01 / Private AI
What a company LLM needs before it becomes useful at work
Architecture notes on knowledge boundaries, retrieval quality, permissions, evaluation, and the human decisions that should stay visible.
Start with the job
A company LLM becomes useful when it helps someone complete a specific piece of work. That could be preparing an account brief, checking a policy, reviewing a case, or deciding what needs attention next. The workflow gives the system a boundary and gives the team a practical way to judge whether it is helping.
Model choice matters, but the surrounding product usually matters more. Retrieval quality, source freshness, permissions, response design, and the route to human review determine whether people can trust the system during real work.
Permissions are part of the product
A private AI system should respect the same access rules as the source systems it reads. A useful answer is still the wrong answer if it reveals information the person should not see. Access control needs to follow the request through retrieval, generation, citations, stored history, and any action an agent can take.
The interface should also make boundaries visible. People need to understand which workspace they are in, which sources were searched, what was excluded, and when an answer may be incomplete.
Evaluate answers and actions separately
An answer can be checked for relevance, evidence, completeness, and clarity. An action needs a different standard because it changes a real system. Teams should define which actions can run automatically, which require confirmation, and which should remain manual.
A strong evaluation set comes from real questions and real edge cases. It should include outdated documents, conflicting policies, restricted sources, vague requests, and situations where the correct response is to ask for more context.