A healthcare billing company needed a HIPAA-compliant voice agent that could call patients about their bills, explain every charge, and take the payment before the call ended. Since December: 61,246 calls, 420 hours of patient conversation, and around 97% of patient conversations across channels completed without human involvement. A connected conversation costs about seven cents in telephony and model inference, before infrastructure, payment processing and people. The patient’s share of the bill was the problem. Statements listed codes and amounts without explaining them, and providers didn’t have enough staff to follow up on every balance. Fifteen weeks after kickoff, the first patient outreach went out. The first AI call followed three days later. Before an agent could explain a bill, we had to assemble the facts behind it: the visit, what insurance covered and why, every payment, every adjustment and every previous conversation. I’ve watched teams skip this step. They connect a voice model to a script, get a convincing demo, and discover it cannot answer the question patients actually ask: why do I owe this? We built shared context across voice, text and email so each channel works from the same case. When a patient calls about last week’s text, the agent already knows what it said. Twelve live tools let the agent look up payment information, send a payment link, arrange a payment plan, generate a statement and manage appointments while the patient is still on the phone. Every action follows rules enforced in code. Verify identity before disclosure. Prevent duplicate charges and bookings when an action retries. Check quiet hours at dispatch. Pause outreach and involve staff when someone disputes a bill, withdraws consent or asks for a person. The outreach plan also changes when the patient’s situation changes. Of 949,563 scheduled outreach actions, the system cancelled 658,657 before they went out, mostly because the patient paid, disputed or changed status. Paying a bill should cancel the reminders still sitting in the queue. Then we built the self-improvement loop. An audit model reviews complete patient communication threads for unanswered questions, inappropriate responses and missed escalations. It has reviewed 5,135 threads and logged 835 findings. Three AI judges examine voice, messaging and patient outcomes, then propose changes to prompts, code or architecture. A human reviews every proposal before anything ships. Each new analysis receives previous proposals with the human decisions attached, including why suggestions were rejected. After a change is marked fixed, a separate model checks fresh conversations to see whether the problem actually went away. The team can trace an issue from the original conversation through to evidence that the change helped. The client’s technical founder ships through the same repository, branch protection and six CI gates as our team. The platform has stayed live since launch. That's what our Velocity Framework was built for.