An AI receptionist call script is a structured set of business rules — not a literal word-for-word screenplay. The AI handles natural conversation; you give it the rules of the road. Tell it the greeting, the questions every caller must answer, the conditions for booking vs. transfer, and the escalation triggers — and it figures out the conversational glue. Trying to script every line is the fastest way to make your AI sound like a 1995 phone tree. This guide walks through the seven parts of a working script, with examples you can copy, and the eight mistakes that make scripts brittle.
1. The greeting
The greeting sets tone in the first three seconds. Keep it under 12 words. Three rules:
- Identify the business
- Identify the assistant (don't pretend to be human — it backfires when caught)
- Open with a question, not a statement
Thanks for calling Acme Plumbing — this is the assistant.
How can I help today?
Avoid "press 1 for…" — you're not an IVR. Avoid "I'm an AI assistant designed to help you with your plumbing needs" — too long, too formal. (Our AI receptionist solutions page shows the greeting templates we ship with.)
2. Business-hours rules
The AI needs to know what time it is in your time zone and what to do at each window:
Hours: Mon–Fri 7am–6pm, Sat 8am–2pm (America/New_York)
If during hours:
- Try to book directly into the calendar
- Transfer to dispatch on emergency keywords
If after hours:
- Book non-emergency appointments for next business day
- On emergency keywords, get callback number and SMS owner
- Don't say "we're closed" — say "we open at 7am"
The phrase "we're closed" loses calls. "We open at 7am, can I book you the first slot?" keeps them.
3. Qualification questions
Pick 3–4 questions every call must answer before booking or transferring. More than 4 and the call feels like an interrogation.
For most service businesses:
- Caller's name
- Best callback number
- Service address or ZIP code (filters out-of-area calls)
- What's going on? (the problem, in their words)
Tell the AI to ask one at a time and confirm back. Batched questions ("can I get your name, address, and phone?") feel robotic.
4. Booking rules
This is where most scripts fail. Be explicit about what the AI can book directly vs. what needs a human.
BOOK DIRECTLY:
- Drain cleaning, faucet repair, water heater diagnostic
- Slots: any open 2-hour window on the calendar
- Confirm: date, time window, address, phone
CALLBACK ONLY (don't book):
- Whole-house repipe
- Commercial jobs
- Anything over $2,000 estimated
NEVER QUOTE A PRICE for:
- Water heater install (varies by tank size and location)
- Sewer line work
Say: "Pricing depends on a few factors — a tech can give you
a firm number after a quick look. Want to book a free estimate?"
The "never quote" list saves you from the AI confidently making up a $400 number for a $1,800 job.
5. Escalation rules
Escalation is when the AI hands the call to a human. Three trigger types:
- Keyword triggers: "supervisor", "manager", "speak to a human", "I want to cancel my account"
- Emergency triggers: "no heat", "leak", "flooding", "smoke", "gas smell", "chest pain" (per industry)
- Confusion triggers: AI hits "I'm not sure how to help with that" twice in one call
Rules around escalation:
ON ESCALATION:
- During hours: warm transfer to (555) 123-4567
- If no one picks up after 4 rings: take message + SMS owner
- After hours, emergency only: SMS owner + give caller ETA
- Never escalate without the caller's name and callback number
6. Handling unknowns
The script must tell the AI what to do when it doesn't know. The default ("I don't have that information") sounds bad. Better:
If asked something not in the knowledge base:
- Acknowledge: "Good question — let me get someone who knows
that to call you back."
- Capture: name, callback number, the specific question
- SMS the owner with the question + caller info
The AI should never invent: prices, hours, service availability, addresses, staff schedules, or warranty terms.
7. A/B testing the script
Once you've got the first version live, change one variable at a time and watch booking rate. Things worth testing:
- Greeting tone (formal vs. friendly)
- Number of qualification questions (3 vs. 4)
- Booking confirmation style ("does Tuesday at 2 work?" vs. "I'll book you Tuesday at 2 unless that doesn't work")
- Asking ZIP before vs. after the problem description
Run each variant for 50 calls minimum. Compare booking rate and call duration. Smaller is usually better — if call duration drops 30 seconds and booking rate holds, you just bought back time and money.
8. Common script mistakes
- Scripting every line. AI sounds robotic. Give rules, not lines.
- Asking 6 qualification questions. Customers hang up at 4.
- No "never quote" list. AI invents prices.
- Vague escalation triggers. Either everything escalates or nothing does.
- No after-hours behavior defined. Calls die in voicemail purgatory.
- Not telling the AI about staff names. Customers ask for "Mike" and get redirected to a generic flow.
- No service-area filter. The AI books appointments 80 miles away.
- Never reviewing transcripts. You can't improve what you don't read.
Three example scripts
Example: Solo HVAC operator
GREETING: "Thanks for calling Smith HVAC — this is the assistant."
