The problem
The answers already exist in handbooks, SOPs and old emails, but finding them means interrupting the one person who knows. New hires ask the same questions for months.
How it works
- Approved documents are synced from Drive or SharePoint on a schedule, split into passages and indexed for search.
- A question comes in through Slack, Teams or a simple web page.
- The most relevant passages are retrieved and the model writes an answer using only those passages.
- Every answer shows its sources and if nothing relevant is found, it says so instead of guessing.
What it needs from you
- A folder of current, approved SOPs and policies (we help clean it up)
- A decision on who can see which documents
- An owner who keeps the source documents up to date
- 20–50 real staff questions to test against
Why it's built this way
Answering only from retrieved passages keeps the model from making things up and makes every answer checkable.
Permissions follow the documents, so people only get answers from what they're allowed to read.
How it stays under control
- Answers only from approved, indexed documents
- Source citations on every answer
- Explicit 'not found' instead of guessing
- Document-level access permissions
- Thumbs up/down feedback reviewed monthly
The path to autonomy
Internal answers can go live quickly because a person is always the reader. Customer-facing use waits until the test questions pass consistently and the 'I don't know' rate is acceptable.