Running ElevenLabs voice agents in production

I have been running conversational AI agents on ElevenLabs in production at SellAnyCar, making automated marketing and outreach calls to customers through our telephony system. Here is what actually matters when you take a voice agent from a demo to a deployed system.

Per-campaign design is where quality is won

The same prompt and conversation flow that works for one campaign fails for another. Every campaign gets its own agent design, tailored to the scenario. The difference in outcome is huge, and it is the part of the work that looks small and is not.

A bad call should become a fix, not a lost customer

When you deploy agents at scale, some calls go badly. The question is whether you find out. I run a monitoring loop over the automated calls, and the underperformers get reworked. This is the same discipline as the chat system. Watch the numbers daily, find the failures, fix the prompts and flows, redeploy.

What close to human actually takes

Getting a voice agent to perform close to a human on the same calls took iteration, not a magic prompt. Every round of review and rework moved the results a step closer. When the business saw the quality hold up, that is what convinced them to scale the agents up.

If you are building voice agents, plan for the loop from day one. The model does the talking. The system around it decides whether anyone keeps listening.