AI Medical Answering Service: What It Does on the Call, and What the Evidence Shows

What You Need to Know About AI Medical Answering Services

• An AI medical answering service answers inbound patient calls with a conversational voice agent. It books, reschedules, cancels, confirms, captures intake and routes calls, then hands anything outside its rules to a person.
• Our search found no peer-reviewed or institutional study reporting containment, resolution or satisfaction rates for administrative voice agents at a US health system. In a national survey, 65.8% of US adults reported low trust in their health system to use AI responsibly.¹
• Three controls decide whether patients accept one: disclosure at the start of the call, and an obvious path to a person. Escalation rules and a business associate agreement (BAA) go in writing before go-live.

What is an AI medical answering service, and how is it different from a traditional one?

An AI medical answering service answers inbound patient calls with a conversational voice agent instead of a human operator. It holds the conversation and completes routine requests such as booking, rescheduling or canceling. It writes the result into the practice record and hands anything outside its configured rules to a person.

A traditional service captures the request and delivers a message that staff work through the next morning. A voice agent finishes the request while the patient is still on the line.

Three shapes are on the market:

• Live-operator services, where people answer, capture the request and page on-call staff
• AI-first services, where a voice agent runs the conversation and escalates by exception
• Hybrid arrangements, where administrative calls go to the agent and clinical calls go to a person

Software for administrative support of a health care facility, including "appointment schedules," is excluded from the device definition at 21 U.S.C. 360j(o)(1)(A).² Booking a visit and advising on a symptom are different regulatory objects.

Among practice leaders polled in February 2026 on AI and automation for patient access, scheduling led at 31% and calls followed at 27%.³ For calls, those leaders focused mainly on voice bots and interactive voice response (IVR) systems for call routing.³

For independent and small multi-site medical practices deciding whether to put an AI on the main phone line, the category question comes before the vendor question.

What happens on the call, from greeting to escalation?

A call runs through six steps, from a greeting that discloses the automated system to escalation for anything outside the configured rules. Design the escalation step first, because it shapes how the whole call feels.

The sequence looks like this:

1. The agent answers and states that the caller is speaking with an automated system.
2. It recognizes the intent behind the caller's first sentence.
3. It verifies identity against the proof points the practice configured, where the request requires it.
4. It completes the transaction: scheduling, rescheduling, cancellation, confirmation, intake capture or a policy question.
5. It documents the result in the practice's record rather than in a message queue.
6. It escalates anything the rules do not cover.

Commure AI Call Center Agents run that sequence on inbound patient calls, with the escalation triggers configured during scoping and implementation.

Design the out-of-scope path before go-live

A call the agent cannot handle at nine on a Saturday evening has four destinations. It can become a structured task written into the record for the next business morning or a transcribed voicemail tasked to a named queue. It can also go by warm transfer into the practice's on-call path or to the live answering service the practice already uses.

The practice writes those rules

Escalation triggers are set during scoping and cover clinical questions, urgent symptoms, distress, repeated recognition failure and any request for a person. Asking a caller to self-declare urgency moves the triage decision to the caller.

The agent's scope stays administrative: it gives no medical advice, performs no clinical triage and takes no payments. Verifying insurance eligibility and benefits stays with staff as well.

What determines whether an AI medical answering service will work at your practice?

Fit turns on five things: the administrative share of calls, the volume and shape of the peak, and EHR write-back. Telephony compatibility and the language profile of the patient panel complete the list. Measure all five on the practice's own call data before scoping begins.

• Call mix. Count what share of inbound calls are scheduling, rescheduling, confirmation and policy questions, and what share are clinical. A line where clinical calls carry most of the volume is a poor first candidate.
• Volume and peak shape. Look at the hourly distribution rather than the monthly total, since hold time is created inside the peak windows.
• EHR write-back. Ask whether the agent books into the schedule or produces a message someone rekeys.
• Telephony. Confirm the agent works with the current phone platform. Compatibility is validated during scoping, and infrastructure replacement is generally not required.
• Language and speech. Confirm coverage for the languages the panel speaks.