QUALIFY (one at a time):
1. Name
2. Callback number
3. Address or ZIP (must be in 30-mile radius of 14620)
4. Heating, cooling, or other?
BOOK DIRECTLY: tune-ups, diagnostics, filter changes
CALLBACK: full system replacement, commercial, anything > $1500
EMERGENCY KEYWORDS: "no heat", "no AC and elderly", "gas smell"
→ SMS owner immediately, give caller "we'll call within 30 min"
NEVER QUOTE: install pricing, repair pricing
SAY INSTEAD: "Diagnostic visit is $89. The tech gives a
firm quote on-site before any work."
Example: Dental office
GREETING: "Bright Smiles Dental, this is the assistant. How can I help?"
QUALIFY:
1. Name + DOB (existing patient lookup)
2. New patient or returning?
3. Reason for visit (cleaning, pain, consult, emergency)
4. Insurance carrier (don't quote coverage — capture only)
BOOK DIRECTLY: cleanings, exams, new-patient consults
CALLBACK: implants, ortho consults, billing questions
EMERGENCY: "tooth knocked out", "swelling", "severe pain"
→ next available slot today, page on-call dentist if after hours
NEVER QUOTE: insurance coverage, procedure pricing
SAY INSTEAD: "Our front desk can verify your benefits and
give you exact numbers — want me to have them call back?"
Example: Law firm intake
GREETING: "Garcia Law, this is the intake assistant. How can I help?"
QUALIFY:
1. Name + callback
2. Type of matter (PI, family, criminal, business, other)
3. Brief description (one or two sentences)
4. Are there any deadlines or court dates?
NEVER:
- Give legal advice
- Quote fees
- Confirm representation
ALWAYS:
- "Nothing said here forms an attorney-client relationship"
- Book a 15-min consult or capture for callback
- Conflict-check question: "Are you currently represented?"
EMERGENCY ROUTING: arrest, restraining order, deadline
in next 48 hours → page on-call attorney
Frequently asked questions
How long should an AI receptionist script be?
The rule set behind the script is usually 1–3 pages. The greeting itself is one or two sentences. Most vendors break the script into structured fields rather than one long document — greeting, hours, services, booking rules, escalation triggers, knowledge-base entries — so "length" is the wrong measure. Aim for completeness instead. A small home-services business will have roughly 12–20 service entries, 4 qualification questions, 5–8 escalation triggers, and a "never quote" list of 3–5 items. A multi-location operation has more knowledge-base entries (one per location) but the same rule structure. If you find yourself writing more than 3 pages, you're probably scripting lines instead of rules.
Should I write the script myself or use the vendor's template?
Start with the vendor's template, then change three things specific to your business: your "never quote" list, your escalation triggers, and your service-area filter. Templates are a strong starting point because they encode the patterns the vendor has seen across thousands of customers — the wording, the order, the recovery flows when the AI gets stuck. What templates can't know is what makes your business different: which jobs you don't take, which customers get priority, which staff need to be paged for what. Spend your time on the differentiators, not on rewriting the greeting from scratch. After 50 calls, you'll have data to refine — that's when custom edits pay off.
Can the AI follow scripts in multiple languages?
Yes — most modern AI receptionists handle Spanish and English natively, with the better ones supporting French, Mandarin, Vietnamese, Tagalog, and 20+ others [VERIFY]. The script logic is the same across languages — the AI translates the rules. What changes is the greeting (you may want a bilingual greeting like "Para español, oprima dos" — except don't, because the AI handles language switching automatically when the caller speaks Spanish). Test the non-English path explicitly during setup. Booking confirmations and follow-up SMS can also be sent in the caller's chosen language. If your customer base is meaningfully bilingual (over 15%), set the greeting to acknowledge both languages from word one.
How do I make the AI sound less robotic?
Three things: shorter sentences, conversational filler, and natural confirmations. Replace "I have successfully scheduled your appointment for Thursday at 2pm" with "Got it — Thursday at 2." Allow contractions ("you're", "we'll"). Tell the AI to acknowledge before answering ("Yeah, totally — let me check that") instead of jumping straight to the answer. Most vendors expose a "personality" or "tone" setting; pick "casual" or "friendly" rather than "professional" unless you're a law firm or medical practice. Also: voice quality matters more than wording. A premium voice model with 11 Labs / OpenAI / Cartesia-grade synthesis will sound natural even with a basic script. A cheap voice will sound robotic no matter how good the script is.
How often should I update the script?
Review weekly for the first month, then monthly afterward. Track three signals: booking rate, escalation rate, and call duration. If booking rate drops, something in the booking rules is off. If escalation rate spikes, an unhandled call type is showing up — add a knowledge-base entry. If duration creeps up, the AI is getting stuck somewhere — usually a qualification question that's confusing callers. Set a recurring 30-minute review on your calendar. Most updates are 60-second changes in the dashboard: adding a service, fixing a price, retiring a phrase. Treat the script as a living document, not a one-time setup task. The teams that get the best results from an AI receptionist are the ones who tune it monthly.
Ready to see what an AI receptionist looks like for your business? → Book a 20-minute demo