Key Results
- 92% recognition accuracy for critical pharmacy phrases and medication names, surpassing the legacy IVR system.
- 85.3% transcription success rate on key medical vocabulary including drug names, alphanumeric member IDs, and clinical terms.
- Scaled to handle 1M+ pharmacy-related calls per day across 7,000+ locations, with up to 14,000 calls per hour in production.
- Higher IVR containment, fewer misrouted calls, and fewer documentation errors—reducing agent load, lowering contact center costs, and supporting HIPAA compliance.
- Future-ready architecture supporting multilingual expansion, including Spanish-language support, without replatforming.
The Challenge: A Legacy IVR That Couldn't Keep Up
With ~120 calls per store per day flowing across thousands of locations, the company's legacy Nuance IVR had become a bottleneck. Complex medication names, dosage instructions, and alphanumeric member IDs were frequently misheard, cascading into misrouting, repeated verification steps, and documentation errors. Noisy retail environments made recognition worse. And as self-service containment stayed low, more callers demanded live agents—driving up costs and wait times, especially during peak refill and vaccination periods.
When Nuance announced the end of life for its on-prem IVR stack, a like-for-like replacement wasn't an option. The company needed materially better medical vocabulary recognition, higher containment, national scale, and HIPAA-eligible infrastructure.
The Solution: Medical-Grade Speech AI for Next-Gen IVR
The company partnered with Deepgram through a leading conversational AI platform to rebuild its IVR speech layer from the ground up, with a major systems integrator supporting implementation.
Deepgram Nova Medical (STT) is trained specifically on medical terminology and phone audio—handling drug names, dosages, clinical concepts, member and plan identifiers, refill numbers, and prescription IDs with the precision required for reliable self-service and accurate routing.
Deepgram Aura (TTS) delivers natural, healthcare-appropriate voice responses that handle sensitive content—benefits explanations, coverage decisions, medication instructions—in a way that feels trustworthy to members.
The deployment included pharmacy-specific tuning for vocabulary, pronunciation dictionaries for medication names, SSML and prosody controls, and number formatting for member IDs and claim numbers. The architecture is cloud-native on AWS, HIPAA-eligible, and designed to support evolving deployment needs without rewriting IVR logic.
Outcomes
Recognition accuracy. 92% accuracy on critical pharmacy phrases and medication names, with an 85.3% transcription success rate on complex medical vocabulary—reducing downstream agent corrections and documentation errors.
National scale. 1M+ pharmacy-related calls handled per day across 7,000+ locations, with up to 14,000 calls per hour in production and a composable architecture ready to expand to additional lines of business.
Containment and cost. Higher recognition accuracy translated directly into more callers completing refills, status checks, and benefits questions inside the automated system—reducing live agent volume, lowering contact center operating costs, and easing store team pressure during peak periods.
Member experience. Instead of rigid menus, members encountered a natural-sounding IVR that understood them the first time. The combination of Nova Medical and Aura produced a smoother, more empathetic interaction befitting a Fortune 50 healthcare brand.
Looking Ahead
With production results strong, the company is expanding. Spanish and additional language support can be added using the same architecture. New workflows—vaccine scheduling, complex benefits questions, post-visit follow-up—can be layered in without compromising latency or accuracy. And as Deepgram advances Nova Medical and Aura, the organization can continue improving quality over time without destabilizing live operations.
Key Takeaways for Healthcare Enterprises
Prioritize domain-specific models. Generic speech engines fall short on pharmacy and medical terminology. Purpose-built models deliver measurable gains where errors affect care and compliance.
Plan for compliance from day one. HIPAA-eligible infrastructure and data residency should be evaluated early—not retrofitted after deployment.
Choose a partner, not just a vendor. Working together on formatting, pronunciation, and SSML details made the difference between a passable IVR and one that genuinely improved member experience at national scale.
