Support calls answered from your own documents, not guesses
Most support lines answer the same thirty questions all day. Where is my order. What are your opening hours. How do I return this. Can I change my booking. A voice agent can answer those from your own documents, in your own words, and say "I don't know" when the answer is not there. The catch is in that last sentence: it is only as good as the documents you give it.
A knowledge base is a shelf, not a brain
It helps to picture it right. The knowledge base is a shelf of documents next to the agent. When a caller asks something, the agent looks along the shelf, finds the passage that matches, and answers from it, often quoting the words you wrote. It is not trained on your documents and it does not "learn" them. It looks them up, every time, live.
That has two consequences. Update the document and the next call uses the new text. And if the answer is not on the shelf, a well-built agent says so and offers a message or a transfer, rather than inventing a returns policy. Pick an engine that checks its own replies for exactly this; we explain why in stopping a voice agent from bluffing.
What to upload
Start with what your team already answers from memory. In rough order of value:
- Opening hours, address, parking, how to find you.
- Prices for standard things, with the date they were last checked.
- Policies: returns, cancellations, deposits, warranties, what happens when someone is late.
- The top twenty questions, each with a two-sentence answer.
- Product or service descriptions in plain words, without the marketing.
- Step-by-step guides for the three things people most often get stuck on.
Formats are not the hard part. PDF, Word, plain text and web pages all work. What matters is that each document is about one topic, has clear headings, and says when it was last reviewed.
Leave out anything you would not read aloud to a stranger: internal notes, staff phone numbers, customer data, drafts. The agent cannot tell a private document from a public one.
Write for a listener, not a reader
Documents are usually written to be read on a screen. The agent reads them aloud over a phone, and that changes what works.
Put the answer first, then the reason. "Returns are free within 30 days. Bring the receipt or your order number." beats three paragraphs of context. Tables do not speak well; turn them into sentences, one per row. Long lists become a blur by the fourth item, so cut them or group them. Write numbers the way you would say them, and spell out abbreviations the first time.
If there is a phrase you want the agent to use, write it in the document as a sentence. If there is a phrase you never want it to use, put that in the agent's instructions instead. Documents are for facts; instructions are for behaviour.
Keeping it current
A knowledge base rots quietly. The price changed, the summer hours ended, the returns window moved, and the agent keeps saying the old thing in a confident voice.
Give every document one owner and a review date, and put the date in the document itself so the agent can say "as of September". When you replace a file, remove the old one. Two versions of the same policy is the fastest way to get two different answers. For anything that changes weekly, point the agent at a live web page instead of a file, and keep that page short.
After every change, run the same five test questions in the browser and listen to the answers. It takes three minutes and catches most mistakes before a caller does.
Close the loop with the calls themselves
The calls tell you what is missing. After each one the agent writes a summary and an outcome. Once a week, look at the calls where it could not answer or had to transfer, and read them. Every "I don't know" is either a document you should write or a question you do not want the agent answering. Both are useful to know.
Watch for the opposite failure too: answers that were technically in a document but did not help, because the document assumed knowledge the caller did not have. Rewrite those in plainer words. And keep the human path open. Some questions should go to a person on the first ask, and the instructions should say which; overflow to a colleague is a feature, not a failure.
Where NordTell fits
NordTell's knowledge base takes PDF, Word, text and web pages, and the agent quotes from them on the call and says when it does not know. With the hybrid engine a second model checks each reply before it stands, and the agent always keeps a way to a person for the questions that need one.