"Training" sounds technical, but for a small business it mostly means writing things down. If you're wondering how to train an AI receptionist, the honest answer is that you don't teach it to talk; the language model already does that. You teach it about your business: what you do, what it costs, when you're open, what you won't do, and when a person needs to step in.
Most of that information already lives in your head, your price list and your staff's habits. The job is to get it into a form the assistant can use. This article walks through it in the order we'd do it, using Klugon's settings as the worked example, and finishes with a testing routine you can repeat every week.
How to train an AI receptionist: what it can and can't learn
Before filling anything in, it helps to know how the assistant uses what you give it.
- It answers from what you've written. Your hours, address, services and documents are given to the assistant, and it searches your knowledge base when a caller asks something specific. If the answer isn't there, a well-set-up assistant says it doesn't know and offers to take a message, rather than guessing.
- It doesn't learn from calls on its own. It won't pick up that you stopped doing Saturday appointments because a caller mentioned it. You change the settings, and it follows them.
- On the phone, it's brief. Klugon's phone assistant is told to keep replies to a sentence or two, to read details back to the caller to confirm them, and not to make things up. Long documents still help, because it searches them for the relevant part, but the caller only hears the short version.
So the quality of your answers depends almost entirely on the quality of what you put in. That's good news: it's in your control.
Step 1: the basics callers ask about most
Start with the facts that come up on nearly every call. In Klugon these live under Settings, and they reach the phone assistant directly.
General
Business name, type, a short description, website, main language and time zone. The description matters more than it looks: a couple of plain sentences such as "Family-run plumbing and heating firm covering Leeds and Bradford. We do boiler servicing, repairs and bathroom fitting, but not new-build work" saves the assistant guessing what you do and don't offer. Get the time zone right too, or "are you open now?" goes wrong.
Contact
Address, phone and email. Callers ask "where are you?" and "can I email you something?" constantly. If parking or access is awkward, put that in the knowledge base as well (see step 4).
Hours
Your opening hours, day by day. Keep them identical to your Google listing, which you can update from Google's Edit your Business Profile help page, including special hours for holidays. Two sources saying different things is how customers end up at a locked door.
One thing to understand: Klugon's assistant answers whenever a call reaches it, at any hour. Your opening hours tell it what to say ("we open at nine tomorrow"); they don't stop it answering. If you only want it to take calls after hours, set that up with call forwarding, which our call forwarding setup guide explains.
Step 2: services and prices
The next most common questions are "do you do…?" and "how much is…?". In Klugon, each service on the Services page has:
| Field | What to put in it |
|---|---|
| Name | What customers call it, not your internal code ("Men's haircut", not "MCUT1") |
| Description | What's included, who it's for, anything to bring or avoid. The page notes that your AI can use this to answer questions |
| Category | Groups similar services |
| Duration | How long to book for |
| Price and currency | The real price, or a "from" price explained in the description |
| Staff who do this | Who can be booked for it |
| Bookable | Whether the assistant can book it, or only talk about it |
Writing service descriptions that answer questions
A good description answers the follow-up question before it's asked. Compare:
- Thin: "Full colour."
- Useful: "Full head colour with a toner and blow-dry. Allow two hours. A patch test is needed at least 48 hours before if you're new to us or haven't had colour here in six months."
The second version lets the assistant handle "how long does it take?" and "do I need anything first?" without a handover.
Prices that vary
If a price depends on the job, say how. "Callout from £75, including the first hour; parts quoted on site" is far better than leaving the price blank, because blank invites the assistant to say it doesn't know, and the caller rings someone else. If you never quote prices on the phone, say that in the description and in your custom instructions (step 5).
Step 3: booking rules
If the assistant books appointments, it needs the same rules your front desk uses. In Klugon, those are in the Availability panel on the Bookings page:
- Start times every: how often slots begin, for example every 15 or 30 minutes.
- Break between appointments: buffer time for cleaning down, travel or notes.
- Minimum notice: how soon a booking can start. The default is 60 minutes; a mobile trade might want a day.
- Book up to (days ahead): the default is 60 days.
- Holidays or days off: closed dates, so nobody gets booked on Christmas Eve.
Klugon books into its own calendar. On voice or social Business plans and above it can also check Google, Outlook or Apple calendars for busy times once connected, so personal appointments block slots too. It doesn't send confirmations or reminders, so if you rely on those, keep them in whatever system already sends them.
