One idea per note. A title that states the answer. Links to the notes a reader needs next. And four fields at the top of every note: owner, status, last reviewed and source.
Keep one idea per note, because search finds passages, not whole documents. Most AI tools answering a question pull back a handful of passages and read only those. Anthropic's engineers give an example. A passage reading "The company's revenue grew by 3% over the previous quarter" is useless alone: which company, which quarter?
Adding that missing context to each passage cut failed retrievals by 49 percent in their tests, and by 67 percent with a reranking step. A note written to stand alone has its context from the start.
A title that states the answer helps both readers. "Refunds after 30 days need the owner's approval" tells a person scanning a list, and a model matching a question, what the note says. "Refunds" only tells them where to look.
Each note links to the notes a reader needs next. The refund note links to the returns policy and to the person who approves exceptions. Links let a person follow the thread, and they give the model related context to pull in.
| Field | What it holds | Why it matters |
|---|---|---|
| Owner | One named person | Someone answers for the note being right |
| Status | Current, draft or out of date | People and models can skip what is no longer true |
| Last reviewed | A date | Shows at a glance which notes nobody has checked |
| Source | Where the fact came from | Lets anyone check it |
How we organise our own second brain
Ours lives in Obsidian as plain Markdown files. A one-page operating manual sits at the top. Our AI assistant reads it before it touches anything, and a person joining would read the same rules.
Raw thoughts land in an inbox. From there they become permanent notes or get thrown away. Each permanent note holds one idea, and its title states the claim rather than the topic.
Required fields at the top of each note record its type, dates, tags, source and how mature it is. Every note links to at least two others. Once a topic reaches about five notes, it gets a map of content: a short index page that walks through them in order. A weekly review empties the inbox and finds notes that nothing links to.
Plain files matter for the same reason we hand clients their code at the end of a project. You can move them to another tool, or read them with no tool at all, and nobody can hold them back from you.
How a company second brain rots, and how to stop it
A company second brain rots in three ways. Nobody owns a page, so nobody fixes it. Changes get announced in chat and never reach the page. Old versions stay searchable beside the new ones.
Rot costs more once a model reads the pages. A person reading a stale page often notices something off, such as an old date or a colleague who has left. A model doesn't hesitate: it answers from what it was given and cites the page.
So when a well-built tool gives a wrong answer, check the note before you blame the model. Usually the note is wrong. The fixes are plain.
- Give every note one owner: the person who would be asked anyway. The returns policy goes to whoever handles returns, month-end to the bookkeeper. When someone leaves, their notes get a new owner before their last day.
- Write it down first, then announce it. GitLab tells staff to document the solution first, then announce it via Slack or email. Its handbook started when the company had ten people, because anyone joining later would never see emails about process changes sent before they arrived.
- Put a review date on every note. Once a month, list the notes past their date and send each owner their list. The owner confirms the note, edits it or marks it out of date.
- Archive out-of-date notes where the model doesn't search. Delete nothing; one day you will need to know what the policy was last year.
- Treat unanswered questions as missing notes. Keep a list of the questions the second brain couldn't answer. That list is next month's writing.
How AI uses it: retrieval and grounded answers
The model doesn't memorise your knowledge base. When someone asks a question, the tool searches your notes and pulls back the few that match. The model writes an answer from them, with a link to each source. That step is called retrieval.
Telling the model to answer only from those notes, and to say so when they don't cover the question, is called grounding.
A grounded answer to a real question looks like this:
- A colleague asks whether a customer can get a refund after 45 days.
- The tool finds the refund note, the returns policy and the exceptions list.
- The model answers from those three notes and links to each.
- If the notes don't cover the case, it says so and names the refund note's owner.
These tools follow the permissions you already have. Microsoft's documentation says Copilot can only summarise or reference content the user is authorised to access, and ChatGPT's company knowledge works the same way. The risk sits in permissions set years ago. A folder shared with everyone for convenience becomes an answer anyone can get by asking, so check your sharing settings before you connect anything.
On its business plans, OpenAI says it doesn't train its models on your data by default. Read the terms of whichever tool you use, because consumer plans can differ.
Once the knowledge is written down, software can act on it. An AI agent can draft a reply to a customer from the refund note, with a person approving it before it goes. Often you don't need an agent for that. A plain workflow that finds the note and drafts the reply is more predictable, as we explain in AI agents vs workflows.
Judgement that depends on knowledge nobody wrote down is on our list of what not to automate with AI. A company second brain is how that knowledge gets written down, and then some of it can come off the list.
Can you use ChatGPT as your company knowledge base?
As the reader, yes. As the knowledge base, no. ChatGPT, Microsoft 365 Copilot, Claude and Notion AI can all search what you connect to them. None of them knows which version of your returns policy you meant to keep.
So keep the notes somewhere your tool can reach, such as a SharePoint site or a Google Drive folder, and connect it. The answers improve the day your notes do, and you can switch tools later without rewriting anything.
A custom GPT with uploaded files is the lightweight version. It works for one team's reference set, but the uploaded copies don't change when the originals do. Someone has to upload them again, which puts you back to needing an owner and a review date.
When you don't need a company second brain yet
You don't need one yet if any of these is true.
- Everyone sits in one room, the owner answers most questions in a minute, and nobody is joining or leaving this year. A shared document with the twenty questions people ask most will do. If you plan to hire, start earlier; GitLab started at ten people.
- Nobody will own it. Without owners, you are building next year's out-of-date drive. Sort out ownership first.
- The knowledge that matters changes daily and already lives in a system, such as your bookings or your CRM. Connect your AI tool to that system directly.
- Your files are already tidy and in one place, and your office suite includes an assistant. Try it for a month before paying anyone.
- You can't yet name ten questions it should answer.
If that is you, you don't need us yet either. The next step costs nothing and tells you when that changes.
Where to start this week
For two weeks, keep a tally of the questions people ask the owner and the longest-serving staff. Every question asked twice is a note. Write the ten most frequent, one idea each, with an owner and a review date.
Then connect the AI tool you already pay for and ask it those ten questions. Where it answers wrong, fix the note, not the prompt. If the tally fills a page, or one person still answers half of it, you have found the case for a company second brain.
Questions
Should we use Obsidian, Notion, SharePoint or Google Drive?
The one your team already opens every day, as long as it passes three checks. You can export everything as plain files, you can set who may read each page, and your AI tool can connect to it. We keep our own in Obsidian because the notes are plain Markdown files we own outright. For a team that lives in Microsoft 365, a SharePoint site is often the better choice, because Copilot already reads it.
Who should be in charge of the whole thing?
One person, usually the operations lead, owns the structure: the maps, the monthly review list and the rule about what goes in. They don't write every note; each note has its own owner. In an owner-run firm, the owner shouldn't be that person. The point is to get the answers out of their head and off their desk.
How we wrote this: drafted with AI, then checked and edited by the people who build our projects. Figures come from our own work or from the linked sources.