PromptsScriptsConfiguration

AI Receptionist Prompting: Prompt Templates and Scripts (2026)

JB
Justas ButkusFounder, AInora
··Updated ·18 min read
Megan·English

Read a sample conversation from a test session

Elmwood Dental Care (private dental practice, Cambridge) is an invented demo business.

Megan (AI): Hello, this is Megan, Elmwood Dental Care's AI assistant. How can I help you today?

Caller: Hi, I'd like to book a check-up.

Megan (AI): I can offer Monday at eight o'clock in the morning, Monday at eight fifteen, or Monday at eight thirty for a check-up, shall I book it?

Megan (AI): I can offer Wednesday at six o'clock in the evening, Wednesday at six fifteen, or Wednesday at six thirty for your check-up, shall I book it?

Caller: The later one, please.

Megan (AI): May I have your full name, please?

Caller: Sophie Turner.

Megan (AI): You're booked for a check-up on Wednesday at six thirty in the evening. May I take your phone number or email address in case we need to contact you?

The most customizable AI virtual receptionist scripts are written as one prompt rather than a fixed call flow: you set the greeting, what the receptionist knows about your business, how it answers FAQs, when it routes a call or hands it to a person, its tone, and industry details such as appointment types or room policies, and the AI composes each reply within those limits. That prompt decides how callers are greeted, how their questions are handled, when a call is handed over and how your brand comes across: a well-written one turns a generic AI into a receptionist that represents you, a weak one frustrates callers. AInora sets up and tunes this prompt for each client as part of its managed service.

This guide covers everything you need to write effective AI receptionist prompts and scripts: greeting structures, FAQ response templates, call routing rules, escalation triggers, tone configuration, industry-specific examples, and the common mistakes that make AI receptionists sound robotic or unreliable. Whether you are setting up your first AI receptionist or optimizing an existing one, these patterns apply across every major AI voice platform.

Prompt-led
Most AI quality depends on prompt
5-7s
Ideal greeting length
Iterate
Practitioners review weekly
6+
Industry templates included

What Is AI Receptionist Prompting?

AI receptionist prompting is the practice of writing the instruction text that tells a voice AI agent who it is, what it knows about your business, how it should speak, and when it must hand the call to a person. It is configuration, not conversation design in the old sense: you are not scripting every line, you are setting the boundaries inside which the model composes its own.

That distinction is the whole discipline. A phone tree is a decision graph and every path is authored. A prompt is a brief, and the model improvises inside it. So the failure modes are different too: a phone tree breaks when the caller wants something the graph does not contain, while a prompt breaks when it is silent on something the caller asks about, and the model fills the silence itself. Most bad AI receptionist calls are not model failures. They are gaps in the prompt.

What Should an AI Receptionist Prompt Contain? (Template)

Every workable AI receptionist prompt carries the same six blocks. Write them in this order, in plain sentences, and the rest of this guide is detail on each one.

BlockWhat it setsWhat breaks without it
1. IdentityWho the agent is, which business it answers for, and whether it discloses that it is AIThe caller does not know who picked up, and disclosure rules are handled by accident
2. KnowledgeHours, address, services, prices you are willing to quote, policies, staff names, the 30 to 50 questions you actually getThe model invents a plausible answer, which is the single most damaging failure on a live call
3. BehaviourWhat the agent is allowed to do: book, reschedule, take a message, quote, transferThe agent either overreaches into commitments you cannot honour, or refuses tasks it should handle
4. EscalationThe exact conditions that trigger a transfer, the destination for each, and the fallback when nobody answersEmergencies queue behind routine calls, and frustrated callers are argued with instead of transferred
5. PersonalityThree or four adjectives, each with a concrete behavioural instruction rather than an adjective aloneTone drifts between calls, and cheerfulness lands on a caller with a serious problem
6. FallbackThe precise wording for I do not know, plus the instruction never to fabricateSilence, or a confident wrong answer that the caller acts on

A copy-ready skeleton, in the order above:

Prompt skeleton

Identity. You are the receptionist for [business], a [type of business] in [city]. If a caller asks whether they are speaking to a person, say plainly that you are an AI assistant for [business].

Knowledge. Hours: [per day]. Address and parking: [detail]. Services: [list, with what each involves]. Policies: [cancellation, payment, insurance]. Common questions and their answers: [30 to 50 pairs, written as spoken sentences].

Behaviour. You may book, reschedule and cancel appointments within [rules]. You may quote [what]. You may not [commit to discounts, give clinical or legal advice, promise a callback time you cannot verify].

Escalation. Transfer to [destination] when [condition]. Transfer immediately if the caller asks for a person. If nobody answers within [n] rings, take a message with name, number and reason, and confirm it back.

Personality. Warm: use the caller's name once they give it. Concise: answer the question asked, then stop. Patient: never rush, never talk over. Professional: complete sentences, no filler, no exclamation marks.

Fallback. If you do not know something, say "I do not have that in front of me, but I can have someone call you back with it." Never guess a price, a date, a name or a policy.

Two rules govern the whole template. Write knowledge as sentences the agent could say aloud, not as data rows, because the model will read them back close to how you wrote them. And write every rule in the positive: "transfer immediately when a caller asks for a person" produces better behaviour than "do not argue with callers who ask for a person", because the first states an action and the second states an absence.

Why Prompts and Scripts Matter for AI Receptionists

Think of the prompt as your AI receptionist's training manual. A human receptionist gets weeks of on-the-job training - learning your business, your customers, your tone, your policies, and your preferences. An AI receptionist gets all of that through the prompt. The more precise and complete your prompt, the better your AI performs.

