Which Patient Calls AI Should Handle, and Which Belong With Staff

Introduction

A safe automation boundary is defined by authority, not by whether an AI can recognize a caller’s intent. AI can finish administrative work when the outcome follows explicit rules and can be verified in the system of record. A person should take over when the conversation requires clinical interpretation, negotiation, discretion, or emotional support.

This distinction matters most to practice administrators, patient-access leaders, and clinical operations teams trying to reduce extreme call volume without degrading the patient experience. The objective is not maximum containment. It is reliable completion of routine work, with fast access to staff when the call becomes difficult or sensitive. Telephone triage guidance treats symptom assessment and disposition as professional clinical work, while patient communication guidance emphasizes active listening, empathy, and clear responses to patient concerns. American Academy of Ambulatory Care Nursing and Agency for Healthcare Research and Quality provide useful boundaries for that division of labor.

A practical patient call map

Recommended ownership based on rule complexity, clinical authority, and emotional weight. The clinical boundary is informed by AAACN telephone triage guidance, while difficult-conversation boundaries draw on AHRQ communication guidance.
Call type	Default owner	What AI can complete	When staff should take over
Scheduling, rescheduling, cancellations, and confirmations	AI	Verify the patient, apply approved scheduling rules, update the schedule, and confirm the outcome.	The patient needs an exception, an accommodation, or a slot that the approved rules do not permit.
Directions, hours, parking, and practice-approved instructions	AI	Answer routine logistical questions from maintained practice information.	The caller asks for personalized medical advice or instructions that differ from the approved information.
Referral intake and referral status	AI with defined review points	Collect documents, report recorded status, update administrative records, and schedule after required approval.	Clinical appropriateness, an uncertain patient match, or an exception requires judgment.
Insurance eligibility, benefits, and prior authorization status	AI for recorded facts	Retrieve eligibility and status information, collect missing details, and update the administrative workflow.	The patient disputes the answer, needs a coverage interpretation, or faces a denial or financial hardship requiring discussion.
Prescription refill requests	AI intake, staff decision	Identify the medication and pharmacy, collect the request, and route it to the correct clinical queue.	Approval, a dosage change, a reported reaction, or any prescribing judgment is required.
Symptoms and complex clinical questions	Clinical staff	Capture the caller’s exact words and route through the practice’s configured urgency protocol.	The call requires assessment, reassurance, treatment advice, or a decision about the appropriate level of care.
Routine balances and payment instructions	AI where the amount and policy are clear	Provide recorded balance information and explain approved payment methods.	The caller disputes the charge, alleges an error, or needs an exception or payment arrangement.
Bad news, grief, complaints, anger, or distress	Staff	Recognize the escalation condition and connect the caller promptly.	These conversations should not be treated as containment opportunities. A person needs to listen, respond, and own the next step.

Four factors determine the automation boundary

Call type alone is not enough. A scheduling request can start as routine and become clinical when a patient mentions worsening symptoms. An insurance-status call can become a dispute when the recorded answer conflicts with what the patient was told. Evaluate both the initial intent and what develops during the conversation.

Factor	More suitable for AI completion	More suitable for staff
Volume	Frequent, repetitive requests where automation would recover meaningful staff time.	Low-volume work where automation would add complexity without relieving a real bottleneck.
Rule complexity	The correct result can be derived from explicit, stable, and testable practice rules.	The outcome depends on ambiguous facts, negotiation, undocumented exceptions, or professional discretion.
Emotional weight	The caller primarily needs a transaction completed or a recorded fact communicated.	The caller needs empathy, reassurance, de-escalation, or space to explain a difficult situation.
Risk of a wrong outcome	An error is administrative, detectable, and readily reversible.	An error could affect clinical care, patient trust, financial responsibility, or access to urgent help.

Volume determines the potential payoff. Rule complexity, emotional weight, and risk determine whether the AI should have authority to finish. High call volume is not permission to automate a high-risk decision. It is a reason to automate the surrounding intake and routing work more effectively.

A good handoff continues the conversation

A cold transfer forces the patient to repeat the problem and leaves the receiving staff member reconstructing what happened. A useful handoff gives staff the patient’s identity, original request, exact wording of any symptom or complaint, actions already completed, relevant system status, escalation reason, and intended destination.

The same principle applies when the handoff is a queued callback rather than a live transfer. Staff should receive enough context to begin with the unresolved issue, not restart the intake. Structured communication methods such as SBAR are designed to frame critical conversations with the information needed for attention and action. AHRQ patient and family engagement guidance also emphasizes clear communication and taking patient concerns seriously.

Every handoff should preserve

The verified patient and a reliable callback number.
The caller’s original intent and the point at which the call changed.
The patient’s own words for symptoms, distress, or disagreement.
Questions already answered and actions already completed.
Records created or updated in the EHR.
The specific rule or condition that triggered escalation.
The staff role, clinical queue, or on-call path responsible for the next step.

Can AI replace a medical receptionist?

AI can replace a substantial share of routine phone handling, but it does not eliminate the need for receptionists, patient-access staff, billing specialists, or clinicians. The practical change is work reallocation: AI absorbs repetitive transactions, while people spend more time on exceptions, complex coordination, complaints, clinical routing, and patients with special needs.