Once it's set, use Check what the AI would offer to preview the slots the assistant would give a caller for a service and date. It uses your saved settings, opening hours and existing bookings, so it's the quickest way to catch a wrong buffer or a missing day off. Our guide to automating appointment booking goes further into booking rules.
Step 4: the knowledge base and FAQs
Everything that isn't hours, contact details or a service goes in the knowledge base. In Klugon you can add knowledge in four ways:
- Upload files. PDF, DOCX, TXT, MD or CSV, up to 20 MB each. Scanned PDFs (pictures of pages) can't be read, so export a text version if you can.
- Paste text. The quickest route for an FAQ list or policy you already have in an email or document.
- Import a web page or scan your site. A single page is read once; if it changes, add it again. The website scan reads your homepage plus up to seven key pages.
- Quick links. A library of links (booking page, location, policies, social profiles) with a label, a category and a short description that helps the assistant know when to share it.
Knowledge is shared across all your assistants, so you only write it once.
What to put in your FAQ
Ask whoever answers your phone today to list the 20 questions they hear most. Then write a plain answer to each, in the words you'd want said on the phone. Typical topics:
- Parking, access and how to find you
- Payment methods, deposits and cancellation fees
- Areas you cover, for trades and mobile services
- What to bring or how to prepare
- Age limits, accessibility, pets, children
- Guarantees, aftercare and what happens if something goes wrong
- What you don't do, and who to try instead
Writing answers the AI can use
- One topic per answer. "Parking" and "cancellation policy" in separate entries, not one long paragraph.
- Put the answer first. "Yes, there's free parking behind the building, entrance on Mill Lane" beats a paragraph of history.
- Use the customer's words. If people ask about "the MOT", don't only call it "annual vehicle inspection".
- Include the "no" answers. A clear "we don't do emergency call-outs after 8pm" is as useful as anything you do offer.
- Date anything temporary. "Closed for refurbishment from 3 to 10 November" ages better than "closed next week".
If you're unsure what callers will ask, our AI receptionist scripts article has examples by business type.
Step 5: custom instructions
Custom instructions are the rules your assistant follows in every conversation. Klugon's settings note that numbered lists work best and that being specific helps. Both the assistant's own instructions and your business-wide instructions reach the phone assistant.
This is where you put your judgement calls: the things a new receptionist would learn in their first week. A sample set for a made-up physiotherapy clinic:
- Never give medical advice. If a caller describes symptoms, say a physiotherapist will assess them at the appointment.
- If someone mentions chest pain, difficulty breathing or a serious injury, tell them to call 999 straight away.
- New patients book an "Initial assessment", never a follow-up.
- Don't quote prices for insurance-funded treatment; take their insurer's name and say we'll confirm.
- If someone asks for a specific physio who isn't available, offer the next available slot with them first, then with anyone.
- Always ask for a mobile number when booking.
Keep each rule to one instruction, and avoid vague ones such as "be helpful", which the model does anyway. If a rule depends on information (prices, policies), put the information in the knowledge base and keep the rule about behaviour.
Step 6: handover rules
No AI receptionist should handle everything. Decide in advance what goes to a person, and write it down.
How handover works in Klugon
Klugon can't transfer a live call to you. Instead, with the Human handoff skill switched on, the assistant takes the caller's details and a note of what they need, saves them as a lead, and flags the conversation so it shows as needing a person in your inbox. On voice or social Business plans and above, Slack or Zapier can notify you. There are no email or push alerts to the owner, so decide who checks the inbox and when.
What to hand over
Write a short list into your custom instructions, for example:
- Complaints and anything involving a refund
- Quotes for jobs that need a site visit
- Questions the assistant can't answer from your knowledge
- Anyone who asks for a person, twice
- Anything urgent or safety-related, with an instruction to call emergency services where relevant
Set expectations about the callback
Tell the assistant what to promise: "someone will call you back within two working hours" only if that's true. It helps to say which number you'll call from, too. Hiya's State of the Call 2026 report found that 86% of calls from unknown numbers go unanswered, so a caller who's expecting you is far more likely to pick up. Our article on AI receptionist handover to staff covers the callback side in more detail.
Collecting the right details
In the Lead capture tab you choose which details to ask for (name, email, phone, company, message), how soon to ask, and whether to note buying intent. For phone calls, a name, a number and a one-line reason are usually enough. If you collect personal data, your privacy information should say so; the ICO's guidance on the right to be informed explains what UK businesses need to tell people.
