How to run the demo call when you sell AI
The demo call that closes is the one where you talk least. Here is the twenty-minute sequence we use, the three questions worth planting, what to do when the agent stumbles in front of the buyer, and the cases where running a live demo is the wrong move.

Hand them the phone in the first two minutes. That is the demo call. The deck you were going to walk through, the explanation of what a language model is, the slide about the future of small business, all of it loses to a plumber asking your AI receptionist what a water heater swap costs and hearing the right number come back in their own price band.
The demo call that closes is the one where you talk least. Below is the sequence we use, what to set up before you dial, what to do when the agent stumbles in front of the buyer, and the cases where running a live demo is the wrong move.
What does a demo call for AI services actually look like?
Twenty minutes, and you are speaking for maybe six of them.
Minutes 0 to 2. Confirm what they do and who answers the phone now. One question: "When someone calls at 7pm on a Saturday, what happens?" You are not discovering anything. You are getting them to say the problem out loud so it is theirs, not yours.
Minutes 2 to 4. Send the link and stop talking. They open it on their own screen. Their business name is already on the page, their trade is already loaded, and the receptionist is waiting.
Minutes 4 to 12. They call it. They ask it things. You stay quiet unless they get stuck. This is the entire value of the call and most reps talk straight through it.
Minutes 12 to 16. Run the website audit against their real domain, then put their own ticket and close rate into the ROI figure. Numbers they supplied are numbers they cannot dispute.
Minutes 16 to 20. Price, and a decision. Not "let me send over some information."
If you have never watched the demo work without narrating over it, the breakdown of what each node does is worth ten minutes before your next call.
Why does handing over the phone beat explaining?
Because the person across the table is, statistically, a skeptic, and you cannot argue someone out of skepticism in twenty minutes.
Pew surveyed 5,119 U.S. adults in February 2026. Eight percent said they have a great deal or quite a bit of confidence in U.S. companies to develop and use AI responsibly. Thirty-eight percent said not too much, and 21 percent said none at all. Seventy-one percent said the increased use of AI will make their personal information less secure. Among the roughly half of adults who have never used an AI chatbot, 45 percent said not trusting them to give accurate information was a major reason (Pew Research Center, chapter on non-users).
That is your buyer. A contractor in his fifties who has read that AI makes things up and has no reason to think yours is different. Your claim that it works is one more claim. A call he placed himself, asking a question only he knew was coming, answered in his trade's language, is evidence. That gap is the whole reason the demo exists.
What to set up in the five minutes before the call
Four things, and none of them take long.
Their business name and their real website URL, typed into the demo before the call starts. A demo that says "Your Business Here" is a template and they will read it as one.
The right vertical pack. There are 100+ trade packs, each carrying services, price bands, and objections written for how that business actually sells. A roofing pack talks about tear-offs and insurance claims. The generic pack does not. Picking a close-enough neighbor because it is faster is the most common way reps lose these calls.
Your remaining voice minutes. The plan includes 100 live voice minutes a month and caps any single call at five minutes. A twelve-minute demo where the buyer stays on the phone the whole time will eat a meaningful slice of your month, so check before a day of back-to-back calls rather than after.
Three questions you want them to ask. Write them down. More on those next.
The three questions worth planting
You do not script the buyer. You do give them a starting point when they freeze, which they will, because most people have never spoken to a machine on purpose.
The price question. "Ask it what a standard service call costs." This is the one that proves the pack is real. A generic bot deflects to "our pricing varies." A loaded pack gives a band that sounds like their band.
The availability question. "Ask it if you can get in this week." This shows booking behavior without anything being booked, which is the point of a sandbox.
The awkward question. "Ask it something you would not expect it to handle." Let them try to break it. If it holds, you have a buyer. If it does not, you find out on a demo call instead of in month two of a signed account. Either outcome is better than not knowing.
What to do when the agent stumbles in front of them
It will, eventually, and the recovery matters more than the stumble.
The most common failure is the agent cutting the caller off. That is turn detection doing its job badly, and it is fixable. Voice systems decide the caller has finished either by measuring silence or by judging the words. OpenAI's Realtime API exposes both: server_vad, which chunks on a silence duration you set, and semantic_vad, which waits until the model believes the sentence is finished and has an eagerness setting from low to high to control how long it will hold (OpenAI voice activity detection guide). A caller who trails off, thinks, and keeps going will get clipped by a short silence timer. Older buyers do this constantly.
