Excluded before indexing, not after
HR, payroll, private drives and anything already password-protected are kept out at ingest. Post-filtering an answer is not a control — a document that reached the model has already left the building.
Governance and your data
Indexed in place. Nothing moves off-site.
This is the first question we are asked in almost every engagement, so it is the first thing we answer. Your systems remain the systems of record. We use read-only connections, preserve source permissions, and document where indexes, model processing and logs run before access is granted.
The controls
Read-only, all the way down
Database access runs through a role that can only select. Statements that try to write are refused at the connection, not caught by convention.
Answers carry their sources
Every answer names the data and documents behind it, so the person reading it can check rather than trust.
A refusal looks like a refusal
When something is out of scope the engine says so and says why. It does not return an empty result dressed up in prose that reads like an answer.
Governance is what lets you roll it out to everyone
Most companies treat an AI policy as a brake. In practice it is the opposite. Clear rules about what the tools may touch are what make it safe to hand them to the whole staff instead of quietly limiting them to two people.
A person approves anything that leaves the building
AI may help estimate or suggest a price, but a qualified person reviews and approves every quote before it reaches a reseller or a customer.
AI can be confidently wrong
We put this in the policy in those words. Figures, dates and job details get verified against the source system.
AI is a tool
Humans use AI; AI does not replace human interaction or human value.
Every engagement produces a written policy
In your language, naming your systems and your folders, short enough that staff actually read it. Ask us and we will walk you through a sample.