# An AI Agent on Your Company Knowledge: What a "Company Brain" Is > An assistant that answers from your own documents and shows its sources. How retrieval works, the permissions check to run first, a 20-question test, and what it costs. - Web page: https://sporesec.com/en/blog/ai-agent-company-knowledge - Language: English - Last updated: 2026-10-17 - Also in עברית: https://sporesec.com/he/blog/ai-agent-company-knowledge.md ## Answer An assistant that answers from your own documents and shows its sources. How retrieval works, the permissions check to run first, a 20-question test, and what it costs. - By Mahdy Lorens Gribkov - Published: 2026-09-20 - Revised: 2026-10-17 - AI Agents, 6 min read - Keywords: AI on company documents, company knowledge base AI, RAG for business, AI assistant internal knowledge A "company brain" is an AI assistant that answers questions from your own documents, tickets, price lists and customer records, and shows where each answer came from. It is not a model trained on your data. It looks up the right passages when someone asks, then writes an answer from them. It is worth building when your team keeps asking each other the same questions, and it only works if your sources are current and the assistant respects who may see what. The second part is where small offices get surprised, so this post spends most of its time there, with a permissions check to run first and a 20-question test to run last. ### How does it work? The technique is called retrieval-augmented generation, or RAG. Anthropic's engineering post of 19 September 2024 gives a plain description: "RAG is a method that retrieves relevant information from a knowledge base and appends it to the user's prompt, significantly enhancing the model's response." In practice: 1. Your documents are split into passages and indexed. 2. When a person asks a question, the system finds the most relevant passages. 3. The model writes an answer using only those passages, and cites them. 4. If nothing relevant is found, a well-built assistant says it does not know instead of inventing an answer. Step 4 is the one that separates a useful assistant from a confident liar, and it is a design decision, not something the model does by default. Step 3 is a feature you ask for: Anthropic's API documentation describes citations as helping you "track and verify information sources in responses". An answer without a pointer to its source is an answer nobody can check. ### What is it good for? - **Internal questions:** "What is our refund policy for custom orders?" "Which template do we send to a late payer?" The answer exists in a document nobody can find. - **New employees:** a way to ask without interrupting a colleague. - **Sales and support:** the current price list, the terms, the technical answer, with a link to the source. - **Repeated research:** pulling the relevant past emails, quotes or tickets for a customer before a call. - **Mixed Hebrew and English material:** contracts in English, internal notes in Hebrew, a price list in a spreadsheet. A person who asks in Hebrew should get the answer from the English contract, with the source shown. ### What should you fix before you connect anything? The assistant reads what it is allowed to read. That sentence is the whole risk. Microsoft's documentation for its own assistant says Microsoft 365 Copilot "only surfaces organizational data to which individual users have at least view permissions", and adds that it is important to use the permission models in services such as SharePoint "to help ensure the right users or groups have the right access to the right content" (page updated 7 January 2026). Google's Gemini privacy page (last updated 22 January 2026) says Gemini "only retrieves relevant content in Workspace that the user has access to", and will not retrieve a document or email the user cannot open. Read that the other way and you see the problem. A company brain does not create new access, but it makes every existing access easy to use. A salary spreadsheet shared with "anyone in the company" by mistake three years ago is now one question away. So, before you connect anything: 1. **List the sources** the assistant will read: folders, mailboxes, the CRM, the ticket system. 2. **For each source, write who can open it today.** Look for sharing set to the whole company or to "anyone with the link", and for folders shared with former employees. 3. **Separate what must stay private:** payroll, contracts with named people, medical or legal files, anything with customer identity numbers. Keep them out of the assistant's sources. 4. **Decide the assistant's rule for each person:** does it answer from what that person can open, or from one shared set that everyone may see? The first needs the assistant to apply your permissions per person. The second is simpler and safer, if the set is chosen with care. 5. **Test it.