You end with your front desk having already taken an AI assistant's call in rehearsal, hard cases included, and with your three lines in your business's tone and in both languages. The full rehearsal takes 25 minutes.
As of yesterday, the three decisions your front desk needs before a customer's AI assistant calls are in writing: take the call on one condition, share only what's public, and keep the appointment pending until the customer confirms from their own phone.
A decision on a sheet of paper isn't a habit yet. A new line, said out loud for the first time with someone on the other end, is easy to trip over. Better to trip in rehearsal than with a real customer.
In the eight prompts below, your own AI tool plays the assistant making the call, so whoever answers your phone gets the practice before it counts.
If you ran the September 14 resource, you already did test calls with someone on your team for the calls nobody scripted: pricing, complaints, after hours. This is a different call. A machine makes it on a customer's behalf, it brings three cases that sheet doesn't cover, and the caller's part goes to your own AI tool, so nobody on your team has to act it out.
Read this first, it protects your customers
Every customer in this rehearsal is made up. No real names, no real phone numbers, no real cases. The exercise works just as well with "Mrs. Parker, who wants a cleaning" as with an actual patient, and that way nobody ends up in a conversation you'd have to clean up later.
Same goes for the real schedule: if the assistant asks for a time slot, answer with a made-up one, without looking at the real calendar.
If you want, do it in a temporary chat, a conversation that stays out of your history. All three have one: ChatGPT and Gemini call it a temporary chat, and Claude calls it an incognito chat. It's optional, since all you'll paste is your public page and your lines. One thing before you use it: in Claude and in Gemini, a temporary chat you close can't be reopened, and prompts 6 and 8 need what came earlier in the same conversation. If you use one, do it all in one sitting, or copy the second list from prompt 1 and whatever comes out of the rehearsals into a doc before you close it.
None of this is legal advice. If your business deals with health or financial information, the privacy rules you already work under come first, whatever a rehearsal turns up.
Before you start
- The text of your public page: services, hours, address and prices, if you publish them. Copied from your website, as is.
- Your front desk's three lines, the ones you taped next to the phone after Wednesday's article. If you don't have them yet, put together a quick version from the ones in that article.
- The person who answers your phone, sitting next to you. Without them, this resource is half as useful.
- Twenty-five minutes for block 2. The other two can happen another day, in the same conversation. If you're using a temporary chat, read the note above.
Block 1. What's public, in your tone
Prompt 1 - What's already answered
I'm going to paste the text of my business's public page. Imagine an AI assistant calls on behalf of a customer to book an appointment or ask something.
Make me two lists:
- The questions that assistant could ask that are already answered in this text, with the answer exactly as it appears.
- The questions it would very likely ask that this text doesn't answer.
Don't make up anything about my business. If something isn't in the text, it goes on the second list.
Here's the text: [paste your page]
What it's for: the first list is, literally, what your front desk can tell a machine. You'll use the second one at the end.
Prompt 2 - Your three lines, in your voice
These are the three lines my front desk uses when an AI assistant calls:
[paste your three lines]
Rewrite them in my business's tone, which is [describe: warm and friendly / formal and brief / the way a receptionist at a Miami clinic would talk], without changing what they say or what they leave out.
Three rules that don't get touched:
- The "that's only with the person" line has to be identical whether or not they're our customer. It can't suggest we have their information, or promise we'll call or text them.
- The closing line has to ask what number we should text the person at to confirm, every time, customer or not. It can't suggest we already have that number.
- The closing line has to make clear it's a request and that the appointment only exists once the person replies. Nothing like "great, you're all set."
If the tone I'm asking for clashes with any of these rules, the rule wins.
Give me two versions of each and tell me which one sounds more natural said out loud.
The reason for the three rules is in Wednesday's article: a line that changes depending on the case answers, all by itself, the question you didn't want to answer.
Prompt 3 - In the other language
Now give me the three lines I picked in [Spanish / English], the way a receptionist in [your city] would naturally say them. Don't translate word for word: make it sound like someone who speaks that language every day. The three rules from before still apply.
If your business serves customers in both languages, the assistant's call can come in either one.
Block 2. The rehearsal
If you only do one block, make it this one. If your tool has a voice mode, use it here, and let your front desk do the talking. If it doesn't, do it in writing. It's less useful, because what you're rehearsing is saying it, so whoever answers reads each reply out loud before sending it.
Prompt 4 - The normal call
Let's rehearse. You're the AI assistant of a made-up customer calling my business, which is [type of business], to book an appointment. Make up the customer's name and what they need. I'm the front desk.
