AI Booking Assistant
Project Field
AI Development & Product Design
Project
For Local Service Businesses
Small local service businesses lose real customers every time a call goes unanswered. Most have no way to capture that interest before it goes to a competitor.
Case Study
● ● ● ● What I Built, and Why It Changed
I started with voice. When a business missed a call, an AI agent would answer, understand what the caller wanted, and collect booking details — service type, date, time — before handing off to staff for confirmation.
The voice assistant was tuned after extensive testing: Deepgram Nova-3 for Swedish transcription, GPT-4o Mini for reasoning, ElevenLabs' multilingual model chosen specifically for pronunciation accuracy over raw speed. I tested it deliberately — ambiguous requests, mid-conversation corrections, background noise, unusual phrasing.
● ● ● ● Where Voice Failed
The AI itself worked reasonably well. The surrounding infrastructure didn't. Carrier limitations made Swedish numbers with reliable voice capability difficult to source. Keypad input was inconsistent. End-to-end latency sat around 1,670ms — enough to make conversations feel unnatural.
Rather than keep tuning a system fighting its own foundation, I rebuilt the same core value as an SMS flow.
● ● ● ● The SMS Version
The conversation logic — understanding what a customer wants and extracting the right details — carried over almost directly. What changed was the delivery mechanism.
A missed call triggers an SMS to the customer. Claude reads the reply, asks natural follow-up questions, and extracts service type, date, and time. Once complete, it confirms and hands off a structured booking request to staff. Built on a real Swedish number via 46elks, tested end-to-end.
● ● ● ● What This Actually Demonstrates
Building the right thing usually means being willing to abandon the first version once the evidence says so. This project is as much about that decision as the AI itself — recognising that a well-built assistant on the wrong infrastructure isn't a small tuning problem. It's the wrong foundation, and rebuilding is sometimes faster and better than fixing.