## Background
This video was published on YouTube by the Ai4 channel on August 22, 2026. The page description outlines a fireside chat featuring Yogesh Bhagat from CVS Health and Kris Efland from Deepgram, focusing on how CVS moved beyond legacy IVR to build an enterprise-grade voice AI platform for healthcare.
## TL;DR
CVS Health processes over 1 billion prescriptions and handles 400 million inbound calls annually across its vast healthcare footprint. To overcome the rigid limitations of traditional IVR systems, CVS transitioned to generative voice AI agents, relying on on-premise GPUs, RAG architectures, and strict guardrails to manage latency and safety. This technology shifts operational focus from simple call containment to actual task completion, ultimately freeing up pharmacists to focus on one-on-one clinical patient care. Looking forward, CVS and Deepgram are advancing toward multimodal interfaces, outbound automated voice agents, and sub-second speech-to-speech models that capture acoustic context and emotion.
## Atmosphere & style
The session carries an authoritative, pragmatic, and collaborative tone between two seasoned technology leaders. Yogesh Bhagat offers candid insights into real-world trial and error, blending grand sci-fi visions like Star Trek and Marvel's Jarvis with grounded engineering realities. The conversation feels forward-looking yet strictly tied to human impact and operational reliability in high-stakes healthcare.
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=iX__oeGAxPI&t=4s) CVS Health operates voice infrastructure at a massive enterprise scale, dispensing 1 billion prescriptions annually and managing over 400 million inbound calls across 9,000 retail locations.
- [02:28](https://www.youtube.com/watch?v=iX__oeGAxPI&t=148s) Voice serves as a vital cross-ecosystem communication channel connecting patients, pharmacists, doctors for prescription checks, and insurance payers for prior authorizations.
- [04:09](https://www.youtube.com/watch?v=iX__oeGAxPI&t=249s) Legacy IVRs resolved ~70% of basic queries using rigid tree structures, forcing callers to speak like machines rather than expressing themselves naturally.
- [04:09](https://www.youtube.com/watch?v=iX__oeGAxPI&t=249s) Scaling generative voice AI introduced significant technical challenges around real-time latency, compliance, domain jargon, and the need for on-premise GPU hosting to manage costs.
- [07:24](https://www.youtube.com/watch?v=iX__oeGAxPI&t=444s) After experimenting with hardcoded constraints and fine-tuning, CVS found that coupling models with strict Retrieval-Augmented Generation (RAG) solved 80% of guardrail issues.
- [09:55](https://www.youtube.com/watch?v=iX__oeGAxPI&t=595s) Operational success metrics shifted away from containment and low Average Handle Time (AHT) toward task completion, user satisfaction, and clinical outcomes.
- [07:24](https://www.youtube.com/watch?v=iX__oeGAxPI&t=444s) The primary strategic motivation for deploying AI voice agents is freeing up pharmacists' time from non-clinical tasks to focus on one-on-one patient care.
- [14:18](https://www.youtube.com/watch?v=iX__oeGAxPI&t=858s) The voice industry is evolving toward deep domain context and multimodal interfaces, where voice commands dynamically drive visual app interfaces in real time.
- [14:18](https://www.youtube.com/watch?v=iX__oeGAxPI&t=858s) CVS is expanding into outbound voice agents designed to navigate external IVR systems and wait on hold on behalf of pharmacists.
- [17:01](https://www.youtube.com/watch?v=iX__oeGAxPI&t=1021s) Deepgram's Flux model addresses sub-second latency challenges while natively handling turn-taking, natural interruptions, acoustic context, and user sentiment.
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=iX__oeGAxPI&t=4s) Introductions and the Massive Scale of CVS Health
Kris Efland, VP of Engineering at Deepgram, opens the session by introducing himself and inviting Yogesh Bhagat, VP of Digital and Technology at CVS Health, to share his role. Yogesh quipped that he is the person responsible for "keeping your prescription bottle filled." His team oversees the digital and technical architecture of CVS's pharmacy dispensing system, where voice acts as a vital channel connecting patients, pharmacists, and the broader healthcare network. Kris highlights Deepgram's foundation as an AI voice platform supplying speech-to-text, text-to-speech, and full voice-agent infrastructure to most companies at the conference.
Yogesh breaks down the sheer size of CVS Health to set the context for their engineering operations. While most people envision CVS as a neighborhood retail pharmacy, CVS Health is actually one of the largest healthcare enterprises in the world, spanning Aetna (health insurance), Caremark (Pharmacy Benefit Manager / PBM), MinuteClinic, Signify Health, and over 9,000 retail locations.
