Writing the FAQ Your AI Receptionist Answers From
How to choose, word and organise the questions and answers an AI receptionist uses, so callers get short, correct replies and unknowns are handled well.
An AI receptionist is only as good as what it can read. If its sources are vague, it will be vague. If they are wrong, it will repeat the mistake politely and fast. The most valuable job in setting one up is not choosing a voice but writing the FAQ.

A phone FAQ is different from a website FAQ. People are listening, not reading. They cannot scan, they cannot click, and they forget the second half of a long sentence. This guide shows how to write for that situation. For the technology behind the lookup, see retrieval for voice agents.
Collect the real questions
Start with what callers ask, not what you wish they asked. Real questions are messier than the ones on your website.
- Listen to a week of calls or ask whoever answers the phone for the top ten questions.
- Check your inbox and chat for the same questions in written form.
- Look at the search terms people use to find you.
- Ask a new member of staff what confused them in week one.
Group the questions into themes such as hours and location, services, booking, prices, payments and problems. Our list of the 20 requests small businesses get most by phone can help you spot gaps.
Write answers for listening
A spoken answer should be short enough to hear once and remember. In Call, an agent is designed to reply in one or two short spoken sentences, without lists or links. Write your sources the same way.
Written answer vs spoken answer
| Written for a web page | Written for the phone | |
|---|---|---|
| Length | A paragraph with details | One or two sentences |
| Structure | Bullet lists and tables | Plain flowing sentences |
| Links | See our page for more | Offer to take a message instead |
| Numbers | 09:00-17:30, Mon-Fri | "We are open from nine to half past five, Monday to Friday" |
| Names | Abbreviations and codes | The spoken name people use |
Make details speakable
Write hours, dates and amounts the way a person would say them. "Opens at nine" sounds right; "09:00" can be read oddly. Spell out unusual names in a form the voice can pronounce, and avoid abbreviations that exist only on paper.
Structure each entry
A good entry has three parts. Keep them separate in your source so you can edit them without breaking the others.
Anatomy of a good entry
- The questionPhrase it the way a caller would say it, and add one or two variants.
- The answerOne or two short sentences with the key fact first.
- The next stepWhat the caller can do now, such as leave a message or ask for a callback.
Variants matter. Callers ask "are you open on Saturday" and "what are your weekend hours" and "can I come in at the weekend". Add each as an alternative to the same answer, or write the answer so the lookup finds it from any of them.
Decide what the agent should not answer
Equally important is what the agent should not say. It is designed to answer only from its instructions and sources, and to say it cannot confirm something when nothing matches. Help it by listing the topics that need a person.
- Prices that vary by job or customer.
- Medical, legal or financial advice.
- Anything that changes by the day, such as stock or availability, unless you keep the source current.
- Complaints and cancellations that need a decision.
For each, write an entry that sets expectations: "That depends on the details, so I will take a message and the team will call you back." This is better than silence, and it teaches the agent your handover. See taking accurate messages with an AI receptionist.
Keep it fresh
A stale FAQ is worse than a missing one. Opening hours change at holidays, services change, staff leave. Give the content an owner and a review date.
Monthly FAQ review
- Opening hours and holiday changes updated
- Any new service or product added
- Questions from the last month's messages added
- Answers callers pushed back on rewritten
- Old promotions and one-off notices removed
- A test call that asks five random questions
Use call history to find gaps. Look for places where the agent said it could not confirm something, and decide whether to add the answer or leave it to a person. For a method, see finding knowledge gaps from calls.
Some sample entries
These are generic examples for a hypothetical small clinic. Adapt the wording to your business.
| Question | Spoken answer |
|---|---|
| Where are you? | We are on the high street, next to the pharmacy. I can take a message if you need directions. |
| Do you take walk-ins? | Most visits are by appointment. I can take your details and the team will confirm a time. |
| What are your hours? | We are open from nine to five on weekdays and closed at weekends. |
| Can I speak to a person? | Of course. The team is not available just now, so I can take a message or note a good time to call you back. |
Notice that none of the answers pretends to confirm a booking. The agent takes a request and passes it on. For this flow see handling appointment requests with an AI receptionist.
Tone and order of the answer
The first sentence of an answer should be the answer. Callers who wait through three sentences of preamble will interrupt, and an agent that allows interruptions has to cope with that. Put the key fact first, then one helpful extra, then the next step. "We are open until five today. Would you like me to take a message?" works better than a long welcome followed by the hours.
Match the tone of your business as well. A funeral director, a skate shop and a law firm all answer "what are your hours" differently. Read giving your AI receptionist the right tone for ways to set this once and keep it consistent across every answer.
Test the FAQ before callers do
Before you set the agent live, test it with questions in different words, including the awkward ones. Ask the same question three ways. Ask something you deliberately left out. Ask something that sounds like a covered topic but is not, such as asking for a price when you only gave hours.
- Does it give the same answer however the question is phrased?
- Does it admit when it does not know, instead of guessing?
- Does it offer a sensible next step each time?
- Does it avoid saying more than your sources say?
- Do numbers and names sound natural when spoken?
Fix what fails by changing the source, then test again. This loop, test then edit then test, is the real work of making an AI receptionist reliable. See testing voice AI for a longer method.
Doing it in Call
In Call, an AI agent reads its instructions, its persona, the script attached to the line and the passages of its knowledge base that match the question. You can add question-and-answer pairs, text snippets and website pages as sources, and keep reusable replies under quick replies. If the agent keeps saying it cannot confirm something, add that fact to a source. Under AI agents, a page shows what the agent learned from calls and lets managers revert a change, and the AI call history lets you read what was said. Read knowledge sources in Call for the details.
Frequently asked questions
How many questions do I need?
Start with the ten to twenty most common questions. That usually covers the majority of calls, and you can add more as real calls show gaps.
Should I copy my website FAQ?
Use it as a starting point, then rewrite each answer for listening. Shorten it, remove links and spell out times and numbers.
What if two answers conflict?
Fix the sources. If your documents disagree, the agent may use either, so keep one source of truth for each fact.
Can the agent learn from calls by itself?
Some systems adjust from calls and saved answers. In Call, you can see what it learned and revert any change, so you stay in control of what it says.
How do I handle prices?
If a price is fixed and public, you can state it. If it varies, say that it depends and offer a message or callback so a person can quote.
How often should I review the content?
Weekly in the first month and monthly after that, and whenever your hours or services change.