The AI chatbot booking engine behind Peak AI’s med spa platform

Why we built an AI chatbot for med spas instead of buying one

Most chatbot tools sold to clinics are generic. They answer questions using whatever the model already knows, they hand off to a human the moment someone asks about pricing, and they were never built to handle the awkward reality of appointment booking: real calendars, real overlapping requests, real medical questions that should never come from a bot. We built our own platform because the booking chatbots we looked at solved the easy 80 percent and left clinics exposed on the hard 20 percent.

An appointment booking chatbot that actually books the appointment

The core of the platform is a booking engine that handles the full appointment flow inside the chat window itself. A visitor can ask about a treatment, check availability, and book a slot without leaving the conversation. State for each booking session sits in Redis, which gives us fast reads and writes without hammering the primary database every time someone types a message. The harder problem was concurrency: what happens when two visitors try to grab the same 2pm slot within the same second. We built the booking logic to be race-safe and tested it under concurrent load, so a slot can only ever be claimed once, no matter how close together the requests arrive.

Why grounding matters

A chatbot that improvises is a liability for a med spa. If a model is left to answer from general training data, it will eventually invent a treatment the clinic does not offer or quote a price that is wrong. We solved this with retrieval: every answer the chatbot gives is grounded in the clinic’s own website content, pulled at query time rather than memorised in advance. If the answer is not on the site, the bot does not guess. This is the difference between a chatbot that sounds confident and one that is actually accurate, and for a medical aesthetics business, accuracy is the only version worth shipping.

Four providers, so one outage does not take the booking flow down

We do not route every conversation through a single model provider. The platform supports OpenAI and Anthropic directly, plus routing through OpenRouter, with failover built in across all four provider paths. If one provider is slow or down, the conversation moves to the next without the visitor noticing anything changed. For a clinic that depends on the chatbot to capture bookings around the clock, a single point of failure at the model layer is not an acceptable design.

Escalating what should never be automated

Medical questions do not belong to a chatbot. We built escalation rules that hand off any medical question to a human member of staff rather than letting the model attempt an answer. The bot is there to book appointments and answer questions the clinic has already published, not to practise medicine.

Proving it on a real site, with a real booking

We did not stop at internal testing. The widget was installed on a real clinic site, embedded cross-origin the same way any client would run it, and a real appointment was booked end to end through the chat window, from first message to confirmed slot. That is the version of the platform we now sell as Peak AI’s chatbot service, from $350 a month.