How to Write an AI Receptionist Call Script
By CodexierPublished 7 min read
An AI receptionist is only as good as the instructions behind it. The technology can hold a conversation; what it should say, what it may book and when it must hand over to you is a business decision nobody else can make. This guide shows how to turn the calls you already get into a script with clear boundaries, in an afternoon, using nothing more than a call log and a spreadsheet.
Collect your ten most common calls
Do not start from a vendor's generic template. Start from your own phone. For two weeks, keep a simple sheet by the phone or in your notes app with four columns: date, who (new customer, existing, supplier, other), what they wanted, and what happened. If you already use a switchboard or a mobile operator's call log, you get the timestamps for free and only need to add the reason. At the end you will have a list that looks nothing like the demo and everything like your business, including the calls you would rather an AI never touched.
Write the greeting and identification
The greeting has three jobs: confirm the caller reached the right company, say that an assistant is answering, and invite the errand. Keep it under ten seconds. Then identify the caller only as far as the errand needs. A price question needs no name; a booking needs a name and a mobile number for the confirmation; a reschedule needs a way to find the existing booking. Asking for everything up front is the fastest way to make a caller hang up.
- Greeting: company name, that an automated assistant is answering, and one sentence about what happens to the call. The privacy side of that sentence is covered in call recording and GDPR.
- Open question: "How can I help you today?" rather than a menu. The model classifies the errand from the answer.
- Identification on demand: ask for a name and number only once the AI knows it will book, change or look something up.
- Confirmation read-back: repeat the number and the time slot before writing anything into the calendar.
Answers, bookings and limits
Now write one block per call type. Each block needs the same four parts: the trigger phrases people use, the questions the AI must ask to complete the errand, the action or answer, and the hand-off rule. The table shows how that looks for a typical salon or workshop; the shape is the same for a clinic, a plumber or an accounting firm.
| Call type | Questions to ask | Action | Hand over when |
|---|---|---|---|
| New booking | Service, preferred day, name, mobile | Offer real slots, book, send SMS | Service not on the list, or caller asks for a specific person who is fully booked |
| Reschedule or cancel | Name, current time | Find booking, move or cancel, confirm | Booking cannot be found, or it is within the late-cancellation window |
| Price question | Which service | Read the list price and any from-price wording | Caller wants a quote for something not on the price list |
| Is it ready or status | Name or order reference | Read status if a system is connected | No system connected: take a message with callback |
| Complaint | None beyond listening | Apologise, take a message | Always, to a named person, same day |
| Supplier or sales call | Company name | Take a message | Never transfers; owner calls back if relevant |
Limits are as important as answers. Anything involving money beyond the list price, health details, or an upset caller belongs with a human.
Write the answers in the words you would use, not in marketing language. If your price list says "from", the AI must say "from" too, and it must never invent a figure. If a service has conditions, such as a patch test before a colour treatment, put the condition in the booking block so it is said every time. The AI will follow the script far more consistently than a busy human, which is a strength only if the script is right.
Hand-off and escalation rules
Escalation is where most AI phone projects succeed or fail. A caller who is stuck in a loop will not call back; a caller who is smoothly passed to a person forgives almost anything. Define three levels and decide, per call type, which one applies.
Level 1: answer and close
Routine bookings and factual questions. The AI completes the errand and sends a confirmation. No human involved.
Level 2: message with callback
The AI takes name, number and a one-line summary, promises a callback window you can actually keep, and posts the message to your inbox or CRM.
Level 3: live transfer
Complaints, emergencies and anything the caller insists on. The AI transfers immediately to a number you define, with a fallback to level 2 if nobody picks up.
Add two global triggers that override everything: the caller says they want a person, and the AI has failed to understand twice in a row. Both should go straight to level 3 or, outside opening hours, to level 2 with an honest message about when someone will call. Our AI receptionist ships with these rules configured; the values, such as who receives transfers and which callback window you promise, are yours to set.
Test calls before go-live
Testing is not optional and it is not the vendor's job alone. Sit down with two colleagues, the script, and a list of at least twenty calls to make. Ten should be the ordinary cases from your log. The other ten should be the awkward ones: a caller who changes their mind halfway, a name the model has never heard, background noise from a car, a child speaking, an English speaker, and someone who asks a question outside the script entirely. Score each call on three things: was the errand completed, did the AI say anything untrue, and did the hand-off fire when it should have. Fix the script, not the caller, and repeat until a full round passes.
When you do not need a script like this: if you receive a handful of calls a week and answer them yourself, an AI receptionist is a solution to a problem you do not have, and a good booking page will serve you better. If every call is a unique consultation, the script would consist entirely of hand-offs, and you should keep a human on the line. We say so on a 15-minute call rather than after selling you a setup.
Frequently asked questions
How long should an AI receptionist script be?
Long enough to cover your ten most common call types with a trigger, questions, action and hand-off rule each, and no longer. In practice that is two to four pages. Anything beyond that is usually edge cases that should simply hand over to a person.
Should the AI use my company's tone of voice?
Yes, within reason. Write the greeting and the standard answers the way your best receptionist would say them, including how you address customers. Avoid jokes and slang; they age badly and are easy to misread on a poor line.
What if the caller speaks English?
Include the English trigger phrases and answers in the same block, or let the platform switch language automatically if it supports it. Test it explicitly: a script that only works in Swedish will fail the first tourist who calls a restaurant.
How often should we update the script?
Review the call summaries after the first week and then monthly. New services, changed prices and seasonal hours all need edits. A script that is six months out of date will confidently give wrong answers, which is worse than voicemail.
Want a second pair of eyes on your script?
Fifteen minutes with an engineer: bring your call log and we sketch the ten blocks and the hand-off rules with you, whether or not you buy the receptionist from us.
Book a free 15-minute call