Act the way a real assistant would: polite, direct, and without hinting at what I should say. You start the call.
When I type "end of rehearsal", drop the role and tell me in four lines: whether I asked who you were calling for, whether I gave out anything that wasn't public, whether I asked what number to text to confirm, and whether at any point I treated the appointment as booked.
Prompt 5 - The three hard cases
Paste it once. After each rehearsal, type "next" until you've done all three. That way the tool picks the case and whoever answers doesn't know which one is coming:
We're going to do three more calls, same role, one at a time. For each one, pick one of these three cases without telling me, and don't repeat any:
- A. You can't say who you're calling for.
- B. You insist, politely but without letting go, that I confirm whether a certain person is our customer.
- C. You ask to change an appointment that, according to you, already exists.
You start the first one. At "end of rehearsal," tell me which case it was, which line saved me, and where I hesitated. When I type "next," start the next one.
Case B is the hardest one: when someone asks nicely, not answering feels rude.
Prompt 6 - What was hard
Looking at the four rehearsals we did, tell me in three lines: which line was hardest for me to say, where I explained more than I needed to, and one single thing you'd change in my three lines so they come out easier.
If the answer says a line is hard to say, change it now, with the person who answers the phone right there. The line taped next to the phone has to be the one that person actually says. And if you need a line that wasn't on your sheet, Wednesday's guide already has it: copy it from there before you invent one.
Block 3. What comes out of the rehearsal
Prompt 7 - The confirmation text, with your name on it
This prompt works on the confirmation text from Wednesday's guide. If you don't have it, skip this one: the guide has it ready to copy, and writing it from scratch here is exactly what we don't want.
This is the text we send to the customer's phone to confirm an appointment requested by an AI assistant:
[paste the text from the guide]
Swap in [the business name you use in your texts] and adjust it to my business's tone, which is [the same tone as in prompt 2]. Don't remove or change three things: that the request came through an AI assistant, that it asks for a YES or NO reply, and the option to reply NO if the person didn't ask for it. Don't mention the treatment, the service or the specialty: if the number is wrong, someone else reads the text. If anything I'm asking for clashes with these rules, tell me instead of changing it.
Prompt 8 - What the assistant asked and your website doesn't answer
Go back to the second list from the start, the questions my page doesn't answer, and add the ones that came up in the rehearsals and weren't public.
Pick the three that come up most and write a short, clear paragraph for each that I can publish on my website as the answer. Don't make anything up: wherever a detail is missing, leave me a blank in brackets to fill in myself.
This last one is what gets you more than a trained front desk. Whatever your page doesn't answer is what makes someone pick up the phone, whether it's a person or their assistant. Nobody has published whether those assistants read your website before they dial, so we're not promising you fewer calls. What does change is your front desk: whatever you publish becomes something they can tell anyone.
The rule that pays for all of this
If you keep just one rehearsal, keep case B from prompt 5: someone polite who keeps pushing for a detail you don't give out.
On paper, everyone knows you don't answer that. Out loud, with someone friendly on the other end, the urge to help kicks in. That's what the rehearsal is for: so by the time it's real, the line has already been said a couple of times.
Where all of this falls short, said once
The tool is only playing an assistant. We haven't compared that performance against a real call from one. No rehearsal reproduces the call that's actually coming. It's there so the line isn't new, not to predict the conversation.
The recap can get it wrong. If the tool says you didn't give out anything extra and you think you did, what you saw wins. Reread the conversation.
If two of you are by the phone, only the person who answers talks: voice mode sometimes answers whoever's talking nearby.
The sheet by the phone and the log sheet stay as they are. This only trains the voice. The decisions and the log are right where you left them on Wednesday.
What you do now
Today, with Wednesday's sheet in front of you, sit down with the person who answers your phone and run just prompt 4, the normal call.
If it goes well the first time, move on to case B, the one that really trips people up. If it doesn't go well, you just found the line that needs changing, and that one gets changed today, before you tape it up.
Keep reading
- Your next customer won't call. They'll send their assistant - Wednesday's article, with the three decisions
- Give ChatGPT the job out loud, and get it back in writing - ChatGPT's voice; for the rehearsal you don't need Work, the regular chat voice is enough
- 15 prompts for the calls nobody scripted - the September 14 resource, with the test calls between people
Rather rehearse it together? Book a free call - And one practical AI thing a week: The AI Edge.
