Suppose you want an assistant that answers questions about internal procedures. People could stop interrupting colleagues or searching through folders. It sounds useful.
Before designing the interface, I would want to know whether the answers can actually be found in the company’s documents. And whether those documents agree.
Start with questions people already ask
Collect a few questions that colleagues deal with at work. How do they approve an unusual discount? Who should receive a particular customer request? Try to find a current answer and a supporting document for each one.
You may find a problem before you reach AI. An old document says one thing, a chat message says another, and a colleague explains that everyone does something else in practice.
Someone needs to decide which instruction applies. You can then give the assistant the correct material and test its answers. With an ambiguous question, watch whether it asks for clarification or simply makes a recommendation.
Decide what evidence you need
Set a time limit for the trial and agree who will check the answers. A specific rule helps: if an answer is not supported by an approved document, count it as wrong during the test, even if it sounds reasonable.
Keep incorrect answers with their source material. After changing the documents or instructions, you can return to the same questions and compare results. Add new questions too, so you are checking more than the examples you used to make improvements.
You may find that the assistant can handle only a narrow set of questions for now. You could limit it to that scope or wait until the documents are updated. Either result tells you what is worth paying for next and which spending would be premature.