{"slug": "the-leak-nobody-measures-what-we-learned-putting-ai-receptionists-on-3-dental", "title": "The Leak Nobody Measures: What We Learned Putting AI Receptionists on 3 Dental Clinics' Phone Lines", "summary": "A developer at DenixLabs deployed an AI voice assistant and chatbot across three dental clinics in Sofia, handling over 1,000 real patient calls and booking roughly 20 to 30 additional hours per clinic each month, mostly from calls that previously went unanswered. The system integrates with each clinic's existing phone number and schedule, using clinic-specific pricing, service durations, and approved phrasing, while routing complex calls to staff with context attached. The developer reported that no receptionists were laid off and that all three clinics saw fewer missed calls in their monthly reports.", "body_md": "A few weeks ago I gave a 20 minute talk at a conference for dental clinic managers in Sofia. Nobody in the room writes code, so I talked about patients and phones, not prompts and latency. I figured the same material might be useful to people here who are building voice agents for small businesses, so this is the developer version of that talk. It's what we shipped, the numbers after three clinics and a bit over a thousand real patient calls, and a few things I'd do differently.\n\nClinics track everything. Revenue, chair time, no-shows, how many implants this quarter. What they don't track is the missed call, because there's nowhere for it to show up. It's not in the schedule and it's not in the monthly report. The practice software has no idea it happened.\n\nAnd the call that gets missed is usually the expensive one. Someone with a toothache doesn't leave a message and wait. They hang up and dial the next clinic on their list. At that point they have no loyalty to anyone, they're just picking whoever answers. The calls get lost exactly when the front desk is busiest, because the receptionist is standing there with a patient in front of her and the chairs are full. And then there's lunch, everything after 18:00, and Saturday mornings, when nobody is on the phone at all. Nobody is at fault here. One person can't be in three places.\n\nI spent almost a third of the talk on this. In hindsight that was the right call. Once an owner accepts that the leak is real and that they literally cannot see it, the rest of the conversation is easy. If you're selling any kind of automation into a small business, the invisible cost is the pitch. The feature list is not.\n\nDenixLabs is a voice assistant and a chatbot that sit on the clinic's existing phone number. I describe it as a communication layer rather than a bot, because that's closer to what it does: it picks up, works out what the patient wants, and either books straight into the schedule or hands the call to a person with the context attached.\n\nThe generic \"dental assistant\" persona is useless. What makes it work is that the assistant knows this specific clinic. It knows every specialist's schedule, which services each of them offers, and how long each service takes, because a cleaning and a root canal don't get the same slot. Prices come from a list the clinic approved, nothing else. The phrasing and tone are the clinic's own, and the owner signs off every phrase before we go live.\n\nDuring the talk I played a 90 second recording of a real patient call, no editing. By the end of it the patient has an appointment with the right specialist, in a slot with the right duration for what she asked for. That clip did more than every slide before it. If you're building in this space, get an unedited real demo as early as you possibly can. Nobody believes the mockup.\n\nThree clinics have the assistant on their phone line every day. Not a pilot, not a sandbox, real patients. We've handled a bit over 1000 patient conversations. Each clinic is getting roughly 20 to 30 extra booked hours a month, most of it from calls that used to ring out. Coverage is 24/7, so evenings, weekends and lunch are covered.\n\nI don't put a revenue figure on that slide, on purpose. Every owner in the room knows their own hourly rate better than I do, and they do the maths in their head faster than I could present it.\n\nNobody got fired\n\nThis was the thing everyone was thinking and nobody said until the Q&A. Is this here to replace my receptionist?\n\nWhat actually happened in the three clinics: the teams stayed, nobody lost their job. What changed is that the receptionist no longer has to choose between the patient at the desk and the phone ringing in her other hand. And at the end of the month, all three clinics have fewer missed calls in their reports.