A 4-chair dental clinic in Kyiv: +38% bookings in 2 months

The clinic has been on the market for 6 years, located in Pechersk, Kyiv, specializing in therapeutic and prosthetic dentistry. 4 chairs, 6 doctors (2 of them carry the main flow), one administrator in the lounge and one on the admin chat. Before MTDK ai there was no booking automation at all — one person held it all together.

Acquisition channels: phone (Kyivstar + Vodafone), Instagram (~12k followers), a Telegram channel (~3k) and a website form. One person ran all of these from 9 to 19, and in the evenings doctors sometimes replied in Direct personally — hence double bookings, forgotten promises and scheduling errors.

An important detail: the clinic used Helsi only for e-declarations, not for managing bookings. So we didn't change the core system — we built an AI administrator for the clinic on top of the existing one. It isn't yet another "dental practice software" or a separate CRM you have to roll out from scratch — it's an automation layer over what already works. This matters for clinics afraid to "break what works": automating patient booking does not require replacing your medical system.

The baseline we recorded before launch, to have something to compare against: ~590 bookings per month, median response time 47 minutes, chair utilization 67%, cancellation rate around 24%.

Dentistry is a tough niche for bookings: a high average ticket, a long treatment cycle, pronounced seasonality (peaks in September and January) and patients who are very sensitive to response speed. Here every missed inquiry is expensive, so automating patient booking delivers a fast and visibly monetary effect.

All three pains had a single root: the process rested on one person during working hours, while patients reached out whenever it suited them — mostly in the evenings and on weekends.

Pain one — evening and weekend inquiries. Patients would write at 21:00 "Any free slots tomorrow at 18:00?" and wait until morning. Up to 40% of such inquiries ended with "didn't wait, booked elsewhere". The owner estimated 80–100 lost patients per month — many of them already paid for via Instagram ads.

Pain two — routine ate the administrator. "How much is a cleaning?", "Do you take cards?", "Can I bring a child?" — about 60% of all conversations are repetitive questions. Instead of working with patients in the lounge, the person sat in chats answering the same things on a loop.

Pain three — broken follow-up. After the first consultation a patient needs a call back in 3–7 days to continue treatment. There was no time to do it manually, so conversion into repeat visits was low — and repeat visits are exactly what drives the margin in dentistry.

The heart of the solution is an AI administrator built on Claude with a custom prompt tailored to this clinic's services. The knowledge base: the price list (47 items), the doctors' schedule (two-way sync with Google Calendar), 32 common questions with the admin's word-for-word answers, and medical ethics rules — the list of topics the AI strictly does NOT discuss (symptoms, diagnoses, treatment advice).

Channels: a Telegram bot, Instagram Direct via the official Meta API, and Viber for reminders. The AI assistant replies instantly across all three. If a request is complex — a complaint, pediatric therapy, adult orthodontics — the AI escalates it to the administrator with the full conversation context, so the human doesn't have to re-ask the patient.

Integrations: Helsi via the official partner API for patient sync, Google Calendar for the doctors' schedule, and WayForPay for online deposits on expensive procedures (used by ~8% of bookings).

A separate lever is reminders and follow-up. The bot automatically reminds about a visit one day and 2 hours before (via Viber, since it's free), and 5 days after the first visit it sends a personal message offering to continue treatment. This exact scenario produced the biggest lift in repeat visits and effectively turned the dental chatbot into a patient-retention manager.

We also defined the tone of voice separately — friendly but professional, without excessive emojis or bureaucratic phrasing, so that in the chat the patient feels the clinic's care rather than "a robot". In a medical niche this matters: trust starts with the very first message.

Dental automation is best explained not by a description but by a real dialog. Here's a typical evening inquiry in Telegram — the kind that used to go unanswered until morning, and now closes into a booking in 4 seconds with no human involved:

No "an operator will reply shortly". The patient immediately gets concrete slots, picks a convenient one and lands in the doctor's schedule — and the clinic doesn't lose them to a competitor who simply replied faster.

Days 1–3. Data gathering: prices, schedules, FAQ. An interview with the administrator — "which questions do you answer 10 times a day, and what exactly do you say, word for word". This is the foundation of the AI prompt.

Days 4–7. Setting up the AI, integrating with Helsi and Google Calendar. Testing in an internal Telegram chat — 50 dialogs, fixing every error and tuning the prompt.

