Building an AI phone receptionist that actually answers the phone. Here's where it stands.
Lisa is a voice agent built for a small home services business that was losing calls after hours and during busy stretches. This is not a finished product story, it's a status update: what's live, what broke on the way, and what's still missing before this can run unsupervised.
Status: in progress. Voice works in both directions. The pieces that turn a phone call into a booked, followed up lead do not exist yet. Read this as a build in motion, not a case study.
2-wayVoice live
6Prompt versions, one session
60+Call scenarios wargamed
0Calendar behind her, so far
01The problem this is solving
A small home services business, the kind that quotes jobs over the phone, was missing calls constantly: after hours, mid-job, during busy periods. Every missed call is a lead that calls the next name on the list instead. The fix isn't a better hold message, it's someone who always picks up, day or night, and asks the right questions from the first sentence.
02What Lisa is, under the hood
A conversational voice agent, telephony on one side, a large language model doing the actual talking on the other. Nothing exotic in the stack, the hard part was never the technology, it was getting the agent to behave correctly on a live phone call, where a wrong answer costs a real customer, not a retry button.
Voice AITelephony integrationPrompt engineering
03Getting her call-ready took six rewrites in one sitting
The prompt went through six versions in a single session, each one graded against a real test call, not a hunch. Every rewrite fixed something a live call had just exposed: address collection, contact completeness, tone in awkward moments, what to do with a request outside scope.
The rules that had to get locked before this could go anywhere near a real customer: fixed business hours, sales calls handled live but anything operational gets a message and a callback, no live transfer to a human under any circumstance, a firm no on certain out-of-scope categories, a scripted and conservative approach to call recording consent. None of this is exciting. All of it is the difference between a demo and something you can point a real phone number at.
04The test that mattered more than any scripted one
During testing, an outbound call landed on a business stakeholder who wasn't expecting it and had no idea it was an early test. She didn't get an obviously robotic call, she got curious and asked to be called back later. That's a low bar in absolute terms and a real signal relative to where voice agents usually fail, awkward pauses, missed context, tone that reads as scripted the moment something unexpected happens.
05Where it actually stands right now
Voice is live both directions, inbound and outbound, proven on real test calls, not simulations. That is the part that's done. Everything past the phone call is not.
LiveInbound and outbound voice calling, both directions tested against real calls
MissingNo real calendar behind her yet, she can't actually book a slot end to end
MissingNo system of record for leads, calls aren't yet turning into tracked, followed-up leads
MissingNo follow-up or nurture layer, she is call-only right now, nothing chases a lead after the call ends
The honest framing: a voice bot that talks well is the easy 20%. The full receptionist is a spine, a lead store and an orchestrator, with the voice agent as one actuator on top of it. Right now the spine doesn't exist. That's the actual project.
06What's next, in order
Build order matters more than feature count here. Each piece below only gets built once the one before it is proven, not in parallel, because a half-built calendar behind a working voice agent is worse than no calendar at all.
NextSpeed to lead. The first slice: a lead store that catches every call and web form the moment it lands, and fires a callback within minutes. Research on this is blunt, a callback inside 5 minutes converts dramatically better than one the next day. This is the highest-leverage piece and it ships first on purpose.
ThenA real calendar. Lisa gets her own dedicated calendar account so a call can end in an actual booked slot, not a promise to follow up. Still deciding which booking tool survives contact with live, in-call scheduling.
ThenReminders and no-shows. Once bookings exist, the next failure mode is the booked slot nobody shows up for. Automated reminders, and a defined recovery path when someone no-shows anyway.
ThenFollow-up and nurture. Most calls don't book on the first try. This is the layer that keeps a lead warm instead of letting it go cold the moment the call ends, without a human having to remember to chase it.
OngoingWeekly call review. Every real call gets listened back against a scoring rubric, on top of the nightly automated eval loop, so prompt drift gets caught before it becomes a customer complaint.
Definition of done for this project: not "the voice sounds good," it's a caller who reaches Lisa at 11pm, books a real slot, gets reminded, and gets followed up on if they don't book at all, with zero human intervention anywhere in that chain. Voice was the visible 20%. This list is the other 80%.