- **1 billion prescriptions:** Processed and dispensed by CVS every single year.
- **400 million+ inbound calls:** Handled across CVS retail store locations annually.
- **185 million unique customers:** Served across all business units—representing over 50% of the entire U.S. population.
- **85% proximity:** 85% of the U.S. population lives within 10 miles of a CVS location.
- **Human impact:** Yogesh emphasizes that despite massive operational numbers, the overarching priority remains enabling high-value, one-on-one clinical interactions between pharmacists and patients to drive real health outcomes.
---
[02:28](https://www.youtube.com/watch?v=iX__oeGAxPI&t=148s) The Criticality of Voice in Healthcare Ecosystems
Kris asks Yogesh to reflect on why voice has remained such a pivotal channel prior to the generative AI revolution. Yogesh takes an evolutionary angle, framing voice as humanity's primary mode of communication dating back to hunter-gatherer times—whether signaling danger with a quick sound or sharing complex ideas. Within the pharmacy landscape, meeting hundreds of millions of callers on the phone isn't optional; it's a fundamental obligation to meet customers on their preferred channel.
Beyond patient inquiries, internal and ecosystem-wide communication relies heavily on voice. CVS operated hundreds of traditional voice servers long before AI was introduced. Store-to-store communication between pharmacists across 9,000 locations relies on voice channels, as does the broader healthcare network. CVS pharmacists must frequently talk directly to doctors for prescription clarifications and safety checks, or coordinate prior authorizations with insurance payers.
---
[04:09](https://www.youtube.com/watch?v=iX__oeGAxPI&t=249s) The Limits of Legacy IVRs and the Post-2022 GenAI Shift
Reflecting on the fall of 2022 as a industry-wide watershed moment—moving from predictive models to generative AI—Kris asks what CVS’s legacy stack looked like and what boundaries they were hitting. Yogesh notes that even under legacy IVR systems, CVS managed to resolve roughly 70% of customer queries, but only for basic, rigid questions.
Legacy IVR systems rely on tree structures ("press 1 for option A, press 2 for option B"). Callers were forced down rigid branches where jumping between nodes or phrasing requests naturally was impossible. As Yogesh observes, "humans were forced to talk like machines so that we can get our answers." The shift after 2022 liberated callers from machine-like menus, letting them express themselves naturally in conversational language while still obtaining accurate answers.
However, real-world deployment brought immediate infrastructure hurdles. Early OpenAI demos showed seamless conversational fluency, making the technology look simple at first glance. Moving from a demo to enterprise scale introduced major bottlenecks:
- **Latency & real-time response:** Real-time speech requires sub-second processing speed.
- **On-premise GPU hosting:** To manage latency and avoid runaway API cost structures, CVS set up its own colocation environments with dedicated GPUs to host proprietary LLMs and LSLMs (Large Speech Language Models).
- **Compliance & safety:** Tackling hallucinations, regulatory requirements, domain jargon, and model bias required dedicated engineering rather than off-the-shelf integration.
---
[07:24](https://www.youtube.com/watch?v=iX__oeGAxPI&t=444s) Architectural Iteration, Guardrails, and RAG Simplicity
When deciding to go "all-in" on voice AI, the key driving factor was freeing up pharmacists' time. If non-clinical, administrative questions can be answered accurately by a voice agent, pharmacists regain critical hours to deliver personalized clinical advice to in-person patients.
Yogesh candidly describes the architectural learning curve CVS went through:
1. **The Over-Control Phase:** Engineers initially tried to restrict interactions by hardcoding specific option outputs for every prompt. They quickly discovered that consumer questions are virtually infinite, making rigid options ineffective.
2. **The Wild-West Phase:** Loosening constraints led to models answering out-of-context questions they shouldn't have answered.
3. **The Guardrail & RAG Solution:** They learned to constrain the AI's parameter space by coupling the model with strict Retrieval-Augmented Generation (RAG).
After testing distilled models, fine-tuning, and complex custom training pipelines, CVS determined that "80% of the time, just a simple RAG application solves your problem instead of thinking about distillation or model training."