Step 7: the greeting and voice
In the assistant's Voice tab you choose one of ten voices, the language, and the call greeting (up to 500 characters). If you leave the greeting blank, callers hear the standard one: "Hello! Thank you for calling" followed by your business name, then "How can I help you today?"
A custom greeting is worth writing. Keep it short, name the business, and if you like, say what the assistant can help with. Our business phone greeting examples has templates.
On the phone, Klugon works in English, German, French, Spanish, Italian, Portuguese, Dutch, Japanese and Arabic. Write your knowledge in your main language; the assistant answers in the phone language you set for it, so check a few answers in each language you expect.
Step 8: test before customers call
Testing is the part people skip, and it's where most problems show up. During Klugon's 7-day trial you can't connect a phone number yet, so you test the voice assistant in your browser. That's no bad thing: you can run dozens of calls in an afternoon.
A test script
Run each of these, and note anything wrong:
- The top ten questions from your FAQ list, asked the way customers ask them, not the way you wrote them.
- A booking for your most common service, then check it appears in your bookings.
- A booking that should fail: a closed day, too short notice, a non-bookable service.
- A price question with an awkward answer ("how much for a boiler?").
- Something you don't do, to check it says no politely.
- A request for a person, to check details are taken and the conversation is flagged.
- A question that isn't in your knowledge, to check it admits it doesn't know.
- A rule from your custom instructions, for example a symptom question for a clinic.
- A caller in a second language, if you serve one.
- A mumbled or interrupted caller, to check it confirms details back.
Fixing what you find
When an answer is wrong, find the cause before rewording anything:
| What happened | Usual fix |
|---|---|
| Wrong fact | Correct it in Settings, Services or the knowledge base |
| Right fact, too vague | Make the source answer more specific, answer first |
| Didn't know something it should | Add an FAQ entry using the caller's wording |
| Broke a rule | Make the custom instruction more specific, one rule per line |
| Offered wrong slots | Check the Availability settings and the preview |
| Didn't hand over | Check the Human handoff skill is on and the handover list is clear |
Keep teaching it: a weekly review
An AI receptionist gets better the same way a new member of staff does: someone checks its work and fills the gaps.
Klugon's Insights page has two lists that make this quick: "Questions without a good answer" and "Most asked questions" over the last 30 days. Once a week:
- Read the questions without a good answer, and add an FAQ entry for each real one.
- Skim the most asked questions, and check the top answers are still right.
- Read a handful of conversations from the inbox, especially any flagged for a person.
- Update anything that's changed: prices, staff, holidays, seasonal hours.
Fifteen minutes a week is usually enough once the first month is past. The first two weeks need more attention, because that's when you discover the questions you didn't think of.
Frequently asked questions
How long does it take to train an AI receptionist?
For a small business with an existing price list and FAQ, the first setup usually takes an afternoon: an hour for settings and services, an hour or two for the knowledge base and rules, and an hour of testing. Expect to keep adding answers for the first couple of weeks.
Do I need technical skills?
No. You fill in forms, upload documents and write plain-language rules. The most useful skill is knowing what your customers ask.
Does the AI learn from calls automatically?
No. It answers from what you've given it. You improve it by reviewing questions it couldn't answer and adding to your knowledge base.
What if a caller asks something it doesn't know?
A well-set-up assistant says it doesn't know and offers to take a message. Klugon's phone assistant is told not to make things up. Those questions then appear on the Insights page so you can add answers.
Can it transfer callers to me?
Not with Klugon. It takes the caller's details and a note, flags the conversation for your team, and on voice Business plans and above can notify Slack or Zapier, so you can call back.
Should I upload my whole website?
The website scan is a quick start, but marketing copy rarely answers practical questions well. Add a short, direct FAQ as well, and update it as new questions come in.
How do I stop it promising things we can't do?
Write the limits into your custom instructions as numbered rules, and add "what we don't do" answers to your knowledge base. Then test those exact questions.
Can I have different assistants for different numbers?
Yes. Each phone number is assigned to one assistant, and each assistant has its own settings, greeting and instructions. Your knowledge base is shared across them.
The businesses that get the most from an AI receptionist aren't the ones with the cleverest settings; they're the ones that wrote down what their best receptionist already knows, then kept it up to date. If you want to try it with your own business, Klugon has a 7-day trial, and the pricing page lists what each plan includes. For the wider picture of covering your phone, see how to never miss a customer call.