The prompt controls three critical dimensions of your AI receptionist's behavior:

  • Knowledge: What the AI knows about your business - services, hours, policies, FAQs, team members, and procedures.
  • Behavior: How the AI acts - when to book, when to transfer, when to take a message, and when to say "I don't have that information."
  • Personality: How the AI sounds - warm and casual, professional and formal, empathetic and patient, or energetic and enthusiastic.

Most businesses underinvest in prompt writing. They spend weeks evaluating AI providers and minutes writing the prompt. This is backwards. The provider gives you the engine - the prompt gives you the driver. Two businesses using the exact same AI platform can have wildly different caller experiences based solely on prompt quality.

What Makes a Good AI Receptionist Greeting? (Best Practices)

The greeting is the first thing every caller hears, and it sets the tone for the entire interaction. A good greeting accomplishes three things in under 7 seconds: it identifies the business, it signals that this is a capable agent (not a basic phone tree), and it invites the caller to state their need.

Greeting Structure

The most effective greeting follows this formula: warm opener + business identification + capability signal + open question.

  • Warm opener: "Good morning" or "Thank you for calling" - brief and natural.
  • Business identification: The name of your business - always include this so the caller knows they reached the right place.
  • Capability signal: A brief phrase that tells the caller what you can help with - "I can help you schedule an appointment, answer questions, or connect you with our team."
  • Open question: "How can I help you today?" - invites the caller to state their need in their own words.

Greeting Examples by Industry

Dental practice: "Good morning, thank you for calling Smile Dental. I can help you with appointments, insurance questions, or anything else you need. How can I help you today?"

Law firm: "Thank you for calling Parker and Associates. I can schedule a consultation, provide office information, or connect you with an attorney. What can I assist you with?"

Home services: "Hi, you've reached Reliable Plumbing. I can schedule a service call, provide an estimate timeline, or help with an emergency. What do you need today?"

Hotel: "Good afternoon, thank you for calling the Grand Hotel. I can help with reservations, guest services, or local recommendations. How may I assist you?"

Keep it under 7 seconds

Long greetings frustrate callers. They called because they have a need, and every second of greeting delays them from expressing it. Aim for 5-7 seconds of total greeting time. Cut any word that does not serve identification, capability signaling, or invitation. Never include your full address, website URL, or promotional messages in the greeting.

Greeting Best-Practice Checklist

Nine rules cover almost every greeting problem that shows up in transcripts. The first five are about what the greeting contains; the last four are about what it must not.

  • Name the business in the first four words. The caller is confirming they dialled correctly before they listen to anything else.
  • Signal capability, not menu. "I can help with appointments, questions, or reaching the team" tells the caller they can speak normally. A numbered list tells them they cannot.
  • End on an open question. "How can I help you today?" hands the turn back. A statement leaves the caller unsure whether to speak.
  • Match the greeting to the clock. Good morning, good afternoon, good evening, and an explicit after-hours line that still offers to help.
  • Warm and concise are one setting, not two. Warmth is carried by word choice, not by length: "I would be happy to help with that" is warm; a second sentence explaining how happy you are is not.
  • No address, website or promotion. Every one of these delays the caller from saying why they rang.
  • No exclamation marks in the prompt. They are read as vocal emphasis and land as forced enthusiasm.
  • No apology for existing. "Sorry to keep you" at the start invents a problem the caller did not have.
  • No question the caller cannot answer yet. Asking for an account number before they have said what they want restarts the call badly.

On disclosure: if your greeting states that the caller is speaking to an AI assistant, put it after the business name and before the open question, in one short clause. Burying it at the end of a long greeting is worse than not saying it, because the caller has already started talking.

A greeting clause is not the whole legal answer. In the EU, Article 50(1) of the AI Act requires providers to design AI systems that talk directly to people so that those people are informed they are interacting with an AI system, unless that is obvious to a reasonably well-informed, observant and circumspect person, and Article 50(5) requires that information to be clear and distinguishable and given at the latest at the first interaction. On a call, that points to the greeting. But the duty is about how the system is designed, so one line does not settle it on its own. It applies from 2 August 2026. Rules outside the EU differ: see AI caller disclosure laws by country.

Time-Aware Greetings

Configure your AI to adjust greetings based on the time of day and day of week. "Good morning" before noon, "Good afternoon" until 5 PM, "Good evening" after 5 PM. On weekends or holidays, add context: "Thank you for calling after hours. I can still help you with appointments and questions." This small detail makes the AI feel more human and less scripted.

FAQ Response Templates: Consistent, Accurate Answers

After the greeting, the most common interaction is answering frequently asked questions: hours, location, services, insurance, preparation instructions, and policies. The prompt should include specific, accurate answers for every question your business gets asked regularly.

How to Structure FAQ Responses

For each FAQ, provide: the answer itself, the ideal conversational phrasing, and what to say if the caller needs more detail. Do not write FAQ responses as bullet points or data dumps - write them as natural sentences the AI would say in conversation.

Bad example: "Hours: Mon-Fri 8-5, Sat 9-1, closed Sunday."

Good example: "We are open Monday through Friday from 8 AM to 5 PM and Saturday mornings from 9 AM to 1 PM. We are closed on Sundays. Would you like to schedule an appointment?"

Notice the good example delivers the information conversationally and immediately offers to take the next step (booking). This is the difference between an AI that answers questions and an AI that converts callers into customers.

Essential FAQ Categories

  • Business hours and location: Hours for each day, address, parking instructions, accessibility information.
  • Services offered: What you do, what you do not do, and what to expect from each service.
  • Insurance and payment: Accepted insurance plans, payment methods, financing options, and billing procedures.
  • Preparation instructions: What to bring, how to prepare, what to wear, where to park, and how early to arrive.
  • Cancellation and rescheduling: Your cancellation policy, rescheduling process, and any associated fees.
  • Emergency handling: What constitutes an emergency, who to contact, and what the caller should do.
  • New patient/client process: Required paperwork, expected duration of first visit, and what to expect.

