Answers from your rulebook
The answers are already written down, in the staff handbook, the policies, the product sheets and the terms. People still email the same questions to the same two people, because finding the right page is slower than asking.
This assistant answers from those documents and nothing else. Each answer cites the page it came from. When the documents do not cover the question, it says so and passes it to a named person rather than improvising.
This is a blueprint, not a client case study. The grounding and citation work behind it is already in use.
General AI fills gaps with guesses
A general chat tool will answer a question about your returns policy whether or not it has read it. For a customer, that is a promise you did not make. For staff, it is HR advice nobody approved.
Security and privacy are the biggest worry small firms name about AI, cited by 44% in Simply Business's 2026 survey. An assistant that can only answer from a fixed set of your own documents is far easier to reason about than one that answers from anywhere.
Four steps, a person in charge
Reads your documents
Takes in the handbook, policies, product information and terms you choose, and nothing else.
Answers with the page
Every reply links to the passage it came from, so the person asking can check it.
Knows its limits
When the documents do not cover a question, it says so and hands over to the right person, with the question attached.
Stays current
When a policy changes, the assistant answers from the new version and the old one stops being quoted.
Same pattern, any size
For a small firm: The staff handbook and onboarding, or the questions customers ask before they buy, answered on your website.
For a larger organisation: Policy and procedure libraries across departments, where a wrong answer becomes a complaint or a claim.
On your data, first
We start with your twenty most asked questions and the documents that should answer them. You see which ones the documents actually cover, which is often the first useful finding.