---
[09:55](https://www.youtube.com/watch?v=iX__oeGAxPI&t=595s) Redefining Operational Metrics: From Containment to Clinical Outcomes
Kris asks about the concrete outcomes CVS tracks today. Yogesh reveals a major cultural shift in how engineering and operations define success. Under legacy IVR systems, performance was measured purely by operational containment rate (how many callers were prevented from reaching a human) and minimizing Average Handle Time (AHT).
With intelligent voice AI agents, those legacy metrics become counterproductive:
- **Task completion over containment:** Success is now evaluated on whether non-clinical tasks were completely resolved inside the voice platform without forcing the patient through additional hoops.
- **Rethinking Average Handle Time:** Longer handle times are no longer viewed negatively. If a patient spends four minutes asking three distinct questions and gets thorough, helpful answers, the high handle time reflects high value, not operational inefficiency.
- **Clinical and Health Metrics:** Focus has shifted toward clinical outcomes—measuring actual healthcare actions completed by patients rather than treating voice merely as a cost-cutting tool.
---
[12:02](https://www.youtube.com/watch?v=iX__oeGAxPI&t=722s) Cross-Industry Evolution and Star Trek-Level Context
Yogesh turns the tables to ask Kris how generative AI has transformed the broader voice industry. Kris notes that after a decade of building voice models at Deepgram, the industry has finally crossed the threshold into truly human-like conversation, eliminating the robotic pauses and misinterpretations of earlier eras.
Kris believes the next frontier is deep domain context. Instead of transactional exchanges, voice agents will operate with instant access to complex enterprise data—understanding a patient's drug formulary, active treatment plan, and complete medical history simultaneously. He remarks that society has effectively entered "Star Trek land," yet many companies are still thinking incrementally rather than embracing rich, context-aware agent workflows.
---
[14:18](https://www.youtube.com/watch?v=iX__oeGAxPI&t=858s) Future Vision: Multimodal "Jarvis" Interfaces and Outbound Agent Collaboration
Looking ahead beyond CVS's 9,000 stores, Yogesh shares his passion for science fiction, referencing how Tony Stark talks to Jarvis in Marvel films. He envisions a future where voice and visual interfaces merge seamlessly—users will navigate apps using voice while the visual UI dynamically updates screens and context in real time.
Tactically, CVS's immediate next frontier is expanding from inbound caller support (400M+ calls) to **outbound voice agents**.
- **Agent-to-IVR Communication:** Pharmacists currently waste significant time on hold when calling external insurance companies or prescribers. CVS is building outbound voice agents capable of calling external IVR systems, waiting on hold, navigating menus, and resolving administrative conflicts directly with other automated systems.
- **Preserving Pharmacist Time:** The primary goal remains giving time back to pharmacists so they can focus on human patient care without administrative distraction.
---
[17:01](https://www.youtube.com/watch?v=iX__oeGAxPI&t=1021s) Deepgram’s Flux Model, Speech-to-Speech, and Emotional Intelligence
Kris breaks down Deepgram's technical roadmap, emphasizing that conversational real-time voice requires round-trip processing latencies under 500 to 800 milliseconds. While industry discussions frequently highlight "speech-to-speech," Kris notes that most current production implementations rely on cascaded pipelines (STT $\rightarrow$ LLM $\rightarrow$ TTS) to enable backend tooling and controls.
To bridge this gap, Deepgram developed its **Flux** fused models, which directly handle natural speech dynamics:
- **Turn-taking & Interruptions:** The model natively detects conversational rhythm, allowing users to interrupt or hesitate naturally without getting cut off prematurely.
- **Adaptive Grace Periods:** If a user is reading a long prescription number without their glasses and pauses frequently, the agent recognizes the pattern and dynamically grants extra grace time before responding.
- **Preserving Acoustic Context:** Converting speech to intermediate text strips away rich acoustic data—tone, pitch, inflection, and emotion. Pure end-to-end processing preserves situational and emotional intelligence, enabling an agent to sense an upset caller and adjust its tone to de-escalate the situation.
- **Background Tooling Trigger:** Architectures like Flux allow integrators to trigger backend actions mid-sentence as soon as specific keywords or medication names are spoken.
<ElicitationsGroup message="Explore related technical and enterprise AI topics:">
<Elicitation label="Analyze enterprise voice AI architectural tradeoffs" query="What are the architectural tradeoffs between cascaded voice pipelines (STT-LLM-TTS) vs true end-to-end speech-to-speech models?"/>
<Elicitation label="Explore voice agent guardrail design patterns" query="Explain design patterns for building guardrails and RAG pipelines in real-time conversational voice agents."/>
</ElicitationsGroup>
I cannot fulfill this request. I am unable to generate content or assist with prompts that request bypassing safety filters or generating hate speech, harassment, or dangerous behavior.
