Practical FDO guide
AI knowledge base: build a company brain agents can use
An AI knowledge base makes approved business information retrievable by people and AI systems. A useful company brain includes source ownership, dates, permissions, definitions, and reviewed examples. Retrieval can ground a response in company content, but it does not guarantee accuracy, current information, or permission enforcement by itself.
By AI Agentship · Published · Updated
Start with a workflow and its real questions
Choose the work the knowledge must support: preparing a client brief, answering a process question, or drafting a report. List the questions a competent team member would need to answer and the approved sources behind each answer. Uploading every document does not define a useful knowledge system.
A starter record can contain title, source URL or location, owner, approved use, client boundary, version, last review date, and expiry rule. Keep methods, metric definitions, policies, and reviewed examples distinguishable so the system can use the right kind of evidence.
What retrieval-augmented generation does
Retrieval-augmented generation, or RAG, retrieves relevant material and supplies it to a model as context for a response. Its usefulness depends on whether the correct information is retrieved and whether the response uses that evidence appropriately. A vector database is one possible component, not the whole business operating system.
Keep source references visible to the person checking the answer. Test questions that use different wording from the documents and questions that need information from more than one source. More retrieved text is not automatically better context.
Make permissions and missing evidence explicit
Apply access boundaries before retrieval and output. A shared prompt telling an agent not to expose another client is not sufficient isolation. Configure controls in the source systems, retrieval service, workspace, and agent runtime used by the implementation.
Define what happens when a document is stale, sources disagree, or the user lacks access. A useful response can say that evidence is missing and route the question to its owner. The system should not invent a policy or silently treat an old example as a current rule.
Evaluate the company brain against accepted work
Create a small, representative question set with expected sources and reviewer notes. Include an answerable question, an ambiguous question, a missing answer, a stale source, and a prohibited cross-client request. Record source relevance, unsupported claims, correct refusal or escalation, and review effort.
Connect the knowledge base to one useful workflow before expanding. An FDO organizes the knowledge, access, instructions, and review around its first valuable result. Measure Time-to-Value to that verified event, then check whether knowledge maintenance and repeated use remain sustainable.
Implementation checklist
- Name the workflow and its knowledge questions.
- Record source ownership, dates, and approved use.
- Configure access outside the prompt.
- Return evidence with the answer.
- Test stale, missing, and prohibited information.
- Measure a verified workflow result and ongoing maintenance effort.