Handling Unknown Questions

Your prompt must include instructions for what the AI says when it does not know the answer. The worst response is making something up. The best response is honest and helpful: "I don't have that specific information, but I can have someone from our team call you back with the answer. Would you like to leave your number?" Configure the AI to never fabricate information - this is a non-negotiable rule in every prompt.

Call Routing Rules and Escalation Triggers

Not every call should be fully handled by the AI. Some need to reach a human - but which ones? Call routing rules define when the AI transfers, who it transfers to, and how it handles the handoff. Getting this right is the difference between an AI receptionist that staff trust and one that creates problems.

Transfer Triggers

Define specific situations where the AI should transfer the call. Common triggers include:

  • Caller requests a specific person: "Can I speak with Dr. Johnson?" - transfer immediately without interrogation.
  • Emergency or urgent situations: Medical emergencies, legal urgency, burst pipes - transfer to the on-call person with priority.
  • Complex or sensitive topics: Billing disputes, complaints, malpractice concerns - transfer to a manager or designated handler.
  • Repeated caller frustration: If the caller says "Let me talk to a real person" or expresses frustration, transfer without argument.
  • Topics outside the AI's scope: When the caller's need is clearly outside what the AI is configured to handle.

Transfer Destinations

Configure different transfer destinations for different situations. A dental practice might route emergencies to the on-call dentist, billing questions to the front office manager, and new patient inquiries to the patient coordinator. Configure fallback destinations: if the primary person does not answer, who is next? If nobody is available, what happens - voicemail, callback scheduling, or message-taking?

Warm vs Cold Transfers

A warm transfer means the AI introduces the caller and provides context to the human: "I have Mrs. Rodriguez on the line. She is calling about her appointment on Thursday and has a question about her insurance coverage." A cold transfer simply connects the call. Warm transfers provide a much better experience for both the caller and the staff member. Configure your AI to use warm transfers whenever the receiving party is available.

The 3-attempt rule

If a caller asks to speak with a human three times, the AI should transfer unconditionally, even if the request does not match any configured transfer trigger. Arguing with a caller who wants a human never ends well. This rule should be in every prompt.

Tone and Personality Configuration

Tone is the emotional quality of how your AI communicates. Personality is the consistent character it projects across all interactions. Together, they determine whether callers perceive your AI receptionist as helpful and professional or cold and robotic.

Defining Your AI's Tone

Choose 3-4 adjectives that describe how you want your AI to sound, then provide specific behavioral instructions for each:

  • Warm: Use the caller's name after they provide it. Say "I would be happy to help with that" instead of "I can do that." Express genuine interest in their needs.
  • Professional: Use complete sentences. Avoid slang and filler words. Provide accurate, specific information. Never speculate or guess.
  • Patient: Never rush the caller. If they need time to think, wait quietly. If they repeat themselves, respond as if hearing it for the first time. If they are confused, rephrase without condescension.
  • Concise: Get to the point. Do not over-explain. Answer the question asked, then pause for the caller's next question. Do not monologue.

Adapting to Caller Mood

A good AI receptionist adjusts its tone based on the caller's emotional state. If the caller sounds rushed, the AI should be brief and efficient. If the caller sounds anxious (medical concern, legal issue, emergency), the AI should be reassuring and calm. If the caller sounds confused, the AI should slow down and simplify. Include these adaptation rules in your prompt.

What to Avoid in Tone Configuration

Do not make the AI overly cheerful - excessive enthusiasm sounds fake and annoys callers with serious concerns. Do not make it apologize excessively - "I'm so sorry about that" after every sentence feels performative. Do not use exclamation marks in prompt examples - they translate to unnatural vocal emphasis. Aim for competent and helpful, not bubbly and eager.

Industry-Specific Prompt Examples

Different industries have different caller expectations, vocabulary, and workflow requirements. Here are prompt configuration patterns for the industries where AI receptionists are most common.

Dental Practice

Key prompt elements: appointment types (cleaning, exam, emergency, consultation), insurance verification questions, new patient paperwork mention, preparation instructions for procedures, and emergency triage protocol. The AI should ask whether the caller is a current patient, which dentist they see, and what type of appointment they need. For emergencies, it should ask about the nature of the issue and the severity of pain to route appropriately.

Law Firm

Key prompt elements: practice areas (family, criminal, personal injury, business), conflict check questions, consultation booking with specific attorney matching, and confidentiality statements. The AI must never provide legal advice - it books consultations and collects case details only. Include a clear instruction: "Do not interpret or advise on legal matters under any circumstances."

Medical Practice

Key prompt elements: appointment types by provider, insurance verification, new patient intake, prescription refill routing, lab result inquiry routing, and urgent/emergency triage. The AI must distinguish between urgent situations (route to nurse) and non-urgent scheduling requests. HIPAA compliance instructions should be in the prompt: never discuss patient records, always verify identity before providing appointment details.

Hospitality (Hotels)

Key prompt elements: room types and amenities, check-in/check-out times, reservation modification, local attractions and dining recommendations, special requests handling, and group booking inquiry routing. The tone should be warm and accommodating. The AI should proactively offer helpful information: "I should mention that we offer complimentary breakfast from 7 to 10 AM."

Home Services (HVAC, Plumbing, Electrical)

Key prompt elements: service area definition, emergency vs routine service classification, basic diagnostic questions ("Is water actively leaking?"), scheduling with estimated arrival windows, and pricing expectation setting ("A diagnostic visit is $89. The technician will provide a full quote on site."). For emergencies, the AI should prioritize speed and collect only the essential details: address, nature of the problem, and availability for same-day service.

