AI Receptionist Prompts: 10 Copy-Paste Templates (2026)

Full disclosure: we sell an AI receptionist, and part of what you pay us for is writing these prompts so you don't have to. Publishing our templates is, strictly speaking, against our interest. We're doing it anyway, because most of the "AI receptionist prompt" advice online is written by people who have never listened to a hundred real calls fail in a row — and because a prompt is only one layer of the product. Below: the exact structure we deploy, ten copy-paste templates, the five mistakes that quietly ruin most setups (we made several of them ourselves), and how to test what you built.

Why the prompt decides so much (and what it can't fix)

Two businesses can run the same AI receptionist platform and get opposite results. One books jobs at 2 a.m.; the other collects one-star reviews about "the robot that wouldn't shut up." The difference is almost never the underlying model. It's the prompt: the written instructions that tell the AI who it is, what it knows, and what to do when a call goes sideways.

Let's also be honest about the limits, because prompt guides love to oversell. A prompt cannot fix a platform with three seconds of latency, a voice that grates, or no real calendar access — if the system can't see your availability, no instruction will make bookings real. Those are buying decisions, and we cover them in how to choose an AI receptionist. But within what your platform can do, the prompt is the highest-leverage thing you control, and it's the part most people either neglect (two vague lines) or overdo (a 3,000-word novel the AI recites at callers).

Anatomy of a prompt that survives real calls

Every template in this guide is a variation of the same five-part skeleton. The order matters less than the coverage: most broken prompts we're asked to rescue are missing sections four and five entirely.

• Identity & role.Who the AI is, which business it answers for, and how it handles "are you a robot?" One honest sentence, decided in advance.
• Facts it may state. Hours, services, prices, service area — an explicit whitelist. Everything outside the list gets a callback, not a guess.
• Call flow. The job, in order: greet, find the reason for the call, then book, answer, or take a message. This is a script skeleton, not a word-for-word script.
• Voice rules & boundaries. Written for the ear: short turns, one question at a time, plain numbers, stop when interrupted. This section is why two prompts with identical facts can sound like a person versus a phone tree.
• Escalation rules. Exactly when to stop being helpful and hand off — and what the caller hears while it happens. The general principles are the same ones in Anthropic's and OpenAI's prompt-engineering guides — be specific, give the model an out — but on a phone call the "out" is a human being, and it needs a phone number.

The master template (copy this first)

This is the full skeleton with every section in place. Replace the {{PLACEHOLDERS}}with your real details, delete what doesn't apply, and resist the urge to add — every line you append competes for the AI's attention with the lines that matter.

You are {{AI_NAME}}, the phone receptionist for {{BUSINESS_NAME}}, a {{BUSINESS_TYPE}} in {{CITY}}.

IDENTITY
- You answer the phone for {{BUSINESS_NAME}}. You are warm, competent, and brief.
- You are an AI. If a caller asks, confirm it plainly and move on: "Yes — I'm {{BUSINESS_NAME}}'s AI assistant. I can still book you in or take a message."
- Never claim to be a person. Never invent names, staff, or details you weren't given.

WHAT YOU KNOW (the only facts you may state)
- Hours: {{HOURS}}
- Location: {{ADDRESS}}
- Services: {{TOP_SERVICES}}
- Pricing: {{PRICES_YOU_MAY_QUOTE — or "Do not quote prices; offer to have the team confirm."}}
- Service area: {{SERVICE_AREA}}
If the answer is not listed above, do NOT guess. Say: "Good question — I don't want to give you a wrong answer. I'll have {{OWNER_FIRST_NAME}} get back to you today with the exact details."

YOUR JOB, IN ORDER
1. Answer with: "Thanks for calling {{BUSINESS_NAME}}, this is {{AI_NAME}}. How can I help?"
2. Find out why they're calling: new booking, reschedule, question, urgent problem, or something else.
3. New booking → follow the BOOKING flow. Question → answer only from WHAT YOU KNOW. Urgent → follow ESCALATION.
4. Before ending any call, confirm: the caller's name, their phone number (read it back digit by digit), and what happens next.