In a small 2026 comparison of two speech recognition systems, word error rates ran higher for non-native English speakers in both, and the gap was larger in Whisper, the only system where it reached statistical significance.⁵ Measure accent and second-language performance on the practice's own calls before go-live.

A patient whose request the system fails to understand may abandon the booking rather than call back. The agent can still record that call as handled, because the caller hung up rather than escalated.

Provider matching, visit-type constraints, location rules and escalation triggers are all configured during scoping against the practice's own templates. Ask the vendor to show the configuration for each of them before signing.

What does the evidence show, and what should you require before it answers a patient call?

The evidence base is small. Our search found one peer-reviewed deployment study of an AI call platform, and it covers a single practice. Three controls follow: disclose the AI at the start, publish a path to a person, and put governance in writing.

What the published evidence covers

The published deployment study of an AI call platform we found covers a single neurology practice.⁶ That study is uncontrolled, and it was written by an author affiliated with the practice.⁶ The uncontrolled single-practice report describes communication backlogs reduced by over 98%.⁶ The same report estimates about $216,500 in annual labor savings, and it reports neither patient satisfaction nor any safety outcome.⁶

We found no peer-reviewed or institutional study reporting containment, resolution or satisfaction rates for administrative voice agents at a US health system, and the prospective evidence has yet to be produced. Prospective studies tailored to the intended use and the level of clinical risk are essential to evaluating the effectiveness and safety of voice AI agents.⁴

The figure of more than 99% accuracy is widely repeated in this category. It traces to a safety evaluation of medical advice that was a preprint, not yet peer reviewed, when a 2025 perspective cited it.⁴

What patients say they want

The surveys below measure trust and notification preferences rather than performance on the call.

In a survey of US adults published in 2025, 65.8% reported low trust in their health system to use AI responsibly.¹ In the same survey, 57.7% reported low trust that their health system would protect them from harm caused by AI.¹

Between 84% and 86% of respondents to a national survey of US adults published in January 2026 want notification or permission across the three AI uses tested.⁷ For administrative uses such as billing and scheduling, 39% want to be notified each time.⁷

ECRI ranked misuse of AI chatbots the top health technology hazard for 2026.⁸

Three controls to require before go-live

Each control is a decision to settle in the script or the contract before the first patient call.

Disclose the automated system at the start of the call

As of September 2026, disclosure duties differ by state.

• California requires an AI disclaimer and instructions for reaching a human on generative AI communications about clinical information.⁹ That section expressly excludes administrative matters, including appointment scheduling and billing.⁹
• Utah requires prominent disclosure by a person providing the services of a regulated occupation in a high-risk interaction, and high-risk interactions include those that collect health data.¹⁰ Its safe harbor asks the system to disclose that it is AI at the outset of and throughout the interaction.¹⁰
• Texas requires a health care practitioner who uses AI for diagnostic purposes to disclose that use to patients.¹¹ A separate provision requires the provider of a health care service or treatment to disclose that the patient is interacting with AI where AI is used in relation to that service, no later than the date the service is first provided.¹²
• Illinois goes the other way for therapy and psychotherapy services delivered by licensed professionals, naming "managing appointment scheduling and reminders" as permitted administrative support while barring AI from therapeutic communication with clients.¹³

Of these state laws, Utah's can reach an administrative call. Collecting health data is enough to trigger it, with or without clinical advice.¹⁰ The Texas provider duty may reach one too, since it covers AI used in relation to a health care service.¹² California excludes administrative matters outright, the Texas diagnostic-use duty is scoped to diagnosis, and Illinois treats scheduling as permitted support.

Disclosure at the start of every call clears Utah's safe harbor, costs one sentence of script, and matches the notification preference patients report.⁷ Under the California, Illinois and Texas diagnostic-use provisions, it is a control the practice adopts rather than a duty imposed on this use.

Publish the path to a person

The control has three parts: a written escalation policy, a route that needs no code, and a reviewed escalation rate. An unannounced zero-out leaves the caller no way to know the route exists.