Salon and Spa

Key prompt elements: services with durations (haircut 30 min, color 90 min, facial 60 min), stylist/therapist preferences, pricing ranges by service, cancellation policy (often 24-48 hours), and product questions. The AI should ask about service history if relevant: "Have you had a color treatment with us before?" to route to the appropriate specialist.

IndustryCritical Prompt ElementBiggest Risk If MissingTone Priority
DentalEmergency triage protocolTrue emergency treated as routine callCalm, reassuring
LegalNo legal advice instructionLiability from AI interpretationProfessional, precise
MedicalHIPAA compliance rulesProtected health information exposureEmpathetic, careful
HospitalityLocal knowledge and recommendationsMissed upsell and poor experienceWarm, accommodating
Home servicesEmergency classification questionsDelayed response to water/gas leakEfficient, practical
Salon/SpaService duration accuracyScheduling errors and overlapFriendly, relaxed

Common Prompt Writing Mistakes to Avoid

These are the mistakes that cause the most problems in practice, and every one of them is visible in the first week of call transcripts. Avoiding them will save you weeks of troubleshooting.

1. Being Too Vague

"Be helpful and answer questions about our business" is not a useful prompt. The AI does not know what your business does, what hours you are open, or what services you offer. Be specific: list your services, your hours, your policies, and the exact answers to every common question. Vague prompts produce vague and often inaccurate responses.

2. Being Too Restrictive

"Only say these exact phrases. Do not deviate from the script." This turns your AI into a basic phone tree. The whole point of an AI receptionist is that it handles natural conversation, including unexpected questions and unusual requests. Give it knowledge and guidelines, not a rigid script. Let it use its language model to construct natural responses.

3. Missing the Fallback

If your prompt does not include instructions for what to do when the AI cannot help, it will either make something up or go silent. Always include a clear fallback: "If you do not know the answer or cannot complete the request, offer to take the caller's information and have a team member follow up."

4. Ignoring Edge Cases

What if the caller speaks a different language? What if they are angry? What if they ask about a competitor? What if a child calls? What if someone is testing the AI? Your prompt should address these scenarios explicitly. Edge cases that are unhandled in the prompt become embarrassing moments in real calls.

5. Overloading the Greeting

Including your full address, website, list of services, and a promotional message in the greeting guarantees that callers start the conversation annoyed. Keep the greeting under 7 seconds. Save detailed information for when the caller actually asks for it.

6. No Update Process

Your business changes - new services, new hours, new staff, new policies. If nobody updates the prompt, the AI gives outdated information. Assign an owner for the prompt (typically the office manager) and schedule monthly reviews. Every time something changes in the business, update the prompt the same day.

The Prompt Optimization Process

Writing the initial prompt is step one. Optimization is the ongoing process that turns a decent AI receptionist into an excellent one. Here is the systematic approach.

1

Review call transcripts weekly

Read at least 10-20 call transcripts per week during the first month. Look for three things: calls where the AI gave incorrect information (knowledge gap), calls where the AI handled the conversation awkwardly (behavior gap), and calls where the AI had the right information but delivered it poorly (tone gap).

2

Categorize issues by type

Group transcript problems into categories: missing FAQ answers, incorrect information, inappropriate escalation (or failure to escalate), tone mismatch, and conversation flow issues. Prioritize fixes by frequency - the issue affecting 20% of calls matters more than the one affecting 1%.

3

Update the prompt with specific fixes

For each issue, add or modify the relevant section of the prompt. If callers keep asking about parking and the AI does not know, add parking information. If the AI is escalating too aggressively, adjust the escalation criteria. Make one change at a time so you can measure the impact.

4

Test the change before deploying

Call your own AI and test the specific scenario you just fixed. Verify the fix works and does not break other scenarios. Regression testing matters - a change to one part of the prompt can affect behavior elsewhere.

5

Track metrics over time

Measure the metrics that matter: call resolution rate (calls fully handled without human intervention), caller satisfaction (if you collect feedback), average call duration, and transfer rate. Improvement in these metrics confirms your prompt optimization is working.

The 90/10 rule

A good prompt handles 90% of calls well on day one. The optimization process is about the other 10% - the edge cases, unusual questions, and specific scenarios that only surface with real call volume. Expect to make significant updates in weeks 1-4 and decreasing updates after that as you cover more ground.

Which Prompting Rules Are Measured, and Which Are Convention?

Prompting guides, this one included, hand out numbers: keep the greeting under seven seconds, write 500 to 2,000 words, review weekly. It is worth being explicit about which of those rest on published evidence and which are practitioner convention, because the difference changes how hard you should hold to them.

RuleBasisConfidenceHow to treat it
Greeting of roughly 5 to 7 secondsPractitioner convention. We are not aware of a published study establishing an optimal AI receptionist greeting length, and this page does not claim one existsLow as a number, high as a directionTreat the direction as sound and the number as a starting point. Measure your own hang-ups during the greeting
Prompt length of 500 to 2,000 wordsPractitioner convention, and platform-dependentLowIgnore the word count. Judge completeness by whether every question in a week of transcripts has an answer in the prompt
Transfer unconditionally on the third request for a humanConvention, chosen deliberately as a conservative defaultLow as a numberThe threshold is arbitrary; the principle is not. Transferring on the first request is defensible
Callers dislike waiting more than they dislike automationMeasured. A Software Advice consumer survey (n=1,100, 2014) found 63% preferred a callback to waiting on holdMedium. Real sample, but the survey is twelve years oldDirectional. Removing the wait matters more than the channel it is removed from
A minority of badly queued calls distort any averageMeasured. ContactBabel 2024 (n=225 UK contact centres) reports average speed to answer at 116 seconds mean against a 48 second medianHigh. Sample and method disclosedRead your own call metrics as medians before you tune a prompt against an average

The honest summary: almost nothing about prompt wording has been studied in a way that would survive a citation check. What is measurable is the output. Your own transcripts, read weekly, are better evidence for your business than any published prompt-length guidance, and they are the only evidence that reflects your callers. For the industry numbers a prompt is ultimately judged against, see the call abandonment rate benchmark.

