Using AI with your own documents: an assistant for your expense policy

Can I take a taxi to the station after a client meeting? How much will the company pay for a hotel? The answers are in the expense policy, and people still ask finance. An AI assistant that knows the policy can answer on the spot, as long as it cites the passage and admits when the policy doesn’t cover something.

This guide walks through the setup using an invented expense policy. The same steps work for an employee handbook, warranty terms or internal work instructions.

In short

• Upload only the current version. Old versions produce contradictory answers.
• Tell the assistant to quote the deciding passage word for word.
• Test with real questions before you release it, including some the document can’t answer.

Step 1: choose the right documents

Upload the version of the policy that applies today, plus at most the FAQ or form that refers directly to it. If the 2024 and 2026 versions sit side by side, the assistant will find two answers to one question. Put the status in the file name, such as “Expense policy, valid from 2026-01-01”. Decide who updates the collection when the policy changes.

Step 2: the instructions

The instructions tell the assistant how to handle the documents. You can use this version as it stands:

You answer employees' questions about expenses.
Use only the documents provided.

For every answer:
1. Start with one sentence: yes, no or "it depends".
2. Then quote the deciding passage word for word, naming the
   document and section.
3. State any conditions in the quote, such as a time or an
   amount limit.

If the answer isn't in the documents, write:
"The expense policy doesn't cover this. Please ask
[contact person, email]."
Don't guess, and don't apply a rule to a similar case.

The word-for-word quote matters most. For long documents, the Anthropic documentation recommends pulling out the relevant quotes first and basing the answer on them. It also says to let the model admit when it doesn’t know. The quote has a practical use for you too, because you can see at once whether the answer fits the rule.

Step 3: test with real questions

Collect ten questions people have actually asked in recent months. Add two the policy doesn’t answer and one where two rules apply at once. A good answer could look like this. We invented both the policy and the answer for this guide:

Question: I’m at a client’s office in Basel until 10 p.m. on Thursday. Can I take a taxi to the station?

Answer: Yes. Section 4.2 of the expense policy says: “Taxi costs are reimbursed if the trip starts after 9 p.m. or if no public transport runs within 30 minutes.” Your trip starts after 9 p.m. Submit the receipt with your expense claim as usual.

And for a question the policy doesn’t cover:

Question: Can I claim my e-bike subscription as an expense?

Answer: The expense policy doesn’t cover this. Please ask the finance team.

Check three things in every test answer:

• Is the quote exactly what the document says? Search for the sentence.
• Is the section number right?
• Does the answer invent a condition that isn’t in the quote, such as “only with your manager’s approval”?

For the two uncovered questions, the only thing that counts is whether the assistant points to the contact person. A plausible invented answer is the worst result here. If any test question fails, adjust the instructions or the documents and test again. Keep the test questions for later changes.

In onAI

In onAI, you set up the policy as a knowledge base, then share an agent with these instructions with the groups that should use it. If your documents are in Microsoft 365, onAI can also search them live with each person’s own permissions. For a first assistant we’d still start with a small, maintained collection, because it’s easier to test.

Start with the document that generates the most questions in your company. You’ll probably find your twelve test questions in the inbox of whoever looks after it.