I want to write this while the lessons are still fresh. We have been running AI receptionists for paying clients for about six months now — dental practices, an HVAC company, two law firms, a personal-training studio, a med-spa, and three professional service firms. The data we have is real. Some of it surprised us. Some of it confirmed what we already suspected. One of it made us change how we sell.
What worked better than we expected
Bilingual coverage was the single biggest unlock. We assumed most callers in Miami would prefer English, and the AI would default there. In practice, when a Spanish-speaking caller hears a Spanish-speaking AI, the call goes on average 40% longer. The caller is more comfortable, asks more questions, and is more likely to book. The clients that turned on Spanish from day one saw 2x the booking rate of the clients that started with English only and added Spanish later.
The second surprise was insurance verification for dental. We thought it would be a small add-on. It turned out to be the entire reason some of the practices called us in the first place. The hold time with insurance companies is brutal, and an AI that can stay on hold for 40 minutes without losing patience has real value.
What we got wrong
We underestimated how much customization each vertical needs. A dental office and a personal injury law firm both have "intake" calls, but the questions are completely different. The first version of our system tried to use one generalist flow with vertical-specific prompt overlays. That did not work. The AI would confidently ask the wrong questions in the wrong order. We rebuilt it to be vertical-specific from the ground up, and the booking rate went up about 25% across the board.
We also over-promised on "sounds exactly like a human." It does not, and pretending it does is a setup for disappointed clients. What it does is sound polite, professional, and clear, which is what 80% of callers actually want. The 20% who want a specific human voice are usually the ones we route to a human anyway.
The numbers that actually matter
Across the clients we have data for, the average recovery rate on previously-lost calls is 64%. Meaning 64% of calls that would have gone to voicemail now get answered, qualified, and either booked or routed to a human who books them. For a typical small business doing 200 to 500 inbound calls a month, that is 60 to 150 additional conversations a month that did not exist before.
Of those recovered calls, the average close rate to a booked appointment or qualified lead is around 22%. So a 300-call-a-month business is now generating 13 to 33 additional booked appointments a month that it was not generating before. The dollar value depends entirely on the vertical, but even at a conservative $200 average customer value, that is $2,600 to $6,600 a month in new revenue against a $250 to $1,500 service fee.
What we changed about how we sell
We stopped saying "AI receptionist" in our first pitch. The phrase triggers a specific mental image (a robot on the phone) that does not match what we actually deliver. We say "an automated line that answers every call in under five seconds, books appointments, and only bothers you when there is a real human question." The technical mechanism does not matter to the buyer. The result matters.
We also stopped quoting on a phone call. The first call is a 15-minute audit where we look at their actual call logs and tell them how many calls they are losing and what those calls are worth. The number is usually higher than the owner expects. Once they see the real number, the price is easy.
What is next
We are starting to track what happens after the call is booked. Does the person actually show up? Do they convert to a customer? If they no-show, why? That is the next layer of the data, and it is where the real product improvements will come from. A 100% booking rate with a 30% show-up rate is not a win. A 60% booking rate with an 85% show-up rate is.
I will write about that in another six months.