What Does a Complete AI Receptionist Prompt Look Like?

A complete AI receptionist prompt is the six-block skeleton above with every block filled in for one business and written as sentences the agent can say aloud. Below are two filled-in prompts for UK businesses, a dental practice in Leeds that sees NHS and private patients and a plumbing and heating firm in Birmingham, followed by a booking script for three objections. Every business fact sits in [square brackets]: replace each one with your own, and delete any line you cannot fill rather than leave the agent to guess.

Complete prompt: dental practice in Leeds (NHS and private)

Identity. You are Jessica, the AI assistant for [practice name], a dental practice in Leeds that sees NHS and private patients. Greet callers with: "[Good morning], [practice name], this is Jessica, the practice's AI assistant. How can I help?" If a caller asks whether they are speaking to a person, say plainly that you are an AI assistant and offer to pass a message to the team.

Knowledge. Opening hours: [per day]. Address: [street], Leeds [postcode]. Parking: [detail]. NHS treatment offered: [which treatments, and whether the practice is taking new NHS patients]. Private treatment offered: [which treatments]. Prices you may quote: [only the NHS and private charges the practice has written down for you]. Dentists and hygienists: [names and the days they work]. Common questions and their answers: [30 to 50 pairs, written as spoken sentences].

Behaviour. You may book, move and cancel [check-ups, hygiene visits and consultations] in [booking system]. Ask whether the caller is an existing patient and whether they want NHS or private care. Ask for the reason for the visit in the caller's own words, and for no more health detail than the booking needs. You may not give clinical advice, diagnose, suggest medicine or say whether a problem can wait.

Escalation. First, before any booking, listen for the four situations the NHS sends to A&E or 999. If the caller describes any of them, read them this NHS instruction straight away: "Go to A&E or call 999 if you have: serious injuries to your face or jaw; heavy bleeding from your mouth that will not stop; severe swelling of your mouth, lips, throat or neck, and you're having difficulty breathing or opening 1 or both eyes; any injury to your head or face that's caused you to lose consciousness, vomit, or have double vision." Pass it on whenever a caller describes anything like these; never judge for yourself whether it is serious enough. Next, check the problem against this list of urgent problems: [the practice's own list]. If it is on the list, offer the earliest urgent slot in [booking system]. If there is none, or the practice is closed, tell them: "You can get advice from NHS 111 or contact a dentist and ask for an appointment." If you cannot tell whether a problem is on the list, do not decide: take name, number and reason, and offer a callback from [the dentist on duty], as you do for any other clinical question. If the caller asks for a person, transfer to [reception line]; if nobody answers within [n] rings, take a message and read it back.

Personality. Calm: short, plain sentences, no exclamation marks, no clinical words the caller did not use. Patient: let the caller finish before you answer. Direct: if a caller is in pain, skip small talk and go straight to the Escalation steps.

Fallback. If you do not know something, say "I do not have that in front of me, but I can ask the team to call you back with it." Never guess a price, a waiting time, a dentist's availability or whether a treatment is suitable.

Complete prompt: plumbing and heating firm in Birmingham

Identity. You are Jessica, the AI assistant for [firm name], a plumbing and heating firm in Birmingham. Greet callers with: "[Good morning], [firm name], this is Jessica, the firm's AI assistant. How can I help?" If a caller asks whether they are speaking to a person, say plainly that you are an AI assistant.

Knowledge. Service area: [towns or postcodes], up to [n] miles from [base]. If a job is outside it, say so and [take a message, or say where else to look]. Services: [list, with what each involves]. Office hours: [per day]. Emergency cover: [hours, or none]. Prices you may quote, in pounds: [call-out fee, hourly rate or fixed prices the firm has written down]. Common questions and their answers: [30 to 50 pairs, written as spoken sentences].

Behaviour. First decide whether the call is an emergency or a routine job, using this list of emergencies: [the firm's own list]. For a routine job, ask for the address and postcode, the problem in the caller's words and [the firm's job-scope questions], then book it into [calendar] with an arrival window of [length]. You may not promise an arrival time the calendar does not show, diagnose a fault or say that a repair is safe to leave.

Escalation. If the caller says they can smell gas, stop the booking. Tell them to call the National Gas Emergency number on 0800 111 999 now, and pass on the gas network's advice: open windows and doors; turn off the gas supply at the meter as well as any gas appliances; do not smoke or use any naked flames; do not touch any electrical switches; do not enter a cellar, even if the gas meter is located there. The emergency line comes first, not [firm name]'s engineers; once the caller has rung it, offer [a follow-up visit or a message]. For a water leak, give [the firm's own first steps, if it wants any given] and transfer to [on-call engineer]. For any other emergency on the list, transfer to [on-call engineer]. If nobody answers within [n] rings, take name, number, address and problem, read them back, and tell the caller [when the engineer will call back].

Personality. Practical: in an emergency, ask only what the engineer needs to get there. Calm: plain words, no exclamation marks, no alarm in your voice. Concise: answer the question asked, then stop.

Fallback. If you do not know something, say "I do not have that in front of me. I can take your details and ask [the office] to call you back." Never guess a price, an arrival time or whether a fault is dangerous.