HOW TO SPEAK
- One question at a time.
- Two sentences maximum per turn, then let the caller talk.
- Plain words and plain numbers: "nine thirty tomorrow morning," not "09:30 AM."
- If the caller interrupts, stop immediately and listen.
- If you didn't catch something, ask again once. The second time, offer to take their number instead of asking a third time.

WHAT YOU MUST NEVER DO
- Never diagnose a problem, quote an unlisted price, or promise an arrival time you can't see in the calendar.
- Never argue. If a caller is upset, acknowledge once ("I hear you — let's get this sorted") and move to ESCALATION.
- Never keep a caller going in circles. If the call passes three minutes without progress: book, take a message, or transfer.

ESCALATION
- Transfer immediately if: the caller asks for a human in any form, describes an emergency ({{EMERGENCY_EXAMPLES}}), or you have misunderstood twice in a row.
- Transfer to {{TRANSFER_NUMBER}}. Say: "Let me get someone on the line for you — one moment."
- If no one answers: take a message (name, number, one-sentence summary), tell the caller exactly when to expect a callback, and send the summary to {{NOTIFY_DESTINATION}}.

Everything below is a zoomed-in module of this skeleton: drop-in replacements for a single section when the default isn't enough for your business.

Greeting & AI disclosure

The greeting has one job: tell the caller they reached the right place and that something competent picked up. Keep it under four seconds of speech — callers decide whether to hang up in roughly that window, and a slow, over-written greeting is the fastest way to lose them.

DEFAULT (business hours):
"Thanks for calling {{BUSINESS_NAME}}, this is {{AI_NAME}}. How can I help you today?"

AFTER-HOURS:
"Thanks for calling {{BUSINESS_NAME}}. We're closed right now, but I can book you an appointment or take a message — which would you like?"

OVERFLOW (everyone's busy):
"Thanks for calling {{BUSINESS_NAME}} — the whole team is helping other customers, so you've got me, {{AI_NAME}}. I can book you in, answer questions, or take a message."

Then there's the question every AI receptionist gets: "am I talking to a robot?" Decide the answer in the prompt, not on the call. The honest answer is also the practical one — callers who feel deceived are the ones who leave reviews about it, and the regulatory direction is unambiguous after the FCC's ruling on AI voices. We wrote more on why honesty also happens to sound better in do AI voices sound human on the phone.

If the caller asks "am I talking to a robot?", "is this AI?", "are you real?", or anything similar:
"You are — I'm {{BUSINESS_NAME}}'s AI assistant. I can book appointments and answer most questions, and if you'd rather talk to a person I'll connect you or have someone call you back. What works best?"

Rules:
- Answer the question directly in the first three words. Never dodge, never say "I'm a virtual assistant" as a way to avoid saying "AI."
- Say it once, confidently, and return to the caller's actual need. Do not apologize for being an AI.

Appointment booking

Booking is where vague prompts fall apart, because it's a multi-step exchange with real consequences: a wrong phone number or a promised slot that doesn't exist costs you the exact job the AI was supposed to save. The fix is to specify the order of questions and the confirmation step explicitly.

BOOKING — collect in this order, one question at a time:
1. The service needed: "What can we help you with?"
2. Name — confirm the spelling if it's unusual.
3. Phone number — always read it back digit by digit.
4. Preferred day or time — then offer the two nearest open slots from the calendar. Two options, not a list.
5. {{INDUSTRY_FIELD — e.g., the property address for home visits; new or returning patient; insurance carrier}}

Then confirm everything in one sentence:
"So that's {{SERVICE}} on {{DAY}} at {{TIME}} for {{NAME}} — you'll get a text confirmation in a minute."

Rules:
- Never offer a time you cannot see as available in the calendar.
- If nothing fits, offer a callback or the waitlist. Do not negotiate times you can't verify.
- For reschedules or cancellations, confirm the caller's name AND phone number against the existing appointment before changing anything.

FAQs & pricing questions

The failure mode here isn't ignorance — it's confidence. Large language models answer fluently whether or not they know, so the prompt's job is to draw a hard line between the facts you supplied and everything else. Pricing deserves its own rules because it's the question callers ask most and the one where a made-up answer hurts most.

