When a dental clinic in the UK came to us with a simple problem, they were losing patients every weekend. We had 48 hours to fix it.
The problem was not complicated. Patients would call on a Saturday afternoon, find the line closed, and book with a competitor instead. The same pattern played out on evenings, bank holidays and lunch breaks. Each missed call was a missed booking, and each missed booking was money walked out the door.
We scoped Sofia in a single 90-minute call. The brief was clear: build something that could handle appointment enquiries on WhatsApp and the clinic website, 24 hours a day, without a human in the loop.
We built Sofia on a Python backend, integrated with Twilio for WhatsApp messaging and the clinic's existing booking system. The AI layer runs on Claude, trained on the clinic's services, pricing, staff availability and tone of voice. Sofia knows how to book, reschedule and cancel appointments. She answers FAQs about treatments, costs and opening hours. When a query falls outside her scope, like a clinical question that needs a dentist, she flags it for the team and follows up.
The build itself took just under 40 hours. Most of that was integration work, not the AI. Getting Twilio to handle inbound WhatsApp messages reliably, syncing with the booking calendar in real time, and testing every edge case, missed payments, double bookings, after-hours escalations, that is where the time went.
We deployed to Railway on a Friday afternoon. By Sunday evening, Sofia had handled three booking enquiries and converted two of them. The clinic owner sent us a message that night: Sofia paid for herself in the first week. We had three bookings on a Sunday evening that we would have lost.
What made Sofia work was not the technology. It was the setup. We spent time understanding how the clinic actually communicated, the words they used, the reassurances patients needed, the specific information that mattered. An AI receptionist trained generically would have failed. Sofia worked because she was built for one clinic, one audience, one job.
The lesson: AI does not need to be complex to be valuable. A narrow, well-trained system deployed fast will outperform a broad system built slowly every time.