For AI to work the way your company works, you do not need to train a model. You need to organize what the company knows in a place the AI can read while it works. We use Tiago Forte's second brain method: capture everything that comes in, organize it into four folders (projects, areas, resources and archives), write each process down only once and split the work with a simple rule. AI takes the volume, which is capturing and organizing, and people keep the judgment, which is deciding and producing.
Do you need to train AI on your company data?#
Almost never. Training a model adjusts the style of the answers, costs a lot and ages with every process change. What makes AI work the way your company works is context: the company knowledge organized in a place the AI reads when it does the task. The real work is organizing, not training.
When a business owner asks how to "train AI on company data", they almost always want something else. They want the AI to know how the company works: who the clients are, how a quote is put together, what a contract looks like, what was agreed in the last meeting. None of that calls for a new model. It calls for that information to be written down and easy to find.
At Epicora we have not trained any model. Even so, the AI writes scopes, opens tasks in our system and drafts contracts our way, because it reads our way written down before it starts. It is the same as with a new hire: nobody expects results on day one without first introducing the company, giving context and showing how the work is done there. We explain the levels of delegation in chat, assistant or AI agent; this article covers what comes before them, which is the place where the AI goes to get its context.
What a company's second brain is#
It is a place where everything the company knows is stored and organized to be used. The method comes from Tiago Forte's book Building a Second Brain, published in 2022, before ChatGPT arrived. It has four steps: capture, organize, distill and express. It was made for people, and AI simply made it cheap to apply.
Tiago Forte calls these four steps CODE:
- Capture: collect what comes in, from everywhere. Meetings, audio, messages, client emails.
- Organize: store things by where they will be useful, not by topic.
- Distill: actually read it and turn it into a decision.
- Express: use it to produce something, such as a proposal, a scope or a plan.
The method was born for the personal life of knowledge workers, and that is why it works so well with AI. AI works with the same material: text someone captured, organized and left ready to use. When the company already has that material in order, the AI finds what it needs. When it does not, the AI fills the gap with what it knows from the internet, and the result comes out generic.
How to capture what the company knows#
Capturing means writing down what is lost in conversation today. The cheapest starting point is the meeting: recorded and transcribed, it becomes a document. Client messages, emails and audio go along with it, into the folder of the project they belong to. What was said now exists in a place the AI can read.
At Epicora, capture started out of curiosity. We began transcribing client meetings to see if it worked, and today that is the start of our sales process. The meeting is recorded, the video call tool generates the transcript, and it goes in whole, unedited, into that client's folder. The summary and decisions go into a separate document written from it.
Two rules keep capture from becoming a mess. The first is to separate what came from outside from what the company wrote: the transcript, the email and the spreadsheet the client sent go into an inputs subfolder, and what the company produces from them sits next to it. The second is not to filter too much while capturing. Deciding what matters is a later step, and the full transcript has already saved more than one decision nobody had written down.
How to organize so the AI finds what it needs#
Organize by where the information will be useful, not by topic. Tiago Forte proposes four folders: projects, which have a start and an end; areas, which never end; resources, which serve as reference; and archives, with what is finished. The AI finds what it needs because each thing in the company has exactly one folder.
Folders by topic work like a library: everything about "contracts" in one place, everything about "clients" in another. When you work on a project, the information is spread across five folders. Folders by use work like a tool: everything a project needs sits in its folder. Our internal repository, which we call cortex, is organized exactly this way.
| Folder | What it holds | In our repository | Example |
|---|---|---|---|
| Projects | what has a start and an end | one folder per client, for the 27 projects underway and the 25 deals in negotiation | a client's scope, prototype and transcripts |
| Areas | responsibilities that never end | governance, team, relationships, weekly priorities | the rules for how the company decides |
| Resources | reusable reference | the 25 standards for how we do each delivery | how to write a scope, how to estimate, how to publish an app |
| Archives | what is finished or lost | delivered projects and deals that did not close | a closed project, out of the way but searchable |
Each project folder has one document that states its current status, and that document is the source of truth. When anyone (a person or the AI) wants to know how a project is going, the answer is in one place. The company's structural decisions are also recorded, with the reason: there are 36 so far, and none has to be argued again because someone forgot why it was made.
Write each way of working down only once#
The knowledge that matters most to the AI is how each process is done, written so a new person could follow it. Each process lives in a single document, and the AI reads that document when it runs. When the process changes, it changes in one place, and every following task already comes out the new way.
We have 25 processes of our operation written as standards: scope, estimate, prototype, app publishing, notifications map, among others. On top of them there are 28 AI routines, and none of them copies the standard into itself. They read the standard when they run. It sounds like a detail, but it avoids the most common mistake of beginners: writing the process into the AI instructions and, months later, having two versions of the same process, the document and the instruction, with the AI following the old one.
An example of how this starts. The person in charge of quality and project management, who is not a developer, spent days copying tasks from the scope into our system, module by module. She taught the AI to do it from thousands of tasks we had already created and dozens of earlier scopes. Today the AI reads the new scope and opens the tasks our way, and another person on the team uses the same routine and helps improve it.
