A knowledge base for a voice agent: writing answers to be spoken

The short answer

A knowledge base for a voice agent should not be written like a web page. On the phone the caller does not see the text; they only hear it: long paragraphs, tables, links and brackets do not work in speech. A good spoken answer gives the main answer in one or two short sentences, keeps numbers in a form that can be pronounced, keeps lists to three items, and asks "shall I go into more detail?" when needed.

The knowledge base also needs two more things: a list of topics the agent must never answer, and a log of questions the agent could not find an answer to. This article shows the spoken answer format, forbidden answers and working with unanswered questions.

Why an answer written for text fails on the phone

A website FAQ answer lets the reader scan for the relevant part, go back, look at a table. On the phone none of that exists: the caller hears the answer once, in sequence, and has to remember it. By the end of a 120-word answer they have forgotten the beginning. "More details are on our website" is useless on the phone — a link cannot be read out.

So the same knowledge can exist in two forms: a full text for the website and chat, and a short spoken version for voice. Some systems have the agent shorten long text itself, but an approved short version is more reliable.

The spoken answer format

1. 1Answer firstThe first sentence answers the question: "Yes, we're open on Saturday from 10:00 to 16:00."
2. 2Then the conditionAn exception or condition in the second sentence: "Hours may change on public holidays."
3. 3A short listIf a list is needed, no more than three items: "You'll need a passport, a certificate and a photo."
4. 4An offerIf there is more detail: "Shall I go through the document requirements in more detail?"

Numbers, dates and names

• Write prices fully and clearly: "forty-five manat", with the currency
• Phone numbers in groups: "012 — 555 — 12 — 34"
• Dates with the day name: "this Friday, 3 October"
• Expand abbreviations: "value added tax" instead of "VAT", at least the first time
• Foreign names as pronounced: note how a brand name is said

A number error is one of the most serious errors on the phone: the caller writes it down and relies on it. Numbers such as prices and dates should be checked separately and updated as soon as they change.

The forbidden answers list

The knowledge base should record not only what the agent will say but also what it will not. This list is the agent's boundary and works together with the scenario.

• Medical, legal and financial advice
• Individual prices, discounts and exceptional terms
• Comparisons and opinions about competitors
• Internal information: headcount, revenue, plans
• Products and services not yet announced
• Any information about other customers

For each forbidden topic the agent needs a ready sentence: "I can't give information on that — shall I connect you with a colleague?" Simply staying silent or changing the subject confuses the caller.

Source owner and version

Every answer needs an owner and a last-reviewed date. Sales owns the price answer, operations owns opening hours, legal or customer service owns the returns policy. When an answer changes, the old version is kept — when a caller says "they told me a different price", you need to see which version the agent used that day.

The unanswered questions log

For every question it cannot find an answer to, the agent does something: hands over, records it, says "we'll confirm and get back to you". The list of those questions is the most valuable source for developing the knowledge base. Review the log once a week and decide for each question: write a new answer, extend an existing one, or add the topic to the forbidden list.

1. 1CollectPull the questions the agent could not answer from the transcripts.
2. 2GroupMerge similar questions and count them.
3. 3DecideNew answer, extension or forbidden.
4. 4CheckTest the new answer with a test call.

Illustrative example: two versions of one answer

Not a real customer case. The website version: "Orders are delivered within 1–2 working days in Baku and 3–5 working days to the regions. Delivery is free on orders over 50 manat; otherwise it costs 5 manat in Baku and 8 manat to the regions. Times may change on public holidays. More details: delivery page."

The spoken version: "It takes one to two working days in Baku and three to five to the regions. Delivery is free on orders over fifty manat. Shall I tell you the price as well?" If the caller wants, the agent gives the second part. The information is the same, but the spoken version can be heard in one breath.

One knowledge base for chat and voice

Keeping separate knowledge bases for voice and chat leads to diverging answers: a price is updated in one place and not the other. A better approach is one source, two formats: the core fact in one place, with the short spoken version beside it. When the fact changes, both formats are checked at the same time.

What to measure

• The size and weekly trend of the unanswered questions log
• The number of answers last reviewed more than 90 days ago
• Complaints about number or price errors
• The average length of spoken answers — in seconds
• The share of correct handoffs on forbidden topics

Common mistakes

• Giving website text to the voice agent as it is
• Keeping links and tables in answers
• Not writing a forbidden topics list
• Not analysing unanswered questions
• Separate, out-of-sync knowledge bases for chat and voice

Limits

Even a well-written knowledge base does not guarantee every answer will be correct: the agent can misunderstand a question or choose an unsuitable answer. So the knowledge base must work together with testing, sample checks and handoff to a person. Ownership and the update routine are a separate governance question.

The knowledge base in Vexvon

In Vexvon AI Call Center the call agent answers from the same knowledge base as the chatbot: a price you write once works in both chats and calls. During a call the agent finds answers with a knowledge-base search tool; on a question it cannot answer, or a difficult one, it passes the call to a live operator. Every call's transcript stays in the panel — that is exactly where the unanswered questions log is collected from.

Ownership and the update routine are covered in knowledge management, and checking the knowledge base during a pilot in the voice AI pilot plan.

First step

Read the answers to your ten most frequent questions aloud and time them. Turn any longer than 20 seconds into a spoken version. More articles are in the reliability & pilots section; to review your knowledge base together, get in touch.