5 things I learned building an AI receptionist that the tutorials skip

Building the thing with Retell or ElevenLabs + Twilio is honestly the easy part. What actually got me was all the operational stuff nobody mentions. Sharing in case it saves someone a headache 🙂

Latency is the silent killer. People hang up if the bot pauses more than about 2 seconds, so get the voice → backend → response loop tight before anything else. Every extra API call is a chance to lose the caller.
Trial phone numbers only work with verified numbers. My demo line worked perfectly for me and nobody else, and I lost a whole day before realizing I had to upgrade the Twilio account. Test with a phone that isn't yours before you show anyone.
A phone line can't depend on your laptop. Put the endpoints somewhere always-on (Vercel/Supabase, or Railway), and make the booking idempotent, because retries happen and double-bookings are the worst.
Say the consent and call-recording line in the opening message. It's required in a lot of places, Canada included, and honestly it sounds more professional, not less.
The part I keep thinking about most isn't the tech, it's proof. A dead webhook doesn't always fail loudly, sometimes it just goes quiet and the owner only finds out from an annoyed customer. So every call should leave a clear trace, and I think the real unlock is a dead-simple weekly recap to the owner: X calls handled, Y booked, Z after-hours caught. Not a dashboard, just the outcome in plain words. (Haven't battle-tested that last one across a bunch of clients, just where my head's at.)
French support exists but isn't great yet, which is mildly annoying since I'm in Quebec 😅

Curious for anyone who's built voice stuff — what surprised you most: latency, reliability, or something else?

0 comments

5 things I learned building an AI receptionist that the tutorials skip