Put the data handling in writing before any call is recorded

A vendor that records calls, stores transcripts or holds messages maintains protected health information for the practice and is a business associate. The Department of Health and Human Services (HHS) limits the conduit exception to transmission-only services.¹⁴ ¹⁵ Any storage must be no more than temporary and incident to that transmission.¹⁵

HHS states that a vendor holding encrypted data it cannot read is still a business associate.¹⁵ The status holds even when the vendor lacks the key.¹⁵

An AI service is a business associate on the same terms as a live one. Sign the BAA before any data moves.

Verify the recording consent rule in every state where the practice takes calls. Ask the vendor how consent is captured and logged.

The clauses in that BAA, and the safeguards behind them, are the working definition of a HIPAA compliant answering service.

What does it look like when it works well?

In a working deployment, the line is answered during the mid-morning peak and at nine in the evening. The appointment lands in the schedule rather than a message queue, the clinical call reaches the on-call path, and staff spend the day on the work in front of them.

Automation of the administrative share adds capacity to the team that is already there. A practice whose call volume grows absorbs that growth without adding an overnight shift or paying peak-window overtime. The call that would have gone unanswered can become a confirmed appointment.

Governance continues as a standing weekly task after launch: review escalated calls, read a sample of transcripts, and confirm that the disclosure line played at the start. Automation comes last in the order of changes among healthcare call center best practices. The changes that come first are instrumenting every line, publishing a service-level target, writing escalation rules by call reason, staffing the peak and building a quality loop.

Track four measures in the practice's own data rather than the vendor's, starting in the first month:

• The escalation rate
• The share of callers who abandon after a transfer
• The bookings that needed rekeying
• The complaints that mention the automated system

An answering service is one use of AI voice agents in healthcare, and the wider category is worth reading before scoping starts.

How should you evaluate an AI medical answering service before you sign?

Run the same ten questions at every vendor and compare the answers side by side. The questions that separate vendors are about disclosure, escalation, EHR write-back, language testing and what the vendor has measured. Ask what the vendor has not measured as well as what it has.

The choice between live, AI and hybrid follows the call mix

Clinical volume points to people. A line where scheduling, rescheduling, confirmation and policy questions carry most of it is the candidate for an AI medical answering service. A mixed line can land on a hybrid split by call type rather than by hour.

In a survey published in April 2026, 22% of US adults said they get health information from AI chatbots at least sometimes.¹⁶ Among those chatbot health users, 23% rated that information as not too accurate or not at all accurate, and 18% rated it highly accurate.¹⁶

The last question is price, which medical answering service pricing breaks down by model.

How do Commure AI Call Center Agents handle inbound patient calls?

Commure AI Call Center Agents answer the inbound line in English and Spanish, 24/7/365, and complete the administrative requests in their configured scope inside the call instead of leaving a message. Complex or non-standard scenarios escalate to staff, and medical or urgent clinical questions can route to the provider on call under routing logic configured during scoping and implementation.

The call types in scope today:

• Appointment scheduling and rescheduling within configured provider, location and visit-type rules
• Appointment cancellation, with notice of the applicable cancellation policy
• Appointment confirmation, after identity verification against configured proof points
• New patient intake, and demographic or insurance updates for existing patients
• Frequently asked questions, and intent-based routing to the front desk, billing, pharmacy or the on-call provider

Insurance information is collected without a benefits or eligibility check, and the agent writes bookings, reschedules and intake into the configured EHR workflow in eClinicalWorks, athenahealth, Epic, ModMed, MEDITECH or AdvancedMD, with other systems scoped for feasibility.

A BAA is signed before any data is shared. The dashboard keeps call activity, call success status and full transcripts, so the practice can audit disclosure and escalation.

Commure Agents are an enterprise deployment that requires scoping, and they suit multi-location groups and regional health systems with a centralized line.

Request a complimentary call center health analysis to see your own call mix. Over about a week, it reviews one week of call transcripts and metadata under a signed BAA. The output is a breakdown by call type, length, category and outcome, plus a model of where automation would help.

‍