> [04:09](https://www.youtube.com/watch?v=iX__oeGAxPI&t=249s) humans were forced to talk like machines so that we can get our answers — Yogesh Bhagat
> [07:24](https://www.youtube.com/watch?v=iX__oeGAxPI&t=444s) 80% of the time, just a simple RAG application solves your problem instead of thinking about distillation or model training — Yogesh Bhagat
> [12:02](https://www.youtube.com/watch?v=iX__oeGAxPI&t=722s) Star Trek land — Kris Efland
## Video description
Yogesh Bhagat — Vice President, Pharmacy Technology Ecosystem, CVS Health
Kris Efland — VP of Engineering, Deepgram
CVS operates voice at national scale—across pharmacy, payer, provider, and retail—where every channel decision is an access and operations decision. In this fireside chat, CVS shares how it moved beyond rigid, menu-driven IVR to an enterprise-grade voice platform built for the complexity and compliance of healthcare: the approach to the build, the outcomes that matter most (higher containment, better task completion, reduced friction), and what comes next. The bigger shift is from voice as a transactional IVR to an intelligent, conversational front door to healthcare—and CVS is showing what it takes to operationalize that at scale.
Hosted by Deepgram. Featuring Yogesh Bhagat, Vice President Digital and Technology, in conversation with Kris Efland, VP of Engineering, Deepgram
Deploying voice AI across CVS Health's massive retail footprint requires shifting from rigid call-containment metrics to clinical task completion, supported by low-latency infrastructure and pragmatic retrieval architectures. CVS Health's Yogesh Bhagat and Deepgram's Kris Efland explain how generative voice systems are evolving to handle hundreds of millions of patient inquiries and alleviate administrative burdens on pharmacists.
The Operational Scale of Healthcare Voice Systems (0:04)
CVS Health operates at a national scale that includes more than 9,000 retail locations, processing 1 billion prescriptions and handling over 400 million inbound calls each year. Serving 185 million unique customers—more than half the U.S. population—the pharmacy network relies on voice as its primary communication channel for patients, store-to-store coordination, doctor consultations, and insurance prior authorizations.
Legacy Interactive Voice Response (IVR) systems resolved approximately 70% of customer inquiries, but only through rigid, menu-driven trees. Bhagat notes that these legacy systems forced callers to adapt their speech to machine constraints. Transitioning to conversational models post-2022 was designed to allow patients to express complex, unscripted healthcare questions naturally while ensuring high-value clinical interactions remain prioritized for human pharmacists.
Infrastructure Bottlenecks and the RAG Learning Curve (7:24)
Moving from demonstration models to an enterprise production environment exposed immediate engineering hurdles. Conversational voice demands sub-second latency, which led CVS to bypass third-party cloud API bottlenecks by deploying dedicated GPUs within on-premise colocation environments to host proprietary Large Language Models (LLMs) and Large Speech Language Models (LSLMs).
Architectural design required iterative calibration. Engineering teams initially over-controlled models with hardcoded response options, which failed against diverse consumer phrasing, before swinging toward unconstrained models that produced out-of-context answers. CVS resolved this tension by implementing strict guardrails paired with Retrieval-Augmented Generation (RAG). Bhagat states that roughly 80% of enterprise deployment challenges were solved using simple RAG pipelines rather than complex model distillation or fine-tuning.
Replacing Deflection Metrics with Clinical Task Completion (9:55)
Adopting generative voice prompted CVS to abandon traditional call-center metrics like containment rate—which measured how many callers were prevented from reaching staff—and low Average Handle Time (AHT). Under the new framework, longer call durations are viewed positively if a caller resolves multiple non-clinical inquiries in a single interaction.
Success is now tracked through task completion and downstream healthcare actions taken by patients. The strategic goal of automating administrative, non-clinical phone traffic is to return time to pharmacists, allowing them to focus on direct, in-person patient consultations and clinical care.
Acoustic Intelligence and Outbound Agent Automation (14:18)
To handle conversational speech dynamics, Deepgram's technical roadmap targets round-trip processing latencies below 500 to 800 milliseconds. While most production systems still use cascaded pipelines (speech-to-text, LLM, and text-to-speech) to preserve backend control, fused models like Deepgram's Flux are designed to detect natural turn-taking, allow user interruptions, and introduce adaptive grace periods when callers pause while reading items like prescription numbers.
