I build the automation a business actually runs on.

Most voice AI is a booking bot for a salon. This one sits inside health insurance, where an agent that says the wrong thing about coverage is not a bad customer experience, it is a compliance problem. Everything it says has to hold up against AI disclosure requirements and CMS marketing rules, and it cannot improvise about plans, prices or eligibility.

I built two. The first works out what a caller actually needs and routes them, built for high inbound volume. The second is an intake line for people who just lost their health insurance along with their job. It qualifies the caller and warm transfers to a licensed human rather than trying to solve it.

I scoped both agents by sitting with non-technical executives, built the flows, presented them back, and coordinated across departments to ship. Nobody handed me a spec.

Building a voice agent is the easy part. Keeping one alive on someone’s real phone line, month after month, is the part most people never reach. This one answers every inbound call, looks up orders against Shopify, handles what it can and escalates what it cannot.

The number that matters is 89%. Their showroom is open Monday to Friday, ten to five, and almost nine in ten calls land outside that window, a lot of them at eleven at night and on weekends. Every one of those used to be a voicemail. They are now qualified, logged, and waiting on someone’s desk in the morning, on enquiries for pieces running from $21,000 to over $300,000.

The part I care about is the loop at the end. Every week I listen to real conversations, find where it handled something badly, and fix it. That is the difference between a demo and a system.

The client needed to reach a narrow technical audience where the usual volume approach fails, because the list is small and the wrong message burns it. So the system spends its effort on targeting and enrichment instead of send volume.

Apify sources the list, n8n orchestrates, Clay enriches each prospect enough that the message can say something true and specific about them, and replies feed back so the targeting improves rather than repeats.