AI Voice Agent Development in Atlanta, GA

Healthcare / Multi-Hospital System

A major Atlanta healthcare system (8 hospitals, 180+ outpatient locations, 4,200 physicians, 2.8M patient encounters annually) was experiencing severe appointment scheduling friction. Scheduling problem: the health system receiving 42,000 appointment scheduling calls per week across its 180+ locations — each call handled by scheduling staff who navigated the scheduling system, verified insurance, matched the patient with the appropriate provider, and confirmed the appointment. Average call handling time: 6.8 minutes. Staff required: 185 FTE scheduling agents across locations and a centralised call centre. The problems: (1) hold times — peak period hold times averaging 12.4 minutes (8 AM-10 AM and 4 PM-6 PM on weekdays). Patient satisfaction surveys consistently citing hold times as the #1 complaint. 18% of callers abandoning before speaking with a scheduling agent — an estimated 7,560 abandoned calls weekly, each representing a potential missed appointment, (2) scheduling accuracy — scheduling agents making errors on 4.8% of appointments — wrong provider, wrong location, wrong appointment type, or scheduling conflicts. Each error requiring patient callback, rescheduling, and sometimes resulting in no-shows when patients arrived at the wrong location or time. Error correction: consuming an estimated 12% of scheduling staff time, (3) after-hours access — scheduling available only during business hours (7 AM-6 PM weekdays, 8 AM-12 PM Saturdays). 31% of scheduling call attempts occurring outside these hours — patients going to voicemail and requiring next-business-day callback. The callback: often missing the patient, requiring multiple attempts, and delaying appointment booking by 1-3 days, (4) insurance verification delays — scheduling staff manually verifying insurance eligibility during the call — adding 1.5-2 minutes to each call and creating errors when verification systems were slow or returned ambiguous results. Verification failures: causing 8% of calls to require a callback after insurance status was confirmed, and (5) provider matching — patients often unsure which provider or specialty they needed. Scheduling staff: not clinically trained to assess whether a patient needed a primary care visit, a specialist referral, or an urgent care visit. Misdirected appointments: contributing to provider no-shows and patient frustration.

We developed and deployed an AI voice agent for appointment scheduling over 16 weeks. Natural language scheduling agent (weeks 1-8). Conversational scheduling: voice agent handling appointment scheduling through natural conversation — patients calling and describing their need ("I need to see a doctor about my knee pain"), the agent determining the appropriate appointment type (orthopedic consultation), finding available appointments based on the patient's location, insurance, and provider preferences, and completing the booking. The conversation: natural and guided, not menu-driven. Patient identification: voice agent identifying patients through date of birth and name verification — matching against the patient database to pull up their record, insurance information, and care history. The identification: secure, efficient, and designed to handle common variations (nicknames, maiden names, pronunciation differences). Clinical triage logic: rules-based clinical triage embedded in the voice agent — determining whether the patient's described need is appropriate for primary care, urgent care, specialist referral, or emergency department. The triage: developed with the health system's clinical team to ensure appropriate care routing. If the patient described symptoms suggesting emergency care: the voice agent advising them to call 911 or visit the nearest ED, not scheduling a routine appointment. Provider matching: the agent matching patients with providers based on clinical need, insurance coverage, location preference, availability, and care continuity (offering their existing provider first when appropriate). The matching: considering the same factors a knowledgeable scheduling agent would, but checking all options simultaneously rather than scrolling through a schedule. Insurance verification (weeks 4-11). Real-time eligibility: the voice agent conducting real-time insurance eligibility verification during the call — checking coverage, deductible status, referral requirements, and copay amounts through payer APIs. The verification: completing in seconds rather than the 1.5-2 minutes of manual verification. Prior authorisation check: for specialist appointments, the agent checking whether prior authorisation was required and, if so, initiating the authorisation request automatically. The check: preventing the scheduling errors that occurred when patients booked specialist appointments without required authorisations. Coverage communication: the agent communicating insurance coverage details to the patient — estimated copay, whether the provider was in-network, and any pre-visit requirements. The communication: setting accurate patient expectations before the appointment. 24/7 availability and integration (weeks 8-16). Round-the-clock scheduling: the voice agent available 24/7 — handling the 31% of scheduling attempts that occurred outside business hours. After-hours appointments: booked and confirmed immediately, with confirmation details sent via text or email. The availability: eliminating the callback cycle that delayed after-hours scheduling by 1-3 days. EMR integration: the voice agent integrated with the health system's Epic EMR — reading appointment templates, checking provider schedules, writing appointment bookings, and accessing patient demographics and insurance. The integration: ensuring that voice-agent-scheduled appointments appeared correctly in provider schedules and patient records. Confirmation and reminders: automated confirmation text/email sent immediately after booking, followed by reminders at 48 hours and 2 hours before the appointment. The communication: reducing no-shows through consistent, timely reminders. Escalation: the voice agent recognising situations requiring human assistance — complex scheduling (multiple appointments across providers), emotional callers (frustrated, upset, or confused patients), and clinical concerns (symptoms suggesting urgent or emergency care). The escalation: transferring to a human scheduling agent with full conversation context so the patient did not need to repeat information.