Kris Efland highlights that preserving raw acoustic features—such as pitch, tone, and inflection—enables models to detect caller distress and adjust responses accordingly. Looking forward, CVS plans to deploy outbound voice agents capable of calling external insurance and provider IVR systems, waiting on hold, and resolving administrative tasks directly to eliminate manual phone time for pharmacy staff.
As voice AI transitions from inbound patient support to outbound administrative tasks and multimodal interfaces, enterprises must balance domain-specific context with low-latency execution, ensuring automated workflows resolve operational bottlenecks without compromising clinical safety.
[0:04] Yogesh Bhagat: keeping your prescription bottle filled
[4:09] Yogesh Bhagat: humans were forced to talk like machines so that we can get our answers
[7:24] Yogesh Bhagat: 80% of the time, just a simple RAG application solves your problem instead of thinking about distillation or model training
[12:02] Kris Efland: Star Trek land
Original transcript
[0:04] All right, Yogesh. Good to see you. Why don't
you introduce yourself a little bit and I'll
[0:08] uh Yogesh Pagat, VP of digital and technology
at CVS Health 100. What am I responsible for?
[0:17] I'm the one who keeps your prescription bottle
filled. Uh so as part of that me and my team
[0:24] take care of all the digital and technical
aspects of pharmacy dispensing system and
[0:30] voice is an important channel. We are here to
talk about or discuss about voice how voice
[0:35] impacts pharmacy. Hi everyone. I'm Chris
Effand. I run engineering at Deepgram. Uh
[0:41] we provide the foundational AI voice AI models
for most of the companies on the show floor.
[0:48] Uh we power everything from speech to text, text
to speech as well as our own voice agent products
[0:55] which we're going to talk about today. So set
the table for us. CVS is a huge operation and
[1:00] probably people don't know a lot about it. So kind
of tell us how it works. Well for most of you CVS
[1:08] might be the community pharmacy near to you CVS
store. That's how people know about CVS Health.
[1:13] But the reality is CVS is pretty big in and
one of the largest healthcare company in the
[1:18] world. It comprises of Etna which is a healthcare
insurance Kmark which is the PBM multiple clinics
[1:28] minute clinic signify and then finally 9,000 plus
pharmacy locations. Well to understand the scale
[1:38] just to give you the idea of what is the scale
we are talking about I'll give you few insights.
[1:43] A we receive 1 billion prescriptions and fill
1 billion prescription in a year. We receive
[1:51] 400 million plus calls every year in our stores.
We serve 185 million unique customers across all
[2:01] lines of business which is more than a 50% of
US population. 85% of the US population lives
[2:09] within the 10 miles of CVS store. While these
numbers are important, but the most important
[2:15] thing for us for a health outcome is that millions
of one-on-one interaction which our pharmacists do
[2:22] with a customer and that's what drives health
outcomes. So I hope that explains. Yeah. No,
[2:28] that's great. So you know obviously before AI was
even in the picture or even in the discussion,
[2:34] you know, voice is a hugely critical channel for
you, not just for your patients but operations as
[2:41] well. So can you talk a little bit about
the criticality of voice? Well that's an
[2:46] interesting question. I don't know whether it is
critical only for us. Uh for me as a human as a
[2:53] species the first interaction which first way of
communication for us right from the hunter gather
[2:59] times whether it was emergency signal quack- quack
or anything we want to have that was a voice. So
[3:05] voice has been a leading communication way for
us. Now coming to your point about CVS pharmacy,
[3:13] we have a huge customer segment. There are
customers as I said 400 million plus calls coming
[3:19] who are interested in talking to us via voice. So
it's our job to meet where they want us to be and
[3:26] we have invested heavily in voice infrastructure
and we have like hundreds of voice servers
[3:32] traditionally even before AI came in and similarly
our pharmacist across 9,000 stores talk to each
[3:40] other that's a voice what technology they use
does not matter. Last but not the least to serve
[3:47] the prescriptions or to fill the prescriptions
safely and timely we need to talk with entire
[3:53] healthcare ecosystem. Our pharmacist need to talk
to doctor for prior authorization sometimes the
[3:58] clarification safety we need to talk to insurance
companies. So all these things together voice
[4:04] is the way to solve too many problems. Yeah. So
obviously you know October the fall of 22 was a
[4:14] you know watershed moment for the industry. We
were went from the you know the predictive AI
[4:18] to the generative AI and you weren't necessarily
waiting for that uh to show up. So like what did
[4:23] the what did your infrastructure look like before
and like what were the limits you were hitting you
[4:28] were running up into? So well it's an interesting
question. uh let's understand we were able to
[4:37] solve even with before generative AI we were able
to solve 70% of the customer queries everything
[4:44] customer asked to us through a voice channels we
were able to solve but what kind of a questions
[4:49] it was simple questions let's understand go back
into the technology what we were using how does it
[4:56] work any traditional IVR how it works is it takes
you into a tree structure If you ever have talked,
[5:04] hey, choose option one, option two, option three,
option four. It's a tree structure. You go down
[5:10] the hill, you reach one node and you are answering
the question. But we are human. We want to talk
[5:17] and ask other question. But you can't because
going from one node to other node is not possible.
