What We’d Change: Lessons From Building an AI Assistant for Patient Billing Calls

Article author and evidence

Written by

Dr. Kainat Amjad, MBBS

Medical Doctor and Practice Growth Strategist, Staffingly, Inc.

Dr. Kainat Amjad is a medical doctor and practice growth strategist who writes about healthcare operations, patient-facing workflows, administrative burden, and practical technology adoption. Her work focuses on helping providers improve the business side of care without weakening accountability, privacy, or the patient experience.

MBBS Practice Growth Healthcare Operations

Connect on LinkedIn

Evidence basis

Primary-source review of AI billing call deployments

Built from Cedar’s engineering research on real billing call transcripts, Notable Chief Medical Officer Aaron Neinstein’s published account of health-system voice AI rollouts, physician forum discussions on AI answering services, and Staffingly’s own internal pilot testing of its billing call assistant. This article is operational information, not legal, compliance, or clinical advice.

Why Are Billing Calls Harder to Automate Than Scheduling Calls?

Scheduling and intake calls follow a predictable pattern. The caller wants an appointment, the system checks availability, and the call ends with a confirmed slot. Billing calls do not follow that pattern because the underlying data is rarely simple, and the patient on the line is often already frustrated before the call starts.

Cedar, the company behind the AI billing agent Kora, studied thousands of real billing calls and found that a question as simple as “why isn’t my bill showing up in the portal” rarely has one answer. The invoice might have been withdrawn, adjusted after a resubmission to insurance, already paid, or sent to collections. In Cedar’s review, roughly one in nine callers had a bill already in collections, and knowing that fact alone was not enough to know whether mentioning it would help or confuse the patient further. A simple rule like “if in collections, share the agency’s contact information” sent a large share of those calls down the wrong resolution path in Cedar’s internal testing.

This is the root cause behind most of the confusion patients report with automated billing lines. The problem is rarely the AI’s conversational ability. It is almost always a data or process gap sitting underneath the conversation, which is exactly why billing calls need a different design approach than scheduling calls.

What Has the Industry Already Learned the Hard Way?

Cedar’s own research shows how much a bad handoff can cost. In their data, 43% of patients said they had to re-explain their situation after being transferred from AI to a human agent, which is exactly the moment a practice most needs to be building trust, not eroding it. The fix is not avoiding automation. It is making sure the human agent receives full context the instant the call arrives, rather than starting the conversation over.

Aaron Neinstein, Chief Medical Officer at Notable, has written about a similar pattern across the health systems he has watched deploy voice AI. His advice is to treat the AI agent like a new hire: define its role, its safety guardrails, its escalation logic, and its success metrics before it ever takes a live call, the same way a practice would onboard a new front-desk employee. He also found that reviewing every single call in the first weeks after launch, then tapering to spot checks once the system earns trust, catches problems faster than months of simulated testing ever could.

Neinstein’s account includes a detail worth sitting with: patients often prefer the AI precisely because it never rushes them. One patient spent close to thirty minutes with an AI agent spelling out complicated medication names letter by letter, something a human agent would struggle to do during a Monday morning call surge, and still rated the interaction a perfect score. For billing calls, the same principle holds. Patience is a feature, not a compromise, as long as the guardrails around what the AI is allowed to decide stay tight. He also notes that decades of frustrating phone trees have primed patients to expect a fight before they even say hello, which is why practices that tell callers up front, in plain language, that they can talk normally rather than bark keywords tend to see calls go more smoothly from the first few seconds.

Sources: Cedar’s engineering research on real patient billing calls and Aaron Neinstein’s published account of health-system voice AI deployments at Notable. These figures describe the cited datasets and individual accounts, not a guaranteed outcome for any specific practice.

What Is Confidence-Based Routing and Why Does It Work Better Than Full Automation?

Confidence-based routing means the assistant does not try to resolve every call. It handles intake, triage, scheduling, and billing FAQ deflection, and it warm-transfers to a live representative whenever the conversation moves outside its confidence range. That design principle matters more than anything else in the model, and it is the same conclusion Cedar and Notable each arrived at independently: the goal was never to automate every call. It was to right-size the work, so callers with a simple question get an answer immediately and callers with a real dispute reach a person without having to ask for one.

For simple questions such as “what is my current balance” or “when is my payment due,” the AI can answer directly from integrated data. For anything involving a dispute, a payment-plan renegotiation, or a claim the patient believes was processed incorrectly, the call moves to a live team member automatically. Any practice evaluating an AI billing assistant should ask the vendor exactly where that line sits and how it is enforced, not just whether warm transfer exists as a feature on a spec sheet.

Language handling is part of the same design problem. An assistant that handles greeting, triage, and language switching as part of the same call flow means a Spanish-speaking patient asking about a balance does not need to be rerouted through a separate system before reaching billing FAQ deflection. Bilingual practices evaluating vendors should test this directly rather than assume it works, because a tool that handles English billing questions well can still fail badly the moment the conversation switches languages mid-call.

Why Must EHR and Phone System Integration Come First?

