Does your startup actually need a “company brain,” or are MCP connectors enough (eval showed this)?
How are you actually setting up your startup to be AI-native?
We’re a small team working mostly through Claude and Codex, where I am the technical guy. Documents live in SharePoint, meeting notes and crm in Notion, and we still need to work with actual Word, Excel and PowerPoint files. What we want is to make our company agent ready (and human of course) so that we can easier automate workflows with agents. My intuition is: data is key. If the data layer is well designed modt of the work is done as every agent will benefit from it.
I keep seeing “company brains” and LLM wikis everywhere. I digged into it and built one, where ocr models read in all data to md, sharepoint files are synced, all put in relations in a knowledge graph. Then connected it via mcp to claude/codex.
When I was done I expected fewer token costs and better quality on specific tasks done by agents. For evaluation i compared it with a very simple approach: just Notion connector (uses notions AI search) and sharepoint connector (m365 mcp, uses sharepoint search) with the same harness.
Results were a bit depressing: the simpler, less complex way was able to get almost the same results with almost the same amount of tokens used. So now I am totally confused as I see llm wikis and company brains videos everywhere hyped up. So what am I missing here?
Has building your own knowledge layer actually improved answers or saved tokens?
Mostly wondering how you keep things current, help agents pick the right document/version, and make sure new info gets filed properly. Ideally without tying everything to one AI provider or turning this into another software project.
Would love to hear what’s working in practice!