The voice agent deployed across the health system over 12 weeks (phased by location type). Call handling performance: calls handled by AI voice agent: 58% of the 42,000 weekly scheduling calls (24,360 calls/week) resolved without human assistance. The remaining 42%: transferred to human scheduling agents, either by patient request or because the scheduling scenario exceeded the voice agent's capability. Call handling time: from 6.8 minutes (human) to 3.4 minutes (voice agent) for fully automated calls. The efficiency: the voice agent completing scheduling faster because it simultaneously checked all provider schedules and insurance in real-time. Patient experience: hold times: eliminated for voice-agent-handled calls (immediate answer). For human-agent calls: hold times reduced from 12.4 minutes to 4.2 minutes — the voice agent handling 58% of volume freed human agents to answer remaining calls faster. Call abandonment: from 18% to 4.2% — the immediate voice agent response preventing the frustration that caused callers to hang up. After-hours scheduling: 31% of scheduling call volume now handled instantly 24/7 — eliminating the callback cycle for after-hours callers. Patient satisfaction: scheduling satisfaction score from 3.1/5 to 4.3/5. Scheduling accuracy: scheduling errors: from 4.8% to 1.1% — the voice agent making fewer errors than human agents because it systematically verified each scheduling parameter. The improvement: reducing rescheduling calls, no-shows from misdirected appointments, and the staff time spent correcting errors. Insurance verification: completed in 4 seconds (voice agent) vs. 1.5-2 minutes (human) — accurate, immediate verification with no ambiguity. Prior authorisation identification: 100% of specialist appointments checked for authorisation requirements — vs. approximately 70% catch rate by human schedulers. Staff impact: scheduling FTE: from 185 to 112 — 73 scheduling positions reduced. Redeployment: 45 schedulers moved to patient services roles (pre-visit preparation, care coordination, patient financial counselling). Natural attrition: 28 positions eliminated through retirement and voluntary departure. The reduction: achieved over 6 months, with no involuntary layoffs. Remaining human schedulers: handling the complex scheduling scenarios that required human judgement — multi-provider appointments, accommodation requests, and emotionally sensitive situations. The human agents: doing higher-complexity, more rewarding work. Financial results: scheduling staff efficiency (73 FTE positions at $42K average fully loaded): $3.07M/year. Reduced call abandonment (7,560 fewer abandoned calls/week × estimated $85 appointment value × 40% conversion): $13.4M/year in recovered appointment revenue. Scheduling error reduction (fewer misdirected appointments, reduced no-shows): $1.8M/year. After-hours scheduling capability (immediate scheduling vs. callback delays): $2.4M/year in recovered scheduling efficiency. Total annual impact: $20.7M/year. Voice agent development: $520,000. Annual operations: $144,000/year. First-year ROI: 3,880%.