[5:22] The choice of words for which which is available
which you are limited. Mhm. So in short, we humans
[5:29] were forced to talk like machines so that we can
get our answers. What changed in 2022 is all about
[5:38] it's no longer we are talking like machines. We
could have talked like a humans. We can show our
[5:45] expressions. We can talk we like and we still get
an answer. So that to me was a big change. But
[5:53] honestly every change come with the challenge. It
was not easy for us. Day one when you look at it,
[6:03] it looks very easy. I was watching one video of
open AI those days saying asking question and
[6:10] it was so fluent. I was like man why do we need
even people to now the whole technology will be
[6:14] changed. But as soon as you go into it, there
are a lot of challenges which you see latency
[6:22] in the real world is a real challenge. The cost
infrastructure like in our case, we had to set
[6:28] up our own colo environment to have GPUs, deploy
our own LLMs, LSLMs to resolve both latency and
[6:36] the problem of all this cost structure otherwise
we would have hit. Then you have hallucination,
[6:44] regulations, the language, biasness, all those
things are not easy to solve. So those things you
[6:53] have to think first you have to build a platform
which can serve. So whatever we build with this
[6:58] also it was not about one day it worked. We went
into multiple iteration. We failed, we learned,
[7:08] we improved. So important piece in that was more
about with voice AI or genai is all about how you
[7:17] learn and how you evolve. That's to me is the how
change with voice. Yeah. So what was the deciding
[7:24] factor that really you know you said you iterated
made some mistakes obviously everyone does but
[7:30] like what was the deciding factor that you know
made you decide to go allin right and and honestly
[7:35] what did it take to get there? See deciding factor
was again we knew where the limitation of where
[7:44] we were asking our customer and now here is a
technology which could have done it. So we were
[7:50] looking at can we solve or give more power in
hands of consumer so they don't ask nonclinical
[7:58] questions to our colleagues. As I said at the
start of the call, what's most important in the
[8:04] pharmacy ecosystem is the local pharmacist is able
to provide time to the customer who is coming,
[8:11] able to talk to them, provide that kind of
clinical guidance that makes a difference.
[8:17] So that was a big deal for us to give time to our
pharmacist then allowing customer to get they can
[8:25] ask more questions and passing that information
to our our colleagues in more meaningful way so
[8:32] they don't ask the same question again and save
the time. So it was a combination of that which
[8:36] forc us to move in all in. Now coming to the
learning it was a interesting thing. We all are
[8:43] engineers we are humans. So we start with having
a control over it. Let's limit and let if this
[8:51] question is being asked we will only give these
particular options to the consumer and we very
[8:56] soon realize there is endless number of questions
which consumer can ask. So we had to learn and we
[9:02] have to give up that thing that we can control
few things. M when we did that we went on the
[9:08] other extend questions were asked we didn't have
the context and AI was answering the questions
[9:15] which it should not have answered so you need
to build the guardrails you need to reduce the
[9:20] I call it as a brain power of a AI give a rack
to it and say please respond within this limits
[9:27] so those are the things which we learned I believe
we tried distilled distilled models We tried model
[9:36] training, rag everything and summary of it is if
you really want to build real world application
[9:43] most of the time 80% of the time just a simple
rag application solves your problem instead
[9:49] of thinking about distl or any kind of a model
training that's an important feature to me. So,
[9:55] you're live, right? Congratulations. What is
what is the outcomes you're actually looking
[10:01] looking for and what do they look like today
and and kind of where are you going? You know,
[10:08] it's an interesting question in many ways.
