An internal AI assistant is only as good as its shelf
The two decisions that matter most in an internal knowledge assistant are what it reads from and who can ask it.
It is tempting to start an internal AI assistant by choosing the model. In my experience the model is one of the less important decisions.
What I have learned
Two choices shape the result far more:
- The source shelf. An assistant that answers from an approved, curated set of documents gives answers people can trust and check. One that reads from everything repeats whatever is outdated or wrong.
- The access scope. Deciding who can ask what, and which documents each group can reach, is a data governance decision. It should be made on purpose, before launch, not discovered later.
Grounding a service assistant in approved documentation, with access scoped by design, turned it from a novelty into something the team could rely on for real questions.
What I would check first
- Who decides which documents go on the shelf, and who removes the outdated ones?
- Can the assistant say "I don't know" when the answer isn't in its sources?
- Does every answer point back to the document it came from?