How to write a prompt for an AI phone receptionist

Voice is not chat

On a phone call there is no screen. The caller can't scroll back, can't see a list and won't wait through a long paragraph. They interrupt, mumble, change their mind and talk over background noise. An AI receptionist's instructions have to be written for that.

This guide walks through the parts of a good receptionist prompt, the knowledge and tools behind it, and a simple way to test it.

1. Decide what the receptionist handles

List why people call you: look at a week of calls or ask whoever answers the phone. A few reasons usually cover most calls. Good jobs for an AI receptionist:

• Questions with clear answers: opening hours, location and parking, prices, services, what to bring, the cancellation policy.
• Booking, moving and cancelling appointments, when it can check your calendar.
• Taking a message with a name, number and reason, so your team can call back prepared.
• Routing: finding out what the caller needs and transferring them to the right person.

Keep for people the calls that need judgement, empathy or authority: complaints, refunds, complex quotes, anything medical, legal or financial, and upset callers. The receptionist's job on these calls is to recognise them early and hand them over.

2. Start with the role and the goal

Tell the agent who it is, who it works for and what a successful call looks like:

You are the friendly front-desk receptionist of Green Street Dental, a family dental practice in Springfield. Your goal is to answer questions about the practice and book, move or cancel appointments. If you can't help, take a message for the team.

One clear goal per agent works better than a long list. If you need very different jobs, such as support and sales, use separate agents.

3. Write an honest first message

The first few seconds set the tone. A good first message:

1. Says the business name, so callers know they reached the right place.
2. Says it is an AI assistant. Callers should never be surprised later, and in many places it is required: in the EU, the AI Act requires people to be told they are interacting with an AI at the latest at the first interaction.
3. Mentions recording, if you record. Several US states and many countries require the consent of everyone on the call.
4. Invites the caller to speak.

For example: "Hi, you've reached Green Street Dental. I'm the practice's AI assistant, and this call is recorded. I can answer questions or book an appointment. How can I help?"

Keep it short: every second before the caller can talk feels long on the phone. And add to the prompt: "If anyone asks whether you are a person or a robot, say honestly that you are an AI assistant."

4. Define the tone

Describe how the agent should sound in a few words: "friendly, calm and efficient; warm but not chatty". Add the language it speaks and what to do when a caller uses another one. Choose a voice that fits the business, and listen to it on a real phone line, not only in the browser.

5. Keep turns short

This is the most important rule for voice. Put it in the prompt in plain words:

• One or two sentences per turn, then let the caller speak.
• One question at a time. "What's your name?" then "And the best number to reach you?", never both at once.
• No lists longer than three items. Offer two or three options, or ask what the caller prefers.
• No formatting. Bullet points, links and symbols mean nothing when spoken. Write numbers, times and dates the way they should be said.

6. Confirm the details that matter

Speech recognition makes mistakes, especially with names, numbers and email addresses. Tell the agent to:

• repeat phone numbers back in small groups of digits;
• spell unusual names and email addresses back;
• confirm the date, time and service before booking: "So that's a cleaning on Tuesday the 14th at 10 in the morning. Shall I book it?"

A ten-second confirmation is much cheaper than a missed appointment.

7. Give it facts, not marketing

The agent can only be as accurate as the information you give it. Build a knowledge base from your service, pricing and FAQ pages, a plain document with the facts staff repeat every day, and your policies in simple sentences.

• Write facts. "We are open Monday to Friday, 8:00 to 18:00, and Saturday 9:00 to 13:00" beats "We're here when you need us".
• Include what you don't do. If you don't treat children or don't open on Sundays, say so, or the agent may guess.
• Keep it current. When prices or hours change, update the knowledge base the same day.
• Say what to do when the answer isn't there: offer to take a message, never invent an answer.

8. Describe the tools and when to use them

Tools let the receptionist act: check availability, book an appointment, add a row to a spreadsheet, transfer the call, end the call. For each tool, tell the agent:

• When to use it: "Check availability before offering times. Never invent times."
• What to collect first: "Before booking, you need the caller's full name, phone number and the service."
• What to say while it works: "Say 'One moment while I check the calendar' first." Silence on a call feels broken.
• What to do if it fails: "If the booking fails, apologise and offer to take a message."

9. Plan the hand-off and the fallbacks

A transfer to a person is not a failure; it is part of the design. Write down what to do when:

• the caller asks for a person: transfer during opening hours, take a message otherwise, and say what is happening ("I'll put you through to our front desk now") instead of going silent;
• the agent doesn't know the answer: say so and offer a message or a transfer;
• the caller is upset: acknowledge it briefly and offer a person;
• the caller asks not to be called again: confirm it and end the call politely;
• the topic is off-limits: "Don't give medical advice. For emergencies, tell the caller to hang up and call the emergency number.";
• the line is silent or noisy: ask once if they are still there, then end the call politely.

10. Decide what to collect after each call

Post-call analysis turns a conversation into data. Pick a few fields, for example:

• reason for the call (booking, question, complaint, other);
• appointment booked (yes or no) and its date;
• callback requested, with the best time;
• caller asked not to be called again (yes or no).

Give each field a clear name and a short description. Fixed choices are easier to count than free text.

Test before callers do

1. Write ten to twenty scenarios: the common reasons people call, a booking from start to finish with a change of mind, a question the knowledge base doesn't answer, a request for a person, "are you a robot?", an opt-out, a caller who interrupts or corrects a wrong number.
2. Make test calls, first in the browser, then on a real phone. Ask someone who didn't write the prompt to call too.
3. Read the transcripts. Did the agent give the business name and the AI disclosure? Were turns short? Did it confirm details, use the tools at the right moment and hand off when it should? Are the post-call fields right?
4. Change one thing at a time and repeat the scenarios that failed.

Then go live gradually: after hours first, then for overflow when the team is busy, then for all calls once you trust it.

Keep improving

An AI receptionist is never finished. Every week, read the short calls, the transferred calls and the calls with negative sentiment, pick the most common problem and fix it: add a missing fact, clarify an instruction, add a tool. After a month or two, most routine calls run smoothly, and your team spends its time on the conversations that really need a person.