Voice agent Vapi + YClients: 24/7 booking without an administrator
A chain of three premium salons in Moscow was losing up to 40% of calls in the evenings and on weekends. A night-shift administrator at ₽95k/mo didn't solve the problem. A voice agent on Vapi + GigaChat with direct YClients booking delivered 97% pickup and brought +₽278k revenue per location.
A 21:30 call goes nowhere
Premium beauty segment works on a simple rule: the client calls and books when they think of it — during lunch, on the way home, in bed before sleep. If no one answers at that moment — they WhatsApp a competitor and never come back.
The three salons' administrators handled incoming calls 10:00-21:00. After 21:00 and on weekends calls went to voicemail no one checked in the morning — by the time they called back, the client had already booked elsewhere.
An attempt to hire a night-shift staff for evenings and weekends at ₽95,000/mo per location didn't solve it: the duty staff didn't know the masters' specifics, got confused with rooms, missed calls during bathroom breaks or lunch. Pickup rate still hovered around 64% across working and non-working hours.
Voice agent on Vapi + GigaChat → direct booking to YClients
Vapi answers calls 24/7. GigaChat 2 Max is the dialogue brain, knows services, masters, rooms, prices, and the schedule of three locations. FastAPI middleware converts agent intents into YClients REST API calls: check open slots, book, reschedule, cancel, send SMS confirmation via SMSC.ru.
Replacement number behind the main, Yandex SpeechKit for STT/TTS in Russian
14 scenarios: new booking, reschedule, cancel, pricing, master consultation
Agent tool calls → YClients REST: slots, bookings, client, history
Direct booking to the right room at the right location, conflict checks
SMSC.ru confirmation to client; confidence < 0.7 → switch to admin
14 scenarios instead of one script
A beauty voice bot isn't "book a haircut". It's a builder: "I want Katya for manicure, but no earlier than Thursday evening", "can I move Tuesday's appointment to Friday", "how much is coloring with a top master", "any open slots for facial massage this week". GigaChat 2 Max confidently handles each of these scenarios because prompt engineering was tailored to the specific YClients account: services, masters, durations, rooms, prices — everything is loaded from the API into the model's system context at session start.
Fallback to administrator on low confidence
If the agent fails to understand the request after 2 clarifications — the dialogue immediately switches to a live administrator (daytime) or goes into the callback queue (night). This is not "a bad booking is better than none": in premium segment a bad experience costs more than a lost booking. The confidence threshold was tuned iteratively over the first two weeks of launch.
YClients tool calls as first-class operations
FastAPI middleware doesn't "make the request for the agent" — it exposes the agent a set of tools:
find_slots,
create_record,
reschedule,
get_master_info.
The agent decides which to call in what order. Middleware validates parameters
and proxies to YClients REST with proper authorization.
SMS confirmation via SMSC.ru
After a successful booking the client gets an SMS with a link to the YClients booking card — they can view/reschedule/cancel via the standard salon interface, as if an administrator had booked it.
Russian providers + predictable hosting
Voice platform: number provisioning, ASR/TTS pipeline, WebSocket to LLM
LLM dialogue brain, tool-calling, supports 14-scenario context
STT/TTS in Russian — premium voices, not 'robot assistant'
Booking/reschedule/cancel directly into salon rooms
Tool server: validation, authorization, routing across 3 locations
SMS confirmation to client with link to YClients booking card
Middleware hosting in RF — FZ-152 compliant
Dialogue logs and quality metrics for iterative prompt tuning
What changed in numbers
across all incoming, including 21:00-10:00 and weekends
per month — bookings outside business hours, 41 on average
recurring per month: Vapi minutes + GigaChat tokens + Selectel
The headline number for the owner — +41 bookings per month per location from "off-hours" time. These aren't clients taken from the day administrators, it's a brand-new stream — those who used to drop after 5 rings.
Recurring cost dropped 12×. The day administrator continues to work in the salon — they're no longer a "phone dispatcher" but a reception consultant who greets clients in person.
Where else the same architecture fits
Universal pattern — "Voice + LLM + CRM API + human fallback". The same architecture (just swap the CRM integration nodes) fits for:
- → Barbershops, nail studios, massage rooms on YClients/Altegio/Beauty Pro
- → Fitness clubs and training studios — class booking, cancel, membership rollover
- → Auto services with booking into Alpha-Auto or own CRM
- → Language schools, tutors, early-childhood centers — trial-lesson booking
- → Delivery / cleaning services — handling orders and reschedules outside business hours
- FastAPI middleware scaffold with tool-call API and dialogue logging
- Prompt engineering for GigaChat with service-catalog loading from CRM into system context
- Confidence-fallback to administrator by two criteria: ASR quality and intent confidence
- Selectel middleware hosting with proper PII storage schema for FZ-152
If you have a CRM with a schedule and missed calls — a voice agent will close them
Vapi + GigaChat + your CRM — 4-6 weeks from kickoff to production. Payback typically fits in 2-3 months from freed administrator hours and intercepted bookings outside business hours.
Аудит за 5 000 ₽ — с конкретным отчётом и сметой
Расскажу что внедрить в вашем бизнесе в первую очередь, какая будет окупаемость, и нужен ли вообще AI для вашей задачи (иногда — нет).
Или просто напишите свой вопрос — отвечу в течение 2 часов