Autonomous AI Voice Agent for RCM: 61,246 Patient Calls
How Limestone helped a US healthtech company deliver 420+ hours of patient conversations through AI, at $0.071 per connected call.
- Healthcare RCM
- AI Voice Agent
- Autonomous Agents
- Industries
- Healthcare, revenue cycle management
- Platform
- Autonomous patient-balance management across voice, SMS, email and mail
- Delivery scope
- AI voice agents, billing integrations, payments, reconciliation and production operations
ABOUT THE PROJECT
Project overview
- Engagement
August 2025 — ongoing
- Delivery team
1 senior AI engineer, plus advisory and delivery lead
- Core technologies
Python, FastAPI, PostgreSQL, Next.js, realtime AI voice agents
The client is a US healthcare technology company building autonomous agents to manage patient balances: explain the bill, follow up, arrange payment and reconcile it back to the provider’s system.
Limestone joined at discovery and architecture. Fifteen weeks later, the first production outreach went live. The first autonomous patient call followed three days later.
By September 2026, the platform was managing 25,338 balances for 11,939 patients, with voice, payments and the financial controls underneath them running in production.
The Challenge
Make patient follow-up economical, without compromising the patient or the books.
-
Small balances. Expensive follow-up.
Billing teams worked patient lists manually, or stopped after the first statement. Every question added another call, callback or handoff. Consistent phone outreach was difficult to justify on small balances.
-
Every billing system worked differently.
Patient context was spread across billing platforms, insurance responses and payment records. Some systems offered APIs; others offered report exports. The agent needed a reliable, complete picture.
-
Real money left no room for approximation.
A successful conversation had to become a correctly allocated payment. Retries could not create duplicate actions. The processor, internal ledger and provider’s billing system all had to agree.
-
Patient trust was part of the product.
The agent spoke on the provider’s behalf in a HIPAA-scoped environment. It had to verify identity, respect consent and quiet hours, and stop when a dispute or request for a human required staff attention.
WHY LIMESTONE
One team accountable for the whole system.
The client needed an engineering partner that could take the product from architecture to live patient operations. An AI voice agent was only one part of the work. Limestone’s scope also covered billing integrations, payments, reconciliation, safeguards and the operations console. We built alongside the client’s technical leadership and kept shipping through a shared repository, protected branches and mandatory quality gates.
WHAT WE BUILT
Solution
The AI voice agent, and the production system underneath it. Autonomous patient conversations connected to billing data, payments and reconciliation, with explicit rules for when a human takes over.
-
Autonomous AI Voice Agents
Outbound billing calls, inbound answers and a dedicated scheduling line. Agents verify identity, explain balances, send payment links and arrange payment plans during the conversation.
-
One Case, Every Channel
Calls, texts and emails use the same patient context: charges, insurance, payments and conversation history. A returning patient can continue where the last conversation ended.
-
Billing System Integrations
REST, FHIR, SOAP and file-based ingestion behind one interface. Malformed records are quarantined, discrepancies are surfaced, and agent notes write back to the provider’s chart.
-
Payments & Reconciliation
Payment links, installments and processor routing connect to exact-decimal accounting and a double-entry ledger. Payments allocate to charge lines; drift checks flag disagreement.
-
Outreach That Adapts
Agents regenerate follow-up when a balance changes. The system cancelled 658,657 scheduled actions before sending, overwhelmingly after a payment, dispute or other change in state.
-
Closed-Loop Self-Improvement
Three AI judges review voice, messaging and patient outcomes, proposing prompt, code and architecture changes for human approval. A separate judge checks new conversations to verify that fixes worked.
INSIDE A HIPAA-SCOPED PRODUCTION ENVIRONMENT
Autonomous where it can act. Human intervention where it must stop.
The AI voice agent verifies patient identity, respects consent and calling hours, and prevents duplicate charges or bookings. Disputes, withdrawn consent, legal concerns or requests for a person pause outreach and bring in staff with the full case history. Call limits and failure alerts keep operations controlled, while testing stays separate from real patient records. Patient information is handled under Business Associate Agreements, with an audit trail and no health information in technical logs.
DELIVERY TIMELINE
Fifteen weeks to production outreach. Then autonomous patient calls.
Discovery began in August 2025. First outreach went live on 4 December; the first autonomous call followed on 7 December. The platform has continued shipping since.
-
AUG 2025
Discovery & Architecture
Define the agent, integration and financial foundations.
-
4-7 DEC 2025
First Live Outreach & Voice
Production messages, then the first autonomous patient call.
-
JAN 2026
Commercial Launch
Inbound voice answers patients when they call back.
-
JUL 2026
Scheduling & Cost Reduction
Dedicated scheduling agent; average messaging cost cut 18%.
-
SEP 2026
61,246 Voice Calls
420 hours of patient conversation; 25,338 balances managed.
HOW IT STAYED TRUSTWORTHY
Every action leaves evidence.
-
FINANCIAL CORRECTNESS
The books have to agree.
Exact-decimal money handling, charge-line allocation and double-entry ledgers sit beneath the agent. Drift detectors flag mismatches. Corrective jobs refuse to run until reconciliation checks pass.
-
PATIENT PROTECTION
Escalation is built into the workflow.
The agent verifies identity before disclosure and stops for disputes, consent changes or requests for a person. Staff receive the full history. Consent, disclosure and operational actions are recorded.
-
CONTINUOUS AI EVALUATION
The system reviews what it said.
AI audits complete communication chains and proposes prompt, code and architecture changes. Humans triage the proposals. A separate judge checks post-change conversations to verify fixes.
-
AUTOMATED TESTS PROTECTING PRODUCTION CHANGES
13,650
-
PRODUCTION AUDIT-LOG RECORDS
6.36M
-
IMPROVEMENT PROPOSALS GENERATED FOR HUMAN TRIAGE
261
Impact & Results
Patient conversations handled. Operating costs reduced. Production outcomes measured.
- 61,246 autonomous AI voice calls, including 21,767 connected conversations and 420 hours of talk time. The platform handles outbound billing calls, inbound questions and appointment scheduling.
- Around 97% of patient conversations complete without human involvement across voice and messaging. Exceptions pause outreach and route to staff with the case history attached.
- $0.071 recorded average telephony and model cost per connected call.
- 18% lower average cost per message after one engineering change. Character sanitization and shorter SMS templates reduced message segmentation, measured across 41,278 messages before and 38,316 after.
- 18,194 balances reached zero; 73% of resolved balances closed within 14 days. Zero-balance outcomes include insurance re-adjudication and other adjustments, as well as patient payments.
- Sixfold growth in monthly communications, from 8,514 to 51,927. Over the reported period, the scheduled-job system recorded 955,633 runs and 10 failures.
-
Patient conversation delivered by AI across 21,767 connected calls
420 hrs
-
Recorded telephony and model inference cost per connected call
$0.071
-
Reduction in average messaging cost, before and after July 2026
18%