The urgent-care wording in the dental prompt is the NHS's own, from its page on how to find an emergency or urgent NHS dentist appointment (guidance for England, last reviewed 14 May 2025). The same page says urgent treatment should be offered "within 24 hours or 7 days, depending on your symptoms"; which of the two applies is a clinical judgement, so the agent never tells a caller which one they fall into. The gas instructions come from SGN's What to do if you smell gas, and Cadent, which manages the 0800 111 999 line on behalf of the gas industry, describes it as free and available 24/7, 365 days a year.

Both prompts are written for the calls the team cannot get to: the agent answers, books and takes messages, and anything that needs a person goes to one with a summary. Transfers reach people when they are free; when they are not, the agent takes a message and reads it back. Both prompts also have the agent say it is an AI assistant in the greeting. The rules on that differ by place: for the EU, see what Article 50(1) of the AI Act asks in the greeting section above; for other countries, the guide to AI caller disclosure laws by country.

Booking Script: Three Objections, Written as Instructions

The booking part of a prompt works best as rules the agent follows rather than lines it reads out, so it can answer in its own words and still stay inside your limits. Three objections are worth writing rules for on any trade line:

Booking script with three objections

"How much will it cost?" Quote only what is in [call-out fee / price list]. If the job is not on it, say that the engineer confirms the price on site, then offer to book the visit. Never work out a figure yourself.

"Can someone come today?" Check [calendar] and offer the earliest slot it actually shows. If today is full, say so and offer that earliest slot. Never promise a time the calendar does not show. If the call is an emergency, follow the Escalation block instead.

"I'm ringing round." Take the caller's name, number and the job in one sentence, offer the earliest slot, and confirm the details back. Do not argue about price or comment on other firms.

Every booking. Before the call ends, read back the name, phone number, address, date and arrival window, and ask the caller to confirm them.

Is there evidence on which script books more jobs? We found no independent published comparison of AI voice scripts for home-service booking calls: a search on 29 October 2026 returned vendor case studies and vendor blogs with self-reported booking rates and no stated method, none of which we would cite as an industry figure. Judge a script by your own transcripts instead, as the measured-versus-convention section above sets out: count the calls that end in a confirmed booking, and read the ones that do not. For the trade detail behind these prompts, see emergency dispatch for HVAC and plumbing firms, the job-scope questions to ask before a quote and NHS and private triage for UK dental practices.

What Questions Should an AI Receptionist Ask Callers?

An AI receptionist should ask why the caller is ringing first (the greeting's open question does this), then their name, then a callback number, and after that only the questions that type of call needs: one at a time, with each answer read back. Taking the number early is practice rather than a finding: if the line drops, someone can ring the caller back.

Call typeAsk, one at a timeBefore the call ends
New bookingWhat is the appointment for? Have you been to us before? Which day suits you? Morning or afternoon? Then offer the earliest real slotsRead back name, number, date and time
Change to an existing appointmentWhich appointment is it? Confirm the caller is the person it belongs to before reading any detail out. Do you want to move it or cancel it? Which new day suits you?Confirm the new time, or the cancellation, back
EmergencyWhat is happening right now? If it matches a safety instruction (a gas smell, the NHS 999 list), give that first. Then: what is the full address and postcode? What number can we reach you on?Say who will call back or arrive, and [when]
Quote requestWhat is the job, in a sentence? Where is it, and is that inside the service area? How soon do you need it? Then the business's own job-scope questionsSay who will price the job and how; read back the number
Message for a named personWho is the message for? What is it about? Is it urgent? When is the best time to call you back?Read the message back word for word

Ask only what the business will use. UK GDPR Article 5(1)(c) requires personal data to be "adequate, relevant and limited to what is necessary in relation to the purposes for which they are processed" (Source: legislation.gov.uk, UK GDPR Article 5), and the ICO's guidance on data minimisation puts it plainly: "You should not have more personal data than you need to achieve your purpose." For a prompt, that means every question has to feed something a person at the business will act on. If nobody reads a field, delete the question that collects it. Never ask for card details on the call unless the business has a payment process built for it, and never ask for more health detail than the booking needs.

Check who you are talking to before reading anything out. Before the agent reads out anything personal, such as an appointment time or an account detail, it should confirm the caller is the person it belongs to. Article 5(1)(f) of the same regulation requires personal data to be "processed in a manner that ensures appropriate security of the personal data," including "protection against unauthorised or unlawful processing". On how hard to check, the closest ICO wording comes from its guidance on responding to access requests, not from a rule about phone calls: "You should be reasonable and proportionate about what you ask for," and "The level of checks you make may depend on the nature of the information and on the possible harm and distress". In a prompt, that becomes a sliding scale: [two details that match the record] before confirming an appointment time, and a stronger check, or a call back to the number on file, before anything more sensitive.

For the job-scope questions a trade asks before pricing work, see quote and estimate request intake; for the information a business has to hand over before the agent can ask any of this, see the AI receptionist onboarding guide.

How Is a Voice Receptionist Prompt Different From a Chat Prompt?

A voice receptionist prompt is written for the ear. A caller cannot scroll back, skim or click, so every reply has to be short, carry one question, and say numbers, dates and names in a form that survives being heard once. A chat prompt can lean on lists, links and formatting; a voice prompt has to rule them out.

One question per turn, and two or three options at most. Short-term memory is small: a review in Behavioral and Brain Sciences found a central capacity limit averaging about four chunks (Cowan, 2001). In a Nielsen Norman Group usability study of voice assistants (17 participants, 2018), participants "could not commit all the information to their working memory" when answers ran long and found it hard to remember all the spoken options, and the authors concluded that "Speech is an inefficient output modality". Neither source gives a number of options for a phone call; capping a spoken turn at two or three is our design choice from those findings.