Cedar’s engineering team was direct about this: without a real data foundation connecting the AI to actual balances, claims, and adjustments, there is not much AI worth building at all. That principle holds regardless of vendor. An agent that sounds confident but is reasoning from stale or incomplete data will eventually tell a patient something wrong, and in a billing context that means a disputed charge or a support call that never should have needed one.

A deployment integrated directly with the practice’s scheduling system and phone routing platform can check a real balance, confirm a real appointment, and route a call based on real account status rather than a static script. Without that integration, the AI is working from incomplete information, which means every answer it gives is a guess dressed up as a fact. Practices that treat integration as a phase-two feature tend to end up with an assistant that sounds capable but answers questions incorrectly because it is disconnected from the system of record.

What Administrative Changes Does This Technology Actually Require?

The part of this technology that surprises most practices has nothing to do with the AI itself. It is how much the surrounding administrative workflow has to change to support it.

One documented example illustrates why this matters. After losing two physicians to retirement, a neurology practice saw call volume climb so high that its medical assistants’ voicemail boxes were receiving more than 200 messages a day, and two experienced MAs quit within six months because of the volume and increasingly frustrated patients. That is the exact situation AI voice automation is built to relieve, but it also shows why the technology cannot simply be switched on and left alone. Someone has to review escalated calls regularly to confirm the routing logic is working. Someone has to reconcile the answers the AI is giving against actual billing rules, since those rules change and a script that was accurate one quarter can be wrong the next. Someone has to own the relationship when a patient calls back frustrated, even when the handoff to a human went exactly as designed.

This is also where pairing AI with a dedicated specialist earns its place. The AI is not left to run without oversight, and the specialist is not stuck manually answering every routine question either. Each side of that pairing does the part it is actually suited for, which is a meaningfully different staffing model than either full automation or a fully human phone line.

What Mistakes Do Practices Make When Evaluating This Technology?

A few patterns show up repeatedly when practices shop for a billing call AI, and most trace back to the same root issue: evaluating the demo instead of the workflow underneath it.

• Testing only easy scenarios: A demo call that asks “what are your hours” will always go well. The real test is a call where the patient disputes a charge or has a bill already sitting in collections. Ask any vendor to run that exact scenario before signing anything.
• Turning on everything at once: Practices that launch full triage, scheduling, and billing automation simultaneously tend to see staff and patients lose trust quickly. A narrower rollout that expands once the first workflow proves stable holds up better.
• Assuming HIPAA compliance is a checkbox: A signed BAA and encrypted data at rest are the minimum, not the whole picture. Ask specifically how call recordings and transcripts are stored, who can access them, and whether any protected health information touches a model that was not built for HIPAA-compliant handling.
• Not defining the escalation trigger in writing: “It transfers to a human when needed” is not a specification. Ask for the exact list of scenarios that trigger a warm transfer and confirm that list matches what actually happens on a live call.
• Underestimating the integration timeline: Conversational setup can happen quickly. EHR and phone system integration takes longer and should be scoped honestly before a go-live date is promised.
• Reviewing calls only at launch, then stopping: The practices getting the most out of this technology review calls constantly in the first weeks, then taper off once the system has earned trust. Reviewing once at launch and never again is how drift goes unnoticed for months.

What Does HIPAA Compliance Actually Require for AI Billing Calls?

Patient billing calls sit squarely inside HIPAA’s reach because they involve protected health information tied to a specific account. Any AI vendor handling these calls needs a signed Business Associate Agreement, documented encryption for data in transit and at rest, and a clear audit trail for every call, transcript, and data access event.

This rules out generic consumer-grade voice AI tools, even ones that are technically capable of holding a conversation. A tool built for retail or hospitality call handling was not designed with protected health information handling, audit logging, or business associate coverage in mind, and retrofitting compliance onto a platform after the fact is a much harder and riskier project than starting with a healthcare-built platform.

What Does a Realistic Implementation Rollout Look Like?

Practices considering this kind of deployment should expect a sequence that looks roughly like the steps below, in this order.

Where Does Staffingly Fit in This Model?

Staffingly runs AI voice automation as part of a broader operational support model for medical practices, paired with a dedicated specialist rather than left to run unsupervised. A billing call assistant built this way handles intake, triage, scheduling, and billing FAQ deflection, integrated directly with practice EHR and phone systems, and warm-transfers to a live team member whenever a call moves outside its confidence range. In internal pilot testing, the system handled more than 50 simultaneous calls in a single test session without dropping the warm-transfer logic that keeps sensitive conversations in human hands.

That combination, a purpose-built voice tool plus a person accountable for the account, is the same pattern documented across other healthcare AI deployments, from Cedar’s engineering research to Notable’s health-system rollouts. A standalone chatbot cannot make an escalation judgment call on its own, and a fully human team cannot scale to cover every call without added headcount. The middle path, where AI handles the routine and reliably hands off the rest, is the one we would build again.

Practices weighing whether to bring AI into patient billing calls should start with the integration and escalation questions before the conversation design. That is the part that determines whether the tool actually works once real patients are calling.