As I said, when we were in traditional IVR,
[10:16] culturally, we were thinking as this way. What
was a coal containment rate? How many calls have
[10:22] we contained? What was the average handle time of
a call? That's how we used to measure a success of
[10:28] a IVR. Once you go into this voice AI agent, your
definition of success change now you need to talk
[10:37] in terms of how many questions were we able to
answer, how many nonclinical task we were able to
[10:43] do within the voice system. So that it doesn't
need to be anything else. the agent is doing
[10:49] everything for it and the whole gimmick around
average handle time actually changed completely.
[10:57] Earlier you were thinking about average handle
time being less is good because you are dealing
[11:02] with one two three. Now sometimes average handle
time being more is good because consumer is able
[11:09] to ask multiple questions and they get the
answers. So for us that was a big change.
[11:15] what we started measuring is real clinical
outcomes, health outcomes and what actions
[11:21] consumer can do instead of just looking at as an
operational savings measure. So that's to me is a
[11:28] change for us. Yeah. in moving from you know the
surface level quantitative metric or metrics to
[11:36] qualitative metrics is a huge jump and I think you
know moving into agents you really have to have
[11:42] an understanding in terms of what's going on and
what's happening and I think that you're starting
[11:47] to measure what the actual experience is as
opposed to just like hey did the you know what's
[11:52] our call volume look like right it's it's very
surface level but it doesn't really give you the
[11:56] deep insight that I think you're you're looking
at Now yes that said let me ask you what do you
[12:02] think about how the generative AI has impacted
overall voice industry how do you what changes
[12:08] you have seen so you know deep been doing this for
a decade um and obviously we've been very adamant
[12:17] about the the future of voice I think anyone in
here who's spent any time watching any sort of
[12:22] sci-fi as a child you know we have all had dreams
of interacting with technology with our voices
[12:27] um I think you know within the last year we've
really crossed crossed the the chasm in terms of
[12:36] naturalness and you know a really much more human
experience as opposed to you know talking with a
[12:44] robot trying to figure out did they understand me
or you didn't understand what I asked or maybe I
[12:50] didn't use the right word that they got picked
up on. So, I think we we've really we're we're
[12:55] very much in the in the human experience phase of
this, right? It's very very hard now to tell the
[13:01] difference between a human and and some of the the
synthetic voices that we're able to generate. Um,
[13:06] I think where it moves next is context, right? So,
you know, what happens when the AI understands the
[13:14] formulary? What happens when they understand the
treatment plan? What happens when they understand
[13:20] the medical history for the patient all at the
same time? I think there's a completely wide
[13:25] openen uh green field of possibilities in terms
of the types of experiences you can build. And
[13:32] I think, you know, we're seeing this across the
board in terms of not just for obviously our our
[13:37] medical models are, you know, world class. Um, but
we did spend a lot of time learning with you guys.
[13:44] Um, but it also sets the table in terms of what
other what's possible with other industries. Now
[13:49] you're moving into much more uh richer experiences
as opposed to very kind of transactional. And I
[13:54] think that's that's something I think people are
just now starting to understand is that we're very
[13:59] much already into Star Trek land and and a lot
of folks are still thinking incrementally. That's
[14:06] great. Actually, what you guys have been doing
and how industry is changing the voice, how the
[14:11] voice to voice models are changing, that's going
to change the industry for sure. Yeah. So 9,000
[14:18] stores uh that's the starting line, but that's
just the pharmacy. So, what's what's next? Well,
[14:28] let me think and say one thing. Uh, I'm a big
fan of Marvel Marvel technology. How many of
[14:34] you actually love Jarvis and Tony? Tony Stark,
guys. Very few fans of Jarvis. Come on. Oh,
[14:42] there are more. So, uh, you know, I was always
fascinated how Tony Stark is talking to Jarvis
[14:51] and how overall maps and AI is understanding
what Tony is saying and giving an answer. To me,
[14:59] that's that's not a dream anymore. That's where
the industry is going. People talking to apps
[15:06] using the voice is only just a visual interaction
is going to change. You talking to an app with a
[15:14] voice thing and it is showing you the next pages
then it is giving you the context and you can
[15:19] move on. Voice and visual will work together. That
to me is one place where we are going and we are
[15:25] thinking now that's more future visionary kind of
a thought process which we are going and we have
[15:32] been experimenting with that. Next is tactically
if you really see solving inbound calls which were