Answer first, keep turns short, and stop when the caller starts talking. A study of conversation in 10 languages found "a general avoidance of overlapping talk and a minimization of silence between conversational turns" (Stivers and colleagues, PNAS, 2009). An agent that talks over a caller, or leaves a long pause before it answers, works against both halves of that finding. How short a turn should be is convention; tell the agent to give the answer in its first sentence and then stop.

Tell the agent how to say numbers. Speech engines have to be told what kind of text they are reading. The W3C's Speech Synthesis Markup Language standard has a say-as element for this, with "dates, times, numbers, acronyms, currency amounts and more" as examples, and its companion note on say-as values defines separate values for a telephone number and for text to be "spoken as a series of alpha-numeric characters". The standard does not tell you how any particular receptionist product reads numbers, so write the way you want them said into the prompt and test it on a call: phone numbers in short groups, then read back; dates as a person says them ("Wednesday the fourth of November", not "04/11"); prices in words ("forty-five pounds"). The grouping and the formats are convention.

Spell names and email addresses back. Have the agent spell them letter by letter and use the ICAO phonetic alphabet (Alfa, Bravo, Charlie and on through Zulu) for any letter that could be misheard: "Is that M for Mike or N for November?"

No lists, symbols, links or markdown in replies. A caller cannot see formatting, so the prompt should rule out bullet points, numbered lists, symbols, links and phrases such as "click here". Read back every critical detail (name, number, date, address) before the call ends. Both are practitioner convention rather than findings, and both cost nothing to follow.

The same split as the measured-versus-convention table above, for the voice rules:

Voice ruleBasisConfidence
One question per turn; two or three options at mostSource: Cowan, Behavioral and Brain Sciences (PubMed)Cowan (2001): short-term memory averages about four chunks. Nielsen Norman Group (2018, 17 participants): people could not hold long spoken answers or all the spoken options. The two-or-three cap is our design choice, not a number either source givesHigh for the memory limit; medium for the usability study; low for the exact cap
Answer first, short turns, stop when the caller talksSource: Stivers et al., PNAS (PubMed)Stivers et al. (2009, 10 languages): conversation avoids overlapping talk and minimises silence between turnsHigh for the finding; turn length is convention
State how numbers, dates, phone numbers and prices are spokenSource: W3C SSML say-as noteW3C SSML 1.1 and its say-as note: speech engines are told whether text is a date, a time, a telephone number or characters to spell outHigh that it matters; the grouping is convention
Spell names and emails back, with the phonetic alphabetSource: SKYbrary: ICAO Phonetic AlphabetICAO phonetic alphabet (SKYbrary reference table). Using it on reception calls is practiceMedium-high for the alphabet; the habit is convention
No lists, symbols, links or markdown in repliesPractitioner convention: a caller cannot see formattingConvention
Read back every critical detail before the call endsPractitioner conventionConvention

These rules are easier to judge by ear than on the page. The guide to testing an AI voice agent before launch covers testing them as real phone calls, the listening checklist for before you buy covers what to listen for, and whether an AI caller sounds robotic covers sounding natural on a call.

Can an AI Write Your AI Receptionist Prompt for You?

Yes, as a first draft. A general AI chat assistant can interview you through the six blocks and turn your answers into sentences an agent can say aloud. What it cannot do is know your business, so every fact in the draft has to be checked by you before the prompt goes anywhere near a caller.

Copy-ready prompt for drafting your receptionist prompt

Interview me one question at a time to write the instructions for an AI phone receptionist for my business. Work through these six blocks in order: Identity, Knowledge, Behaviour, Escalation, Personality, Fallback. Wait for my answer before you ask the next question. Write the result as sentences the receptionist could say aloud, with no lists, symbols or exclamation marks. Never invent a fact about my business: no hours, prices, names, services or policies I have not given you. Mark every fact I did not give you as [CHECK], and list all the [CHECK] items at the end.

Then check every fact yourself. NIST's Generative AI Profile (NIST AI 600-1, July 2024) defines confabulation as "the production of confidently stated but erroneous or false content" by which users may be misled, and calls it "a natural result of the way generative models are designed". A drafted prompt that invents your Saturday hours or a call-out fee is the exact failure the Fallback block exists to prevent, written into the instructions themselves. Read every line against what you know, and treat any [CHECK] you cannot fill as a line to delete.

Machine-written instructions can be good. In a 2022 research preprint, Large Language Models Are Human-Level Prompt Engineers, instructions generated automatically by a language model achieved "better or comparable performance to the instructions generated by human annotators on 19/24 tasks". Those were text benchmark tasks, not phone calls, so the result shows that a machine draft can be a sound starting point, not that a drafted receptionist prompt will work on your line. Test it as a phone call before launch, as the pre-launch testing guide sets out.

Keep real callers out of the drafting chat. Do not paste callers' names, numbers or real call recordings into a general chat tool to draft or test the prompt; invented examples do the same job. Drafting instructions does not need anyone's personal data, and the UK GDPR data minimisation principle limits personal data to what is "adequate, relevant and limited to what is necessary" for the purpose. The GDPR guide for AI receptionists in Europe covers the wider picture.

For the full sequence around the prompt (call flows, FAQ base, escalation, scheduling and testing), see the AI receptionist setup guide; for a plain explanation of how the model behind an agent learns to hold a conversation, see how voice AI is trained for conversations. If you would rather start one step earlier, our free prompt interviews you about your company and maps the calls, messages and follow-ups your team repeats, before anything is written.