Days 8–10. Launch on the Telegram channel — the lowest risk, because these are "our" loyal clients who forgive small awkwardness. We monitor every dialog in parallel.

Days 11–14. If stable — we connect Instagram Direct. This is the largest channel and a mistake here is costly, so for the first 3 days we run in "AI + human sees every dialog" mode.

Day 14+. Full autonomy. The administrator moves to lounge work and complex cases, and the human admin chat shrinks to 4 hours a day — for VIPs and non-standard situations.

Two months after the full launch we tallied up the numbers. These aren't marketing roundings — they're data from Helsi, Google Calendar and the bot statistics:

Bookings per month: 590 → 814 (+38%). The main growth came precisely from evening (19:00–23:00) and weekend inquiries that used to be lost.

Response time dropped from 47 minutes to 4 seconds — a turning-point number for conversion: the client hasn't had time to leave for a competitor. Cancellations fell by 52% thanks to reminders one day and 2 hours before, plus a UAH 200–500 deposit on expensive procedures.

Chair utilization grew from 67% to 89% — not a cosmetic metric but money that simply wasn't being billed before. And the administrator's hours in chat shrank from 12 to 4 a day: the person moved to half-time in chat and full-time in the lounge.

We measured the evening segment separately: before launch only ~6% of bookings closed after 19:00, afterward — 22%. In other words, the time that used to be "dead" became the clinic's third-largest revenue channel. That's exactly where the bulk of the +38% is hidden.

Repeat visits grew by +24% — and that may be the most important number in the case: continued treatment is what drives the margin in dentistry. The automatic 5-day follow-up after the first visit — previously done by hand and often forgotten — did the work.

To be fair: the AI didn't fully replace the human — it took over the routine and the night shifts, while the administrator focused on what she does better than any algorithm. Here's how the key operational points changed before and after launch:

For the first 5 days the AI tried to give medical hints ("if it hurts, it's most likely a cavity"). In a medical niche that's unacceptable. We fixed the prompt with a hard rule: any question about a symptom → "a doctor needs to assess this, I can book you a consultation".

At first the AI dumped all free slots at once. A patient who asked for "any convenient one" got 47 options and got lost. We fixed it: the bot offers the 3 nearest slots, asks "does this work", and only shows more after a refusal.

The Helsi integration on the test stage caused a 30–40 second delay (requests were synchronous). We moved it to an async queue — now the AI replies instantly while Helsi sync runs in the background.

One doctor didn't want the bot to book to him without a prior phone consultation — his personal working style. We added a separate prompt rule for his schedule. The takeaway is simple: the AI administrator must adapt to the clinic, not the other way around.

First: an AI administrator works exactly as well as its knowledge base is good. 80% of success is not "AI magic" but a clear price list, an up-to-date doctors' schedule and word-for-word answers to common questions. Budget 2–3 days of prep before launch.

Second: don't launch automation on your biggest channel first. Start with Telegram or a "warm" audience, run the scenarios on real dialogs, and only then connect Instagram Direct, where the cost of a mistake is higher.

Third: define the boundaries clearly. In a medical niche the AI must not give treatment advice — only book, remind and answer organizational questions. Everything else is an instant escalation to a doctor. Boundaries are exactly what separate a reliable AI assistant from a risk to the clinic's reputation.

Now the main question of any owner — how much it costs and when it pays off. For this clinic we used the "Enterprise" plan — €200/mo (≈UAH 10,300) plus a one-time setup of €1500 (≈UAH 76,900): the Helsi integration is harder than average, so they took the top plan.

Additional revenue: +224 bookings per month × an average ticket of UAH 1,850 = +UAH 414,400 in monthly turnover. At a ~45% margin (including materials and the doctor's work) that's +UAH 186,480 net.

Payback of the full package (subscription + setup in the first month) — about 2 weeks. From the second month on, the monthly fee pays for itself in 1–2 working days of the clinic. For comparison: hiring a night-shift administrator would cost ~UAH 25,000–30,000/mo plus taxes — and that's just one channel, with no reminders, follow-up or analytics.

The owner's verbatim takeaway: "if we'd known earlier, we wouldn't have hired a second administrator a year ago". That second salary (~UAH 28,000) was effectively paying for work the AI administrator now does faster, with no weekends and cheaper. If inquiries get lost in your clinic in the evenings, it's not a staffing problem — it's a process problem. And it's solved in 14 days.