I build AI phone receptionists for Australian small businesses - tradies, clinics, salons. Not demos, production systems that answer real customer calls. Here are the five things I got wrong before getting them right, in case it saves you the same tuition.
My first version was technically impressive and unusable: 2.5 seconds of silence between the caller finishing and the AI starting. Humans interpret silence as a broken line and hang up. Everything else - model quality, prompt cleverness - matters less than getting response time under a second. We ended up streaming transcription, starting the LLM call on partial transcripts, and pre-generating the first filler phrase. Unglamorous plumbing, completely decisive.
The demo question everyone asks is "how smart is the AI?" The production question that matters is "what happens when it doesn't know?" Early versions tried to answer everything, which meant confidently wrong answers to real customers. Now the system has a defined set of things it may do (answer FAQs, book appointments, take messages) and everything else routes to a human, fast, with a summary. Callers forgive a handover. They do not forgive being trapped with a bot that cannot help.
The AI part was maybe a fifth of the effort. The rest was calendars, CRMs, practice management systems, booking tools - every small business runs on a different stack, each with its own API quirks, auth flows and edge cases. A receptionist that can't actually write into the business's real booking system is a toy. Budget accordingly.
The honest benchmark data says current models complete about 12% of paid professional tasks to a client-payable standard. That matches production reality: AI is a strong junior assistant, not an unsupervised employee. So every action the receptionist takes is logged and checkable, bookings land in a calendar the owner already watches, and anything unusual flags a human. Design for review, not for trust.
Not one customer has bought "AI". They buy: the phone answered at 9pm, no more lost Saturday bookings, fewer interruptions on a job site. The technology is incidental to the sale and should be incidental to most of your design decisions too. If a simpler, dumber system would answer the phone reliably, use it.
I write more about practical AI for small businesses - what actually deploys versus what demos well - on our insights page.