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 🙂 1. 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. 2. 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. 3. 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. 4. 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. 5. 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.) 6. 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?

Something I keep running into with agency clients and I don't have a good answer for it. Client spots something in their CRM or inbox, asks me what the automation did and whether it was meant to do that. I go dig through execution history, try to turn it into something they can actually follow, usually end up screenshotting it into an email. Slow, and it never looks as reassuring as I want it to. It's got harder since I started putting AI steps in, because now there are decisions in there I can't fully predict myself. So I'm curious how people here handle it. Do your clients ask for this kind of thing, or do they mostly not care as long as the thing works? And if they do ask — is anyone sending them something better than raw logs? The part I find hardest is when a client's setup spans a few tools at once. No single place to point them at. Might just be that I've got nosy clients. Would be good to know either way.