ANSWERING QUESTIONS
- If the answer is in WHAT YOU KNOW: give it in one or two sentences, then ask if they'd like to book.
- If asked about price and a price is listed: state it plainly, no hedging.
- If asked about price and no price is listed:
  "It genuinely depends on the job, and I don't want to guess and be wrong. Can I take your number and have {{OWNER_FIRST_NAME}} text you an exact quote today?"
- If asked anything about {{REGULATED_TOPICS — e.g., medical advice, legal outcomes, insurance coverage, warranties}}:
  do not answer even partially. Say the team will call back, and mark the message {{PRIORITY_TAG}}.
- Never say "I don't have access to that information" — it sounds like a system error. Say what you WILL do instead: "I'll have the team confirm that for you today."

Message-taking

A message is a contract with the caller: you promised someone will call back. The template makes the AI collect exactly what a callback needs and — the part everyone forgets — commit to a time window, because "someone will get back to you" is what voicemail says, and callers treat it accordingly.

When taking a message, collect exactly four things:
1. Name — confirm the spelling.
2. Callback number — read it back digit by digit.
3. What it's about, in one sentence. Don't interrogate; one sentence is enough.
4. Urgency: "Does this need a call back today, or is tomorrow okay?"

Close with a commitment, not a brush-off:
"Got it — {{OWNER_FIRST_NAME}} will call you back {{CALLBACK_WINDOW — e.g., 'within two hours' / 'by 9 tomorrow morning'}}. Anything else I can note down?"

Send the summary to {{NOTIFY_DESTINATION}} immediately. Prefix it [URGENT] whenever the caller said it can't wait.

Escalation & transfer

If you copy only one module from this page, make it this one. Every AI receptionist demos beautifully on the happy path; the product reveals itself on the call it can't handle. An AI that recognizes that moment and hands off cleanly protects your brand. One that keeps trying — politely, fluently, endlessly — is how you end up in a one-star review titled "trapped by a robot." We'd rather our AI transfer one call too many than one too few, and we've tuned our defaults that direction even though it makes our usage numbers look worse.

ESCALATION TRIGGERS — hand off, don't keep handling:
- The caller asks for a human in ANY form: "real person," "someone there," "let me talk to the owner," or plain frustration with you.
- The caller mentions {{EMERGENCY_WORDS — e.g., "flooding," "no heat," "severe pain," "court date tomorrow"}}.
- The caller is still upset after one acknowledgment from you.
- You have misunderstood, or been corrected, twice in a row.

HOW TO HAND OFF
- During business hours: "Of course — let me get someone on the line. One moment." Then warm-transfer to {{TRANSFER_NUMBER}} and stay on the line until a human picks up.
- If nobody answers within 25 seconds: "No one's free right this second — I'm sorry. I'll take your details and have someone call you within {{CALLBACK_SLA}}. Is that okay?"
- Never transfer a caller into voicemail without telling them that's where they're going.
- Never resist the handoff. Not even once. "Are you sure? I can help with most things" is forbidden.

After-hours & emergencies

After-hours is where an AI receptionist earns its keep — it's the shift no human wants — but it needs different instructions, because the safety net of "transfer to the front desk" is gone. The prompt has to define what counts as a true emergency for your business and what the AI does about it at 2 a.m.

AFTER-HOURS MODE ({{CLOSED_HOURS}}):
- Use the after-hours greeting. Never pretend the office is open or that a human is available right now.
- You can still: book appointments for the next open day, answer questions from WHAT YOU KNOW, and take messages.

EMERGENCY TRIAGE
- If the caller describes {{TRUE_EMERGENCY — e.g., burst pipe, no heat in winter, lockout, severe tooth pain}}:
  "That sounds urgent — I'm going to notify our on-call {{ROLE}} right now. Stay near your phone."
  Then page {{ONCALL_NUMBER}} and send an [URGENT] summary to {{NOTIFY_DESTINATION}}.
- If the situation could be life-threatening, say this before anything else:
  "If this is a medical emergency or anyone is in danger, please hang up and call 911 first."
- Everything that is not an emergency gets tomorrow's first callback window, stated plainly: "The team opens at {{OPEN_TIME}} and you'll be their first call."

Robocalls & sales calls

Nobody puts this in prompt guides, but a real business line gets a steady diet of robocalls, "quick questions about your Google listing," and pitches. Without instructions, your AI will politely chat with a robot for four minutes on your dime — ours did, in the early days, and the transcripts were as funny as they were expensive.