The second failure is silence. People notice conversational gaps at a resolution most of us would not believe. Stivers and colleagues measured turn-taking across ten languages from indigenous communities to major world languages and found every one of them avoided overlap and minimized silence between turns, with average gaps sitting within a 250 millisecond band of the cross-language mean (PNAS, 2009). A pause your buyer cannot name still registers as wrong.
When either happens, do not apologize and do not explain the architecture. Say "try that again, it clipped you," and let them. A system that misbehaves and recovers reads as real. A demo that never misses reads as a recording.
One caution. A smooth answer is not proof that anything happened. OpenAI's own voice agent guidance puts it plainly, warning that "a natural-sounding response does not prove that a tool ran or that application state changed," and telling developers not to substitute a backend timer for end-to-end response time (OpenAI voice agents guide). On a demo call that translates directly. When the agent says it booked the appointment, nothing was booked. Say so in the moment. The sandbox disclosure on the page says so too, and a buyer who catches the gap himself after you glossed over it has learned something about you, not about the software.
When you should not run the demo call
Three cases, and pushing through them costs you the deal.
The phone is not their problem. Some businesses answer every call inside two rings because the owner's wife does it and does it well. A missed-call recovery pitch to that business is noise. Sell them something else or nothing.
They want to see their own data. The demo runs against a sandbox. No calendar, no CRM, no messaging system is connected, and that is deliberate so nothing touches their production systems before they have signed anything. A buyer who wants to watch a booking hit his real calendar is asking for the build, not the demo. Move him to pricing and timeline instead of trying to stretch the sandbox.
The policy is complicated. Multi-turn instruction following is still the weakest axis of these systems. On MultiChallenge, converted to audio, gpt-realtime scored 30.5 percent, up from 20.6 percent on the prior model (OpenAI). That is real progress on a hard benchmark and it is also a number under a third. If a prospect's intake has eleven conditional rules about insurance verification, do not demo all eleven. Demo three, and be honest that the rest gets built and tested during fulfillment rather than improvised on a sales call.
There is a fourth case that is not a failure. A warm referral who already trusts you does not need twenty minutes of proof. Ask what they want, quote it, and give the time back.
How to close on the call
Ask for the decision while the agent is still on their screen. "That is the receptionist. Yours would be answering your line by the end of next week. It is $2,400 a month. Do you want it?"
Note what is missing from that sentence. There is no build fee. You are not asking them to fund development before they have seen anything work, which is the biggest reason these deals stall. How the no-build-fee structure changes what you can charge is worth understanding before you quote, because the number you say out loud is yours to set.
If they say yes, they go into the client portal that day. If they say "send me something," you lost the call at minute four, when you explained instead of handing over the phone.
FAQ
How long should an AI demo call be? Twenty minutes, and under thirty in every case. The demo does the persuading, so extra time is you talking, which works against you. Book it as twenty so nobody is watching a clock.
Should I demo live on a call or just send the link? Send the link during a call, at least the first twenty times. You want to hear what they ask and where they hesitate, because that becomes your objection list for every future call in that trade. After that, sending the link cold starts to work, and you can watch opens and calls from the dashboard.
What happens if I hit my voice cap in the middle of a demo? The live parts pause and the prospect sees a neutral message saying the demo is not accepting calls right now. They never see a billing notice or your plan name, and there is no overage invoice. It is still a bad look, so track your minutes during heavy weeks.
What do I say when they ask which AI it uses? Tell them and move on. Nobody buying a receptionist cares about the model name, and a long answer signals that the technology is the product rather than the result. If they push, the answer to "why can't I just use ChatGPT" is the useful version of that conversation.
Do this before your next call
Open the live demo and run it on yourself as the buyer. Pick the trade you sell into most, call the receptionist, and ask the three questions your last prospect asked you. Write down where it impressed you and where it did not. That list is your demo call script, and it beats anything anyone else can write for you.
If you are still deciding, start here at $49 a month. Pricing goes up on October 1. And if you have not booked one of these calls yet, the first ten clients post covers how to get the meetings.
