Your most exposed IP may be how your company knows what to do
Your most exposed IP may be how your company knows what to do.
Companies worry about sending the confidential PDF, the source code or the customer list to an LLM provider. Those are the obvious secrets.
The less obvious one lives in people. A senior consultant knows which question to ask next. An engineer knows when a specification that looks right is wrong. A salesperson can tell which "great" opportunity will never close. Almost none of it is written down.
Now agents sit inside those workflows. Every time an expert corrects one, explains why the obvious answer is wrong, or shows it the next step, that judgment goes into someone else's system.
So "sovereign AI" needs a stricter meaning than "our documents aren't retained." Using AI should never teach another party how your organization works. That means protecting three things:
Data: they don't keep our documents.
Knowledge: they don't learn our facts, heuristics and experience.
Capability: they don't learn to reproduce what we are uniquely good at.
The third matters more as agents move from answering questions to watching and running whole processes.
The answer is a Map of Work: a private, structured record of your processes, decisions, exceptions, corrections and outcomes. Start with one recurring activity. Capture how your best people actually do it, where they depart from the standard process, and why.
Then treat the LLM like a CPU. Anthropic Claude, OpenAI GPT and Google Gemini become interchangeable, and each receives only the context a single task needs. Models do better with more context, so this is a real trade. Minimizing data is no longer enough. Minimize context.
Documents tell you what a company knows. A Map of Work captures what it knows how to do.
I wrote up how to start building one.