The next step, which we are taking now, is writing down every process in the company, including the non-technical ones, such as handling leads, contracts and office routines. We learned two things right away. Processes are written by role, never by person, because a document describing what one person does dies when that person leaves. And there are two different queues: a process that lives in someone's head costs half an hour of conversation to write down, while a process that does not exist yet costs a decision by the partners.
Who does each part: AI or people#
AI takes the volume and people keep the judgment. Capturing and organizing are volume work, and that goes to the AI. Distilling and expressing, which mean deciding and producing, stay with whoever answers for the result. If the AI distills for you, you stop learning from your own business.
| Step | What it is | Who does it at Epicora | Example |
|---|---|---|---|
| Capture | record what comes in | the AI, from the recording and the messages | the meeting transcript goes into the client folder |
| Organize | store it in the right folder | the AI, following the folder rules | the minutes and decisions go into the project, dated |
| Distill | turn it into a decision | a person, with the AI helping to read | what does this meeting change in the scope? |
| Express | produce something with it | a person leading, the AI writing the first draft | the proposal, the revised scope, the reply to the client |
This split also protects the company from a trap that appears when everything works well. A system that decides on its own gets more and more comfortable to use, and the team slowly loses the habit of thinking about what it delivers. Keeping distilling and expressing with people is what guarantees someone still understands the business. For the tasks where the AI acts on its own, the rule for what it may do without asking is in how to make AI follow your company's method.
What changes in practice when knowledge is organized#
The speed at which knowledge becomes delivery changes. In our quoting process, the meeting transcript, the messages and the client emails go into the project folder, and assistants that follow our method help write the scope, build the prototype screens and estimate the effort. Today the client receives all of that together with the quote.
Before, clients signed based on a document of a few pages and only after closing received the full scope to approve. The discovery stage that came after signing no longer exists, because what it produced is now ready before the proposal. None of this is done by the AI alone: each stage has a person leading it, and the AI does the heavy lifting of reading, organizing and writing the first version.
The same repository feeds other routines. Every morning, a routine reads the official artificial intelligence sources, matches each piece of news against the projects in the projects folder and emails only what could help one of them. It can only make that match because the projects are described in a place it reads. In customer service, the same principle shows up differently: the service agent of Agência OKSE answers only from the official material it was given, and it does not make up what is not in that material.
Where to start without a technology project#
Start with one folder per client and transcripts of your meetings. Then write down one process that repeats every week, the way you would explain it to a new hire, and test it with the AI. Do not start by buying a tool, or by dumping everything the company has into one place. Organization comes before volume.
Three mistakes show up often among beginners:
- Buying the tool before the method. A knowledge base platform with everything mixed together delivers less than simple, well-organized folders.
- Putting everything in, including what should not go. Use the company account in the AI tool, because on business plans what you send is not used to train the model, and keep personal data only where the task needs it.
- Trying to document the whole company before using it. A written process only improves when someone uses it. One process a week, tested with the AI, yields more than a complete manual nobody opens.
When that knowledge needs to connect to the company systems, such as the ERP, the CRM or the task tool, and the AI starts acting on them, that is development work, which is what we do in AI and automation.
Frequently asked questions#
Do I need to train an AI model on my company data?#
Almost never. Training a model adjusts the style of the answers, costs a lot and has to be redone every time a process changes. What makes AI work the way your company works is context: organized documents it reads when it does the task. We have not trained any model, and the AI writes scopes, opens tasks and drafts contracts our way, because it reads our way written down.
Which tool should I use to store the company documents?#
The tool matters less than the organization. The method works in Google Drive folders, in Notion or in a repository of text files, which is what we use. What must exist in any of them are the four folders (projects, areas, resources and archives), one folder per project and each process written in a single document. An expensive tool with everything mixed together delivers less than simple, well-organized folders.
Can I put client documents in this knowledge base?#
Yes, with two rules. The first is to use the company account in the AI tool, because on business plans what you send is not used to train the model. The second is to keep personal data only when the task needs it: a scope needs the client's problem, not anyone's ID number. The rest of the company knowledge can and should go in.
How long does it take to set this up in my company?#
The structure takes an afternoon: the four folders and the first project folder. What takes time is writing the processes, because many of them live in someone's head or do not exist yet. That is why it pays to start with one process that repeats every week and grow from there, instead of trying to document the whole company before using it.
Will the AI get confused by outdated documents?#
It will, if the old document sits next to the current one. That is why there are two rules. Whatever is finished goes to the archives folder, which gets it out of the way without losing it. And each piece of information lives in one document only: when a rule changes, it changes in one place, and no old copy is left for the AI to read as if it were the right one.
Sources#
- Forte Labs, "Building a Second Brain: The Definitive Introductory Guide", by Tiago Forte, with the CODE method (Capture, Organize, Distill, Express): https://fortelabs.com/blog/basboverview/
- Forte Labs, "The PARA Method", by Tiago Forte, with the four folders (Projects, Areas, Resources, Archives): https://fortelabs.com/blog/para/
Next step#
If your company knowledge is already written down and the next step is for the AI to act on your systems, that is the work we do in AI and automation. And if you are still deciding how much to delegate to the AI in each task, the criteria are in chat, assistant or AI agent.