\n\nIf I were doing the talk again I'd put this before the numbers. In my experience the thing that blocks adoption is not the owner being sceptical, it's the staff being scared, and the owner picks up on that.\n\nThese clinics have no IT department. Onboarding has to be a conversation, not an integration project, so it looks like this. One call where we go through the schedule, the services and the durations. We configure the assistant with their phrases, their prices and their tone. The owner listens to the voice and approves every phrase, and if they don't like it we don't launch. Then we go live in parallel: the AI only takes the calls nobody picked up. Two weeks later we sit down, look at the numbers together, and the clinic decides.\n\nThat parallel launch is the one design decision I'd defend the hardest. The first two weeks are always the roughest, least tuned version of the assistant, and having it only catch unanswered calls takes almost all the risk out of that period. The real value shows up from tuning on real conversations after launch, and you can't tune something you were too nervous to launch.\n\nThere's a catch I learned the hard way, though. If a clinic stays in that backup mode forever and the receptionist keeps answering nearly everything, the assistant barely sees any traffic and the economic effect looks weak. One of our clinics ran like this for months and the numbers were unimpressive for exactly that reason. Backup mode is a launch strategy. It's not where you want to end up.\n\nThis section got more nods than anything else, and almost none of it is about model capability. It's product decisions.\n\nIt does not give diagnoses or treatment advice. Any medical question gets handed to a human, no exceptions. Prices are read from the approved list; the assistant never makes one up. Every booking is visible to reception immediately, and every conversation is kept as a text transcript. We don't store audio. Each clinic has a data processing agreement with us and can export their data whenever they want.\n\nIn a regulated vertical the list of things your agent won't do is a feature. Present it as one.\n\nThe eight questions I got\n\nIn roughly the order they came up. Swap the nouns and you'll get the same list in any small business vertical:\n\nAnd finally, you're a small company, what happens if you're gone tomorrow? The last two are the ones founders don't prepare for. Have real answers. Including for the case where the honest answer is that it doesn't make sense for a clinic that size, which I said out loud to at least one person that day.\n\nThe same layer is now running as a reservation agent for a hotel, integrated with Clock PMS+ for availability, bookings and guest profiles. We're also doing outbound now, not just inbound. What won't change is the approach: one vertical at a time, integrated with whatever software that vertical actually runs on, built all the way through. I have no interest in a generic assistant that does a little bit of everything.\n\nIf you're building something similar, or you work at a clinic software or PMS vendor and want to talk integrations, I'm at [denixlabs.com](https://denixlabs.com/en). Check out our product page for dental clinics at [denixlabs.com/en/industrii/dentalni-kliniki](https://denixlabs.com/en/industrii/dentalni-kliniki). Happy to compare notes in the comments.", "url": "https://wpnews.pro/news/the-leak-nobody-measures-what-we-learned-putting-ai-receptionists-on-3-dental", "canonical_source": "https://dev.to/denis_zahariev_9e82025ca2/the-leak-nobody-measures-what-we-learned-putting-ai-receptionists-on-3-dental-clinics-phone-lines-3mk6", "published_at": "2026-09-17 18:11:04+00:00", "updated_at": "2026-09-17 18:22:54.877073+00:00", "lang": "en", "topics": ["ai-agents", "ai-products", "natural-language-processing", "ai-tools"], "entities": ["DenixLabs", "Sofia"], "alternates": {"html": "https://wpnews.pro/news/the-leak-nobody-measures-what-we-learned-putting-ai-receptionists-on-3-dental", "markdown": "https://wpnews.pro/news/the-leak-nobody-measures-what-we-learned-putting-ai-receptionists-on-3-dental.md", "text": "https://wpnews.pro/news/the-leak-nobody-measures-what-we-learned-putting-ai-receptionists-on-3-dental.txt", "jsonld": "https://wpnews.pro/news/the-leak-nobody-measures-what-we-learned-putting-ai-receptionists-on-3-dental.jsonld"}}