Frequently Asked Questions

AI receptionist prompting is the practice of writing the instruction text that tells a voice AI agent who it is, what it knows about your business, how it should speak, and when it must hand the call to a person. It is configuration rather than line-by-line scripting: you set the boundaries and the model composes inside them, which is why gaps in the prompt, not model errors, cause most bad calls.

Six blocks, in this order: identity (who the agent is and whether it discloses being AI), knowledge (hours, services, policies and the 30 to 50 questions you actually get, written as spoken sentences), behaviour (what it may book, quote or promise), escalation (the exact transfer triggers, destinations and fallbacks), personality (three or four adjectives, each with a concrete behavioural instruction), and fallback (the exact wording for "I do not know" plus a rule never to fabricate).

Name the business in the first four words, signal capability instead of reading a menu, end on an open question, and adapt to the time of day. Keep it short: no address, no website, no promotion, no exclamation marks and no question the caller cannot answer yet. If you disclose that the caller is speaking to an AI assistant, place that clause after the business name and before the open question. The commonly quoted 5 to 7 second target is practitioner convention rather than a published finding, so treat the direction as sound and measure your own hang-up rate during the greeting.

An AI receptionist prompt is the configuration text that defines how your AI phone agent behaves - its personality, knowledge about your business, conversation rules, and escalation criteria. Think of it as the training manual for your AI receptionist.

Most effective prompts are 500-2,000 words. Short enough to stay focused, long enough to cover all common scenarios. The prompt should include your business information, FAQ answers, call routing rules, tone guidelines, and escalation criteria.

You should use the same base template but customize location-specific details: address, phone number, hours, staff names, and unique services. Keep the tone, personality, and general procedures consistent across locations for brand coherence.

Review weekly during the first month, then monthly after that. Update immediately whenever your business changes: new services, new hours, new staff, new policies, or new seasonal offerings. An outdated prompt gives callers wrong information.

The escalation rules - knowing when to transfer to a human. Getting this wrong either frustrates callers (too few transfers) or overloads your staff (too many transfers). Start conservative (transfer more often) and reduce transfer triggers as confidence grows.

Write the prompt in conversational language, not bullet points. Include example phrases the AI should use. Define 3-4 personality adjectives with specific behavioral instructions. Avoid corporate jargon, excessive formality, and robotic phrasing. Read your prompt aloud - if it sounds unnatural to you, it will sound unnatural to callers.

Regulations vary by region, but transparency is generally the best practice. A brief disclosure in the greeting ("You are speaking with our AI assistant") is the usual way to tell callers without hurting the caller experience, though in the EU the AI Act places the duty on how the system is designed (see the answer on what the EU AI Act requires). Most callers do not mind - they care about getting help efficiently.

Article 50(1) of the EU AI Act requires providers to ensure that AI systems intended to interact directly with people are designed and developed so that those people are informed they are interacting with an AI system, unless that is obvious to a reasonably well-informed, observant and circumspect person. Article 50(5) requires that information to be given in a clear and distinguishable manner, at the latest at the time of the first interaction; on a phone call, that points to the greeting. The duty is about how the system is designed, so one greeting line does not settle it on its own. It applies from 2 August 2026. Rules outside the EU differ by country.Source: EU AI Act: Article 50

Most AI platforms support separate prompts for each language. Write a complete prompt in each language your business serves, maintaining the same information and rules but adapting tone and phrasing for cultural norms. The AI detects the caller's language and uses the appropriate prompt.

Include explicit instructions: acknowledge the frustration ("I understand this is frustrating"), do not argue or defend, offer to transfer to a manager if the issue is beyond the AI's scope, and never take a condescending or dismissive tone. If the caller is abusive, the AI should remain calm and offer a transfer.

Some AI platforms support A/B testing natively. If yours does not, you can test manually by running one version for a week, measuring metrics, then switching. Compare call resolution rate, average duration, and transfer rate between versions to determine which performs better.

Yes. This guide gives a six-block skeleton and two complete, filled-in prompts for UK businesses, a dental practice in Leeds that sees NHS and private patients and a plumbing and heating firm in Birmingham, plus a booking script for three objections: the price question, the same-day request and the caller who is ringing round. Every business fact sits in square brackets. Replace each one with your own and delete any line you cannot fill, so the agent never has to guess.

Tell it in the prompt. Phone numbers in short groups, read back for the caller to confirm; dates the way a person says them; prices in words; names and email addresses spelled back letter by letter, using the phonetic alphabet (Alfa, Bravo, Charlie) for letters that could be misheard. Speech engines need this stated: the W3C speech synthesis standard and its say-as note let authors mark dates, times, telephone numbers and spelled-out characters as different kinds of text. The exact grouping is convention.Source: W3C SSML say-as note

Yes, as a first draft. Ask a general AI chat assistant to interview you one question at a time through the six blocks, write the result as spoken sentences, never invent a fact about your business, and mark anything you did not give it as [CHECK]. Then check every fact yourself: NIST describes confabulation, confidently stated but false content, as a natural result of the way generative models are designed. Use invented examples, never real callers' details.Source: NIST AI 600-1 (2024)

The reason for the call first (the greeting's open question), then the caller's name, then a callback number in case the line drops, then only the questions that type of call needs, one at a time, with each answer read back. Ask only what the business will use: UK GDPR requires personal data to be adequate, relevant and limited to what is necessary. Before reading out anything personal, such as an appointment time, the agent should confirm the caller is the person it belongs to.Source: ICO: data minimisation

JB
Justas Butkus

Founder & CEO, AInora

Building AI voice agents that let businesses serve more clients with the same team. The agents call back every new lead, re-contact past customers and take the calls the team cannot get to, and the work lands in the client's own CRM.

View all articles

Ready to try AI for your business?

Hear how AInora sounds handling a real business call. Try the live voice demo or book a consultation.