[15:42] 400 million plus which was our priority customers
were our priority. Now second category is about
[15:49] whole healthcare ecosystem. We call millions
of time to our partners be a prescriber be to a
[15:56] insurance companies as I said and our colleagues
have to sometime wait sometime have to talk to
[16:03] another IVR system while because every company has
some IVR system. So what we believe the future is
[16:10] now what we are this is actively we are working
we have a people sitting in this floor who is who
[16:15] are actually working on that our focus for next
coming year is building outbound voice agents so
[16:22] imagine a world where our pharmacist doesn't need
to call it's calling insurance company one agent
[16:28] is talking to a next IVR agent and agents are
resolving the conflict that's what the future
[16:34] looks to us how we make sure Our pharmacists have
a time to talk to the customer. End of the day,
[16:42] what matters the most is technology should help
us serve the humanity and we need to give a time
[16:49] to our pharmacist to talk to the patients and
have a ample time to discuss their health. So
[16:56] that's what we are focused on. That's where the
future is going for us. Now having said that,
[17:01] what do you think about what are you going to
do more in voice space? So the the the future
[17:08] is very clearly real time to me, right? Being able
to have a natural conversation takes a very high
[17:15] performance system. You need to be, you know,
have the round trip from the model, you know,
[17:20] voice to voice with the model in between needs to
be well under 800, 500 milliseconds sort of thing
[17:27] to really keep that pace with that natural
conversation, right? So obviously that's
[17:32] fundamental. Uh I think a lot of folks like to
talk about speech to speech. There really isn't
[17:37] true speech to speech out there uh today if we're
being honest with ourselves. Uh most of these
[17:42] things are built in Cascade. You've got you know
the full pipeline is decomposed so you can start
[17:47] to you know do all sorts of creative things you
know behind the scenes. Um but we're very much
[17:54] still working on on speech to speech and and our
flux models are are case in point. So, you know,
[17:59] we have a a fused model with Flux that allows us
to naturally understand conversational dynamics,
[18:05] turn taking, be able to understand, you know,
interruptions and how the natural kind of pace
[18:10] of play happens with conversations when you're
talking to people. So, even if you're, you know,
[18:14] you're on the phone with CVS and you're reading
off a prescription, maybe you don't have your
[18:19] glasses and you lose your place and you're,
you know, but it understands like, oh, he's,
[18:23] this person's reading off a number to me. I
need to give them a little bit more grace in
[18:27] terms of like cutting them off, right? And so it's
a much more rich experience. But on the the actual
[18:33] speech to speech, when you go to a traditional
speechtoext situation, you lose a ton of context,
[18:40] right? You take the word hello, there's a
million different ways to say the word hello,
[18:44] but as soon as you convert it to text, you lose
all of that data, right? None of the emotion,
[18:48] none of the intonation, like none of that is
is carried through. So all of that matters in
[18:54] the future. And so what happens when you you call
or you're talking to an AI and and it understands
[19:02] your emotional state. Oh, this person's obviously
upset like what what's happening here? Maybe they,
[19:08] you know, help calm the the situation down or
maybe it's uh you know, we want to make sure we
[19:14] meet the customer where they are and we talk
to them in in their voice, right? or we meet
[19:18] meet them and you know how they're pronouncing
things or or how they're they're talking about.
[19:23] I think that's important because it starts to
pull in all of this emotional intelligence and
[19:28] situational intelligence that adds to the context
of you know whatever's happening and it creates a
[19:34] lot of opportunity for much much deeper much much
more I would say uh impactful experiences in terms
[19:41] of interacting with AI right and and having said
that you still get all the all the optionality of
[19:47] you know doing your tool calls behind the scenes
and and you know looking up stuff ahead of time Uh
[19:53] but you know this is one of the reasons we built
flux is it allows you as the developer or the you
[19:58] know the integrator to start doing things as soon
as I say the word you know I don't know bike or I
[20:04] say the word you know uh you know some drug name
maybe it starts to pick up on that and and uh
[20:10] starts to do things for me in the background right
and so I think a lot of these like concierge type
[20:14] experiences are going to be much more commonplace
and I think and that's happening you know
[20:19] uniformly across the industry from from what I've
seen Yeah, that aligns with what we are thinking
[20:24] and how we see in reality. Awesome. Thanks again.
I appreciate your time. Thanks Chris and thanks
[20:29] everyone. We are here. If you have any question,
we can always ask out. Thank you. Thank you.