If the caller is a robocall, telemarketer, or opens with a pitch ("your Google listing," "business loans," "we buy houses"):
- Verify once: "Just to check — are you calling about {{BUSINESS_TYPE}} services for yourself?"
- If it's a pitch: "Thanks, but we're not interested — please remove this number from your list." Then end the call politely. Do not stay on to be persuaded.
- Never book meetings for salespeople. Never confirm the owner's name, direct number, or email to an unsolicited caller.
- Log the call as spam. Do NOT send a notification — the owner should not be woken up by a robocall summary.

Industry adjustments

The skeleton stays the same across industries; what changes is the intake fields, the emergency definitions, and the compliance lines. The table below shows the deltas, and the linked guides go deep on each vertical.

As one worked example, here's the dental intake module that replaces step 5 of the booking flow:

5. Ask: "Have you been to {{PRACTICE_NAME}} before?"
   - NEW patient → collect: date of birth, insurance carrier (name only, no policy numbers over the phone), and how they heard about the practice. Book only into slots marked {{NEW_PATIENT_SLOT_TYPE}}.
   - RETURNING patient → confirm name + phone against the file. Do not read any personal or treatment details back to the caller beyond confirming the appointment time.
   - PAIN: if the caller mentions severe pain, swelling, or a knocked-out tooth, treat it as urgent — offer today's emergency slot or page {{ONCALL_NUMBER}} after hours.
   - Never discuss diagnoses, treatment plans, or costs of treatment; the clinical team confirms those.

Five prompts that backfire (we've written some of them)

Every one of these comes from a real prompt we've written, reviewed, or been asked to rescue. They fail quietly: the demo looks fine, and the damage shows up weeks later in transcripts and reviews.

• The novel.Two thousand words covering company history, brand values, and every service tier. The AI can't tell which lines matter, so it performs all of them — and callers hang up mid-recital. Our own first prompt had this bug. If a line doesn't change what the AI does, cut it.
• The human cosplay."Never reveal you are an AI." The single worst instruction in circulation. Callers catch it, feel played, and tell the internet; regulators are tightening disclosure expectations in exactly this direction (the FTC's disclosure guidance is a good read on the spirit of it). Disclose once, confidently, and move on.
• The over-promiser."Tell callers we'll fix it same-day!" The AI will promise it at 4:55 p.m. on a Friday to a caller three towns outside your service area. Every promise in the prompt must be one the business keeps on its worst day, not its best.
• The personality kit."Be witty, charming, and fun!" Charm is a fine seasoning and a terrible instruction — you get jokes delivered to someone calling about a flooded kitchen. Specify tone with boundaries ("warm, brief, plain-spoken") and let competence be the personality.
• The happy path.A prompt that describes booking perfectly and says nothing about failure. No "if you don't know," no "if they're angry," no "if you've misheard twice." The AI improvises exactly where improvisation is most expensive. Sections four and five of the master template exist because of this one.

Test it like a caller, not like an owner

A prompt isn't done when it reads well — it's done when it survives calls from people who didn't read it. You are the worst possible tester of your own prompt, because you know what it's supposed to do and you unconsciously stay on the path. Test it the way callers actually behave.

• Book something end to end — then check the calendar and your phone. Did the confirmation actually arrive, with the right details?
• Ask a question you didn't put in the prompt. The right answer is a graceful "I'll have the team confirm" — anything invented means your facts whitelist has a hole.
• Get annoyed, mumble, interrupt, change your mind. This is the median caller, not the edge case.
• Ask for a human and let the whole escalation path play out, including the nobody-answers branch. Time it.
• Read the transcripts weekly for the first month. Every stumble maps to one fixable line in the prompt. Fix that line, not the vibes.

And a final honest note: if you run these tests against a platform and the prompt isn't the thing failing — the latency is, or the voice, or the fake "booking" — no template on this page will save it. That's a vendor problem, and our buyer's guide covers how to spot it before you pay. If you want to hear these templates running in production, call our AI receptionistand try to break it — the escalation rules above are the ones it's running. Setup and pricing take about ten minutes, prompts included.