How to Train an AI Receptionist on Your Business FAQ

Nobody wants to hear it, but the truth about any AI phone assistant is boring and slightly deflating: it is only as good as what you tell it. The clever part — understanding a caller, holding a conversation, sounding human — is largely solved for you. The part that decides whether callers hang up happy or hang up confused is the part you own: the answers. This is a guide to writing them well.

When people ask how to train an AI receptionist, they usually picture something technical — uploading data, tweaking settings, maybe a bit of coding. It isn't that. The work is closer to writing the crib sheet you'd hand a smart new hire on their first morning: here is what people call about, here is what to say, and here is what to never say. Do that job carefully and the assistant sounds like it has worked at your business for years. Skip it, and it sounds like it read your website once and is now bluffing.

The good news is that you already own everything you need. The questions callers ask are not a mystery — you and your staff answer them fifty times a week. The trick is getting them out of everyone's heads and into a form the assistant can actually use. Let's walk through it.

Why the FAQ is basically the whole job

It helps to be clear about what "training" means here, because the phrase makes it sound more mysterious than it is. An assistant like Vunoon doesn't learn by watching your calls or absorbing your business by osmosis. It answers from a knowledge base — a structured set of facts and answers you give it during setup. Everything it says traces back to something you wrote down.

That's a feature, not a limitation. It means the assistant never invents a price, never guesses your parking situation, never confidently tells someone you're open on a Sunday when you're not. But it also means a gap in your knowledge base becomes a gap in the conversation. If a caller asks something you never covered, the assistant has to fall back on taking a message — which is fine occasionally and embarrassing if it happens on the most common question you get.

So the goal of this whole exercise is unglamorous and specific: assemble the twenty to forty things your callers actually ask, and write down what a great human receptionist would say to each. Get that right and you've done ninety percent of the work of a good AI phone assistant.

Step one: mine the questions people actually ask

The instinct here is to sit down and try to imagine every question a caller might have. Don't. You'll produce a neat, logical list that has almost nothing to do with what real people phone about. Real questions are messier, more repetitive, and more mundane than the ones you'd invent. So instead of guessing, go and collect the evidence — it's sitting in four places.

Your own memory, and your staff's

You and whoever answers the phone are a living archive of frequently asked questions. The problem is that the knowledge is so automatic you've stopped noticing it. The fix is to make it visible. For one week, keep a sheet by the phone — paper is fine — and put a tally mark every time someone asks something, plus a scribble of what it was. You are not aiming for elegance; you're aiming for a raw frequency count.

By Friday you'll have a picture that surprises you. It's rarely the interesting questions that dominate. It's "are you open now," "do you take walk-ins," "where do I park," "do you do X," and "how much is Y roughly." The boring questions are the whole ball game, because they're the ones that repeat. Rank your list by tally marks and you have your priority order handed to you.

Your inboxes and message threads

People ask the same things by phone, email, and message — so your email inbox, your booking-form messages, and your WhatsApp or SMS threads are a searchable transcript of your FAQ. Skim the last two or three months. You're looking for the questions that recur across different customers: the same worry phrased five slightly different ways is a signal that it belongs in the knowledge base.

Pay special attention to the questions you find yourself copying and pasting an answer to. If you've built a canned reply for something, that's your subconscious telling you it's a frequently asked question. Those canned replies are, quite literally, draft answers you can lift almost word for word.

Reviews, and what people type into search

Two smaller but revealing sources. First, your online reviews — especially the lukewarm ones. When someone writes "lovely place but I couldn't work out if I needed to book ahead," they've just told you about a missing FAQ entry that's costing you goodwill. Second, if you have a website with any analytics, the search terms people type on your own site are questions in disguise. A spike in searches for one thing means it isn't answered clearly enough anywhere.

Between these four sources — memory, inboxes, reviews, and search — you'll rarely need to invent a single question. Almost everything worth answering has already been asked. Your job is collection, not creation.

Step two: write answers in spoken language, not website language

Here is the mistake almost everyone makes, and it's an understandable one: they copy the answers straight off their website. It seems efficient. The information is correct, it's already written, why not reuse it? Because website language and spoken language are different languages, and a phone assistant lives entirely in the second one.