** Ask as a person without access, and make sure the answer is "I cannot see that" and not the content. Whether your systems also let you log who asked what is worth deciding now. If your files live in Microsoft 365 or Google Workspace, the [comparison of the two suites](https://sporesec.com/en/blog/microsoft-365-vs-google-workspace-small-office) covers who should hold the admin account that controls this. ### What does it need to work? - **Sources that are current.** The assistant is only as right as the documents. An outdated price list produces a confidently outdated answer. Give each source an owner and a "last reviewed" date, and decide how often it is reviewed. - **A defined scope.** Start with one team and one set of questions. "Everything the company knows" is a project that never finishes. - **A test set.** Before you launch, write 20 real questions with the answers you expect. A balanced set has: eight plain lookups ("what is the warranty on product X?"), four that need two documents together, three that are not in the documents at all (the right answer is "I do not know"), three permission checks, and two where two documents disagree. Add Hebrew questions against English documents and the reverse, and dates, numbers and shekel amounts in the formats your team uses. Run the set, read the results and fix the failures. Run it again after every change. This is how you learn whether it works. - **A named owner** who reads the questions it could not answer. Those are your list of missing documents. ### What can go wrong? - **People see what they should not.** See the section above: the assistant inherits your permissions, including the mistakes. - **Stale or contradictory sources.** Two versions of the same policy give two different answers. Clean up before you connect. - **A wrong answer taken as fact.** Require citations, so a person can check the source. For anything with legal, financial or medical weight, the answer is a draft for a person to confirm. - **Customer personal data going somewhere it should not.** If the sources contain personal data of your customers, the Israeli Privacy Protection Law and its regulations apply to how you store it and which vendors process it. Treat the vendor contract and where data is processed as part of the design, and ask counsel for the legal reading. ### What does it cost? We build this as an autonomous AI agent: from ₪25,000, plus an optional ₪750 a month for light maintenance, plus VAT as required by law, with the scope fixed before work starts. For comparison with a simpler tool, read [AI agents versus automation](https://sporesec.com/en/blog/ai-agents-vs-automation): if your question can be answered by a rule, you do not need an agent. See the [AI agents service](https://sporesec.com/en/services/ai-agents), or [book a free intro call](https://sporesec.com/en/book) and bring your 20 questions. ### Sources - [Introducing Contextual Retrieval, Anthropic Engineering](https://www.anthropic.com/engineering/contextual-retrieval), published 19 September 2024, as captured by Common Crawl on 10 May 2026. The RAG definition above is quoted from this post. - [Citations, Claude API Docs](https://platform.claude.com/docs/en/build-with-claude/citations), as captured by Common Crawl on 12 June 2026. - [Data, Privacy, and Security for Microsoft 365 Copilot, Microsoft Learn](https://learn.microsoft.com/en-us/copilot/microsoft-365/microsoft-365-copilot-privacy), page updated 7 January 2026, as captured by Common Crawl on 16 February 2026. - [Generative AI in Google Workspace Privacy Hub, Google Workspace Admin Help](https://support.google.com/a/answer/15706919), last updated 22 January 2026, as captured by Common Crawl on 16 February 2026. ## Questions and answers ### Is it a model trained on my data? No. It retrieves the relevant passages from your documents when someone asks, then writes the answer from them and cites them. Anthropic describes RAG as retrieving relevant information from a knowledge base and appending it to the prompt. ### What is the main security risk? Access control. Microsoft (page updated 7 January 2026) and Google (22 January 2026) both say their assistants only retrieve what the user can already open. So the assistant inherits your sharing mistakes: check who can open each source, including links shared with everyone, before you connect it. ### What does it cost? From ₪25,000 plus an optional ₪750 a month, plus VAT as required by law. ## Next step - Book a free intro call: https://sporesec.com/en/book.md - Blog: https://sporesec.com/en/blog.md --- - Contact: contact@sporesec.com | WhatsApp https://wa.me/972532322318 | https://sporesec.com/en/contact.md - Book a free intro call: https://sporesec.com/en/book.md - Machine-readable site data: https://sporesec.com/llms.txt | https://sporesec.com/agents/en/index.json