How to give AI your company context (without a six-month project)

A short answer first: you do not train a model on your company. You connect the material you already have to a system that retrieves from it, keep each person's permissions intact, and require every answer to name the document it came from. That is it. Fine-tuning a model on internal documents is the expensive way to get a worse result.

The reason people expect it to be harder is that the first attempt usually is. Someone pastes a policy into ChatGPT, gets a good answer, and concludes the company needs "an AI project". Six months later there is a data-lake proposal and nobody can ask a question yet.

What "context" actually means here

Your company already knows the answer to almost every question it gets asked. The knowledge exists — in a drive, in a closed thread, in the head of the person who negotiated that contract. What is missing is a way to ask.

So the goal is not to teach a model your business. It is to make your own material answerable, in the tools people already use, with the permissions you already have.

Four steps

1. Pick one painful question. Not a category — a question. "What delivery time do we promise?" "Which clients are on the old pricing?" The one where somebody interrupts a colleague every week.
2. Connect the sources that answer it. Google Drive, Notion, GitHub, Linear, Slack. Not everything you own; the two or three places that answer *that* question.
3. Ask it, and check the citation. An answer without a source is a guess with confidence. If the system cannot show you the document, you cannot deploy it to anyone who is not you.
4. Only then widen it. Add the next question, the next source, the next team. Value arrives on day one and compounds; a big-bang rollout does neither.

Three mistakes that cost the most

Fine-tuning a model on internal documents. It is slow, it is expensive per update, the model cannot cite anything, and it will state a superseded policy with the same confidence as the current one. Retrieval is not the cheap alternative to fine-tuning here — it is the correct one.

Ignoring permissions until later. The moment an AI assistant answers from company material, it can leak. Salary tables, unsigned contracts, a client's data in a support thread. Permissions cannot be added afterwards, because "afterwards" is the day someone asks the wrong question and gets an answer. The rule worth insisting on: an agent asking on someone's behalf inherits that person's access, never more.

Trusting a chat log as memory. Conversation history is not organisational memory. It is per-person, per-tool, and it disappears. What you want persisted is the decision and its source, not the transcript that produced it.

What ValorBrain does about it

We built the product around exactly the two constraints above: answers cite the document they came from, and permissions are enforced per person — including when an AI agent asks on someone's behalf. It connects to the sources listed above, and it answers inside Claude, Cursor and any MCP client as well as in its own interface.

On public benchmarks, when the answer exists in the material, we find it in the top 10 results 96.58% of the time (LoCoMo, 1986 questions). We publish where we fall short too.

Documents and vectors are stored in Brazil under the LGPD (answer generation uses a cloud model; the subprocessor is listed publicly), and on-premise runs entirely on your own infrastructure. Starting is free: free plan, no credit card. Five minutes to connect the first source is a fair estimate; a six-month project is not.