Brose
Tier one automotive supplierA company assistant 20,000 people actually use
An internal AI assistant for Brose staff across 76 locations in 25 countries. Ask it a work question, get an answer from the company own documents.
Click any box in the diagram to see why it is there
Problem
Brose is a tier one automotive supplier with people spread across 76 locations in 25 countries. The answers they needed were already written down, in documents most of them could not find, in departments with different rules about who may read what.
Approach
One assistant, four departments behind it. Finance, HR, Controlling and Shopfloor each have their own material, and the assistant answers from the company documents rather than from open ended general knowledge. Employees can also act through it, so raising an incident, a security incident or a support ticket and checking its status happens in the same place as asking the question.
Architecture
Permissions are enforced at retrieval, not at the answer. Azure AD groups decide which documents can even enter the context for a given person, so the model is never holding material the reader is not cleared for.
Constraints and scale
About 50,000 messages a day, counting questions and answers together. An evaluation framework tests both answer quality and access correctness, and a failure on either one blocks the release.
At a glance
- Answers from company documents for Finance, HR, Controlling and Shopfloor, rather than guessing from the open internet.
- It does things, not just answers. Staff report a problem, raise a support ticket and check its progress without leaving the chat.
- It only reads what you are allowed to read. Permissions are applied at the moment documents are fetched, using the company existing Azure AD groups, so nothing you are not cleared for ever reaches the answer.
- Every release is tested for two things before it ships: is the answer correct, and did it respect access rules. Failing either one blocks the release.