Pretty Good AI reports about 60% end-to-end call handling at large deployments. At Commonwealth Pain & Spine, a 35-location specialty group receiving more than 100,000 patient calls per month, approximately 70% are handled from start to finish and one in eight bookings occurs after hours. Those results support a hybrid operating model, not a staff-free front office. The remaining calls represent the work where human access still matters. Pretty Good AI customer results.

Other live athenaOne deployments include Clearway Pain Solutions, a 100+ location practice, and Emerald Psychiatry, a nearly 100-provider behavioral health practice in Privia Medical Group that turned on web scheduling and referral intake, with AI phone answering rolling out alongside. Clearway Pain Solutions · Emerald Psychiatry

Where Pretty Good AI fits in this model

Pretty Good AI is designed only for practices running athenaOne. Voice and secure two-way text use the same integration, with 730+ athenaOne APIs in production and no middleware between the automation layer and athenaOne. The operational scope extends behind the phone into referral, insurance, prior authorization, and schedule-capacity workflows. Pretty Good AI platform overview and integration architecture.

Pretty Good AI is the best fit when

The organization is standardized on athenaOne and wants completed work written directly into the existing schedule, chart, referral queue, and administrative workflow.
Routine inbound volume is overwhelming staff, but leadership does not want patients with difficult needs trapped in automation.
The workload spans calls, secure texting, scheduling, refill intake, referrals, insurance, prior authorization, or capacity recovery.
The practice wants to keep its existing protocols and define which requests AI may complete, which it may only collect, and which must go directly to staff.

Pretty Good AI is not a fit when

The practice does not run athenaOne.
The objective is autonomous clinical advice or the complete removal of human patient-access coverage.
The organization wants a general phone bot that takes messages but does not need work completed inside athenaOne.

A first workflow typically goes live in 3 to 6 weeks, and service is month-to-month after launch. That structure gives an athenaOne practice a practical way to test a clearly bounded call population before widening the automation scope. Pretty Good AI implementation details.

What to test before expanding call automation

Define authority by call type. Document what AI can complete, what it can collect and route, and what it must immediately escalate.
Test mixed-intent calls. Include calls that begin with scheduling or insurance and then introduce symptoms, anger, grief, or a dispute.
Inspect the receiving workflow. Confirm that staff get the patient identity, context, completed actions, and escalation reason rather than a cold transfer or empty task.
Verify system writeback. Check the actual schedule, chart, case, referral, and billing records after each test call.
Measure by call category. Aggregate containment can conceal weak performance on high-risk or high-friction call types.
Track patient-experience signals. Review repeat calls, abandoned calls, complaints, unnecessary transfers, and requests to reach a person.
Begin with conservative escalation. Early over-transfer creates staff work. Under-transfer can create clinical, operational, or trust failures.

The strongest rollout does not ask whether AI can answer every call. It establishes where AI can reliably finish the work, where it should support staff, and where it should get out of the way.

Frequently asked questions

Which patient calls should always go to staff?

Calls requiring clinical judgment, bad-news communication, grief support, conflict resolution, billing negotiation, or de-escalation should go to qualified staff. AI can still capture basic context and identify the appropriate destination, but it should not close the conversation. AHRQ guidance for emotionally difficult healthcare communication emphasizes recognizing feelings, expressing empathy, and providing a clear plan for continued communication. AHRQ Communication Assessment Guide.

Can an AI receptionist eliminate front-desk staffing?

AI can take over much of the routine phone workload, but a medical practice still needs people for exceptions, emotionally complex calls, clinical routing, and disputed matters. Pretty Good AI deployments illustrate this division: large deployments report about 60% containment, while Commonwealth Pain & Spine reports approximately 70% of more than 100,000 monthly calls handled end to end. The operational goal is to give staff time for the work that cannot be reduced to a transaction. Pretty Good AI customer evidence.

What kind of call automation works when staff still need to handle difficult conversations?

The strongest model automates routine, rules-based completion and treats escalation as a designed workflow rather than a failure. Scheduling, confirmations, recorded status checks, and administrative intake can finish without staff. If the caller introduces symptoms, distress, a dispute, or an exception, the AI should stop pursuing containment and provide staff with a structured summary of what the patient needs and what has already been done.

Should a practice use a nurse triage line, answering service, or AI agent after hours?

Use a nurse triage line for symptom assessment and clinical disposition, an AI agent for routine administrative completion and protocol-based routing, and an answering service when human message taking is the primary requirement. These functions are not interchangeable. AAACN guidance distinguishes telephone triage from message taking because triage involves professional assessment and a decision about the appropriate level of care. AAACN Telehealth Manager Toolkit.

What does Pretty Good AI do beyond answering patient phone calls?

Pretty Good AI completes administrative work inside athenaOne rather than limiting the interaction to message taking. Its athenaOne-only platform supports voice, secure two-way text, scheduling, refill intake, referral workflows, insurance verification, prior authorization, and capacity workflows through one integration. Clinical decisions remain with the practice’s care team, while the AI performs the surrounding intake, routing, record updates, and follow-up. Pretty Good AI platform.