Website copy is written to be scanned by the eye. It's dense, it front-loads keywords, it uses phrases no human says aloud — "We offer a comprehensive range of services tailored to your individual needs." Read that sentence out loud on a phone call and the caller will physically wince. A spoken answer to "what do you do" is shorter, warmer, and specific: "We do haircuts, colour, and treatments — mostly for women, but plenty of men come in too. What were you after?"

The technique that fixes this is almost silly in how well it works: say the answer out loud and transcribe yourself. Imagine the caller has just asked the question. Answer them as if they were real. Then write down exactly what came out of your mouth, contractions and all. That transcript is your answer. It'll be a little scruffier than written prose, and that's the point — it's meant to be heard, not read.

1. 1
   Take one question

   Pick the top question from your ranked list. Just one — don't try to draft the whole set in a single sitting.
2. 2
   Answer it out loud to an imaginary caller

   Speak naturally, the way you would at the counter. Don't perform; just answer.
3. 3
   Write down what you actually said

   Contractions, casual phrasing, the lot. Resist the urge to "clean it up" into formal prose.
4. 4
   Trim the waffle, keep the warmth

   Cut any repetition, but leave the natural, human phrasing intact. Aim for two or three spoken sentences.
5. 5
   Add a natural next move where it fits

   Many answers should end by nudging the call forward: "Shall I book you in?" or "Want me to have someone call you back?"

One detail that separates a good knowledge base from a great one: let answers end with an action. A human receptionist doesn't just answer "yes, we do that" and fall silent — they say "yes, we do that, would you like to come in this week?" Teaching your assistant to do the same turns a Q&A machine into something that actually books work. Where an answer naturally leads somewhere — a booking, a callback, a message — write that nudge into it.

The facts that hold everything up

Before the nuanced answers, there's a layer of plain facts the assistant needs as scaffolding — the things it should never have to improvise. These aren't really "answers" so much as the ground truth every answer stands on. Get them exact, because a confident wrong fact is worse than a graceful "let me take a message."

• Opening hours — including the exceptions: late nights, half-days, the fact that you're shut between one and two on Wednesdays.
• Location and access — the address as a human would describe it, plus parking, the tricky entrance, the buzzer that doesn't work.
• What you do and don't do — your actual services, and just as importantly the things people assume you do but don't.
• Rough pricing, where you're comfortable stating it — even "it starts from around this and depends on the job" is far more useful than silence.
• How to book, pay, and cancel — the practical mechanics people phone about constantly.
• Who to reach for what — which enquiries the assistant should take a message for versus answer itself.

Notice that several of these include the awkward exceptions — the Wednesday lunch closure, the entrance round the back. Those exceptions are exactly what callers phone to check, and exactly what a website tends to leave out. A knowledge base that includes the messy real-world details is worth more than a tidy one that omits them.

Step three: decide what stays off-limits

This is the step almost nobody thinks of, and it's the one that separates a trustworthy assistant from a liability. Just as important as deciding what the assistant should say is deciding what it must never say. An assistant that will confidently answer anything is not a strength — it's a way to make promises you can't keep and give information you shouldn't.

Some things you genuinely can't answer over the phone without risk, and the honest, professional move is to route them to a human. A physiotherapy clinic's assistant should never diagnose a caller's back pain; it should book them in to be seen. A law firm's assistant should never offer even a whiff of legal advice; it should take the details and have a solicitor call back. The boundary isn't a weakness to hide — knowing what not to answer is what makes callers trust the answers you do give.

Draw your off-limits list deliberately. For most small businesses it covers a handful of categories: anything that amounts to professional advice (medical, legal, financial), firm commitments the assistant has no authority to make (custom quotes, guarantees, deadlines), sensitive personal matters, and complaints — which almost always deserve a human ear. For each of these, the instruction is the same: don't guess, take a message, and hand it to a person.

