> ## Content Index > Fetch the complete content index at: https://hoeijmakers.net/llms.txt > Use this file to discover other available public pages before exploring further. # The company does not live in the LLM - URL: https://hoeijmakers.net/the-company-does-not-live-in-the-llm/ - Published: 2026-08-31T18:18:49.000Z - Updated: 2026-08-31T18:18:49.000Z - Description: A knowledge-intensive company should keep durable organisational knowledge outside the model, and let AI do what it is unusually good at: reason across it. - Author: Rob Hoeijmakers - Tags: AI in Practice, AI Strategy Over the past few weeks I have been using AI harder than I had before. Not just for writing, research or occasional assistance, but as part of the actual machinery of my company. I now have AI systems working with me on finance, quality, legal matters, security and product development. They use tools. They read operational systems. They develop expertise. Some of them have become genuinely useful. The obvious response is to give them more context. More instructions, more documents, more memory, more accumulated understanding of the company. For a while, I did. And then I realised I was creating a different problem. Parts of the company were starting to live inside the LLMs. A model would know why we had made a particular decision. Memory preserved something from an earlier conversation. Another agent had developed a good understanding of a particular domain. I could usually reconstruct where something had come from, but increasingly I had to ask the AI to tell me. The systems knew more about the company. I had less oversight of what exactly they knew. So I started again with the information. ## Taking the company out The question I had been asking was what an AI agent needed to know to do its job. I replaced it with a more basic one: what does the company need to know? Finance is a body of organisational knowledge. A CFO is a role that uses it. Legal knowledge exists independently of whichever model happens to be acting as Legal Lead. The same applies to quality, security, product knowledge and company strategy. That distinction matters particularly in a small knowledge-intensive company. Finance, legal, security and quality are essential capabilities, but they are not primarily what our customers pay us for. Traditionally, making each of them stronger also means adding people, suppliers and administrative complexity. AI changes that equation. A small organisation can have much deeper capability in these functions without building a department around each one. But only if the knowledge belongs to the organisation rather than to the AI performing the role. So I started extracting durable company knowledge into explicit domains. And much of it ended up somewhere I had not expected. Git. ## Old technology, new division of labour Git was created to manage software source code, but its basic properties work remarkably well for organisational knowledge. A file has a location. A change has a history. Versions can be compared. Earlier states can be recovered. Proposed changes can be inspected before they are accepted. I use GitHub, but Git is the more important choice. The knowledge itself is mostly Markdown: lightweight, readable by humans and LLMs, and able to carry metadata. ``` --- status: current owner: security review_cadence: quarterly last_reviewed: 2026-08-29 --- ``` That creates a useful division of labour. A deterministic system can establish that something has not been reviewed for six months. An LLM can help determine whether it still makes sense. The distinction goes much further than documentation. My accounting system establishes whether an invoice was paid. Project systems contain current operational state. Documents hold contracts and evidence. Email and meetings introduce new information. People add context through prompts. The LLM does not need to contain any of this. For a particular task, the relevant state can be assembled into context. The model can then do what it is unusually good at: interpret, connect, reason and synthesise. What comes out is usually an artefact: a report, analysis, spreadsheet, document, image, proposed decision or action. Sometimes the work produces something the organisation has genuinely learned. That can be reviewed and incorporated into its durable knowledge. **state → context → reasoning and synthesis → artefact → action → learning → validated state** The model participates in the loop. It does not contain the loop. ## A more capable small company There is a pleasing irony in all this. LLMs have helped me build the deterministic structures that make me less dependent on LLMs. Doing this manually would have been tedious and I probably would not have done it nearly as thoroughly. But the result is not primarily about independence from ChatGPT or Claude. It is about capability. I can make finance stronger without pretending an AI is my accountant. I can develop much deeper legal and security capabilities without confusing an agent with the underlying organisational knowledge. I can change models without rebuilding the company around them. And the same information remains available to me. I can inspect what the company knows, see where it came from and change it without asking an AI what it remembers. This feels like a particularly strong setup for a knowledge-intensive company that wants to remain relatively small while increasing the quality and breadth of what it can do. Keep durable knowledge explicit. Keep operational state in systems that can establish it. Let probabilistic systems reason across both. The more capable the models become, the more convinced I am that this is where they belong. In the reasoning seat.