AI Voice Agent Development in New York

Healthcare / Multi-Hospital Health System

A New York-based health system (18 hospitals, 4,200 physicians, 12 million patient encounters annually, serving the New York metro area) faced critical challenges in patient access and scheduling. (1) Call volume overwhelm: the health system's central scheduling operation received 2.8 million inbound calls annually — appointment scheduling, rescheduling, cancellation, physician referrals, and general inquiries. 380 scheduling agents handled these calls across 3 call centres. Average handle time: 6.2 minutes. Peak hours (8-10 AM weekdays) saw 40-minute hold times — causing 28% of callers to abandon before reaching an agent. Abandoned calls represented lost appointments: the health system estimated that each abandoned scheduling call cost $285 in lost revenue (average appointment revenue) — totalling $22.4M in annual revenue leakage from abandoned calls. (2) Scheduling complexity: the system managed scheduling for 4,200 physicians across 450 specialties and sub-specialties, 85 locations, and complex scheduling rules (provider availability, insurance network participation, new patient vs. established patient slots, referral requirements, preparation instructions, and equipment availability for procedures). Human agents required 6 months of training to handle the full complexity — and even experienced agents made scheduling errors (wrong provider, wrong location, missing referral requirements) at a 4.2% rate. (3) After-hours access: the scheduling operation ran 7 AM to 8 PM weekdays and 8 AM to 2 PM Saturdays — leaving 76 hours per week with no scheduling access. Patients calling after hours reached a voicemail system with a 62% callback completion rate — meaning 38% of after-hours callers never successfully scheduled their appointment. (4) Language diversity: the New York metro area patient population spoke 200+ languages. The health system offered live interpreter services for the top 15 languages, but interpreter wait times averaged 8 minutes — extending call times and frustrating patients. Spanish, Mandarin, Cantonese, Russian, and Korean were the most common non-English languages. (5) No-show rates: the system experienced a 14% no-show rate — partially attributable to inadequate appointment reminders (automated reminder calls had a 45% answer rate) and scheduling mismatches (patients scheduled at inconvenient times because agents were under pressure to fill slots rather than find optimal times).

We built a patient scheduling voice agent over 14 weeks, integrated with the health system's Epic EHR and scheduling infrastructure. (1) Natural conversation scheduling: the voice agent handled appointment scheduling through natural conversation — understanding patient requests ("I need to see a cardiologist," "I want to reschedule my appointment with Dr. Chen," "My doctor referred me to a specialist for my knee"), collecting necessary information (insurance, symptoms, preferences for time and location), and booking appointments that matched all scheduling rules. Multi-turn conversation: the agent managed complex scheduling conversations — "I need a follow-up with my endocrinologist, but I also need lab work done before that appointment, and I'd like both at the same location if possible." The agent checked provider availability, lab scheduling, and location options — presenting the patient with workable combinations. (2) Full scheduling rule engine: the agent enforced all 4,200 providers' scheduling rules — new patient vs. established patient slots, insurance network verification (real-time eligibility check via the health system's clearinghouse integration), referral requirements (checking whether a specialist referral was on file before scheduling), preparation instructions (fasting requirements, medication holds, arrival time), and procedure-specific requirements (equipment availability, anaesthesia scheduling, pre-procedure testing). Error rate: 0.8% scheduling errors (from 4.2% with human agents) — the voice agent consistently applied scheduling rules without the shortcuts that human agents sometimes took under time pressure. (3) 24/7 availability: the voice agent operated continuously — patients could schedule, reschedule, or cancel appointments at any time. After-hours volume: 35% of voice agent interactions occurred outside the previous scheduling operation hours (evenings, nights, weekends) — demand that had been unmet or poorly served by the voicemail system. (4) Multilingual capability: the voice agent operated natively in English, Spanish, Mandarin, Cantonese, Russian, and Korean — with real-time translation capability for 15 additional languages. No interpreter wait time — the agent switched to the patient's preferred language immediately upon detection or request. Language quality: native-speaker-level fluency in all 6 primary languages — not stilted machine translation but natural, culturally appropriate conversation in each language. (5) Smart scheduling and reminders: appointment optimisation — the agent considered patient preferences, historical patterns (when patients typically preferred appointments), and health system goals (filling underutilised time slots, reducing wait times for high-demand specialties). Proactive reminders: the agent called patients 48 hours and 2 hours before appointments — conversational reminders that could reschedule on the spot if the patient could not make the appointment, rather than simple automated messages.

Call handling: the voice agent handled 72% of scheduling calls without human intervention — 2.0 million of 2.8 million annual calls. The remaining 28% (complex cases, patient preference for human agent, escalation) transferred to human agents with full conversation context. Hold times: eliminated for voice-agent-handled calls (immediate answer). Human agent queue: average wait time from 40 minutes (peak) to under 3 minutes — because human agents handled only 28% of volume. Abandoned calls: from 28% to 4% — recovering approximately $19.2M in previously lost appointment revenue. After-hours access: 35% of appointments now scheduled outside business hours — capturing demand that was previously lost. Patient access: time from "patient decides to call" to "appointment confirmed" reduced from average 22 minutes (including hold time) to average 4.5 minutes. Scheduling accuracy: error rate from 4.2% to 0.8% — reducing patient frustration, rescheduling overhead, and wasted provider time from incorrectly scheduled appointments. Multilingual access: interpreter wait time eliminated for 6 primary languages — patient satisfaction among non-English speakers improved from 3.2 to 4.4 (out of 5). No-show rate: from 14% to 8.5% — driven by smart scheduling (better appointment-patient matching) and conversational reminders (65% answer rate vs. 45% for automated messages, with real-time rescheduling capability). Revenue impact of reduced no-shows: $15.7M annually (5.5 percentage point improvement on 12 million encounters at $285 average appointment revenue). Agent redeployment: 180 of 380 scheduling agents redeployed to complex patient navigation, insurance authorisation, and care coordination roles — higher-value work that improved patient outcomes and could not be automated. Annual value: $19.2M (recovered revenue from eliminated abandoned calls), $15.7M (reduced no-shows), $12.8M (agent cost optimisation — 180 agents redeployed from scheduling to higher-value roles), $4.2M (after-hours revenue capture), $2.1M (scheduling accuracy — reduced rescheduling and wasted provider time) = $54.0M. Development cost: $420,000. Payback: under 3 days.