• Professional advice — medical, legal, financial, or anything a caller could act on to their harm. Book them in or take the details.
• Firm commitments — exact custom quotes, guaranteed timelines, promises about outcomes. State ranges if you like; never invent specifics.
• Sensitive or emotional situations — bereavements, disputes, distress. A human should handle these, and the assistant should say so kindly.
• Complaints — never argue, never dismiss. Apologise briefly, capture the detail, and get it in front of you fast.

There's a reassuring reason a well-built assistant stays inside these lines by default: it answers from your knowledge base rather than from the open internet. If you didn't give it a price for a bespoke job, it won't manufacture one — it takes a message instead. Its natural instinct is caution, and your off-limits rules simply reinforce where that caution should be firmest. It also won't pretend to be human if a caller asks it directly, which is exactly the kind of honesty that keeps trust intact.

Step four: test it the way a real caller would

Once you've fed in your questions, answers, and boundaries, resist the urge to declare victory. The gap between what you think you wrote and how it actually sounds on a call is always bigger than you expect. So test it — and test it like a slightly awkward customer, not like the person who built it.

The fastest way is to simply talk to it. With Vunoon you can test the assistant by having a conversation with it before you ever forward a real number — you ask it things, it answers, and you hear immediately where it's crisp and where it flounders. Go in with your ranked question list and work down it. Then, crucially, go off-script: ask the weird, half-formed, slightly rude questions real callers ask, because that's where the gaps hide.

1. 1
   Run your top ten questions verbatim

   Ask exactly what your tally sheet says people ask. Confirm the answers sound right out loud, not just correct on paper.
2. 2
   Ask them again, badly

   Rephrase each question the way a distracted, non-expert caller would. "So like, do you do the thing with the... you know, the colour?" It should still land.
3. 3
   Probe the boundaries

   Deliberately ask something off-limits. Make sure the assistant declines gracefully and takes a message rather than winging it.
4. 4
   Try to get it lost

   Ask something obscure you didn't cover. A good response is an honest "let me take a message" — a bad one is a confident guess.
5. 5
   Listen for tone, not just facts

   Even correct answers can sound cold. Note anywhere it feels stiff, and rewrite that answer more warmly.

Every place it stumbles is a to-do item, not a verdict. Stumbles are the whole point of testing — each one is a missing answer or a badly worded one that you can now fix in a couple of minutes. It's far cheaper to find these gaps yourself, in a test, than to discover them in a lost booking three weeks later.

Step five: review the gaps every month

A knowledge base is not a document you finish. It's a thing you keep. Your business changes — a new service, a price adjustment, holiday hours, a room that's out of use for renovations — and every change is a chance for the assistant to say something that used to be true and now isn't. The maintenance is light, but it isn't zero.

The mechanism that makes this easy is already built into how a good assistant works. Vunoon sends you a summary and transcript of every call, so you have a running record of exactly what people asked and how the assistant handled it. That record is your monthly review, done for you. You don't have to guess where the gaps are — the calls tell you.

Once a month, put twenty minutes in the calendar and skim recent call summaries with two questions in mind. First: what got asked that wasn't in the knowledge base? Any question that ended in "let me take a message" when it shouldn't have is a missing answer — add it. Second: what did the assistant say that's no longer true? Prices, hours, services you've dropped or added. Fix those and you're done for the month.

This monthly rhythm is what turns a decent assistant into one that quietly gets better all year. The first version you build will be good. The version six months later — sharpened by real calls, with every recurring gap filled — is the one that genuinely sounds like it has always worked for you.

What a realistic first hour looks like

It's easy to read all this and imagine a project that swallows a weekend. It doesn't have to. Imagine a two-chair hair salon setting this up. In the first ten minutes the owner scribbles down the fifteen questions she answers every single day — hours, prices roughly, do you do men's cuts, do you take walk-ins, where to park. That covers the overwhelming majority of her calls before she's even started on the rare ones.

Over the next half hour she writes each answer the way she'd say it, ending the booking-related ones with "shall I get you in the diary?" She spends ten minutes deciding her boundaries — no medical advice about scalp conditions, take a message and she'll call back. Then she talks to the assistant for fifteen minutes, catches three clumsy answers, fixes them, and forwards her number. Total: a bit over an hour, most of it just writing down things she already knows. That's the honest scope of the job.