Hello, Is Anyone There? Reimagining Healthcare Scheduling with AI

This transcript was generated automatically and may contain errors.

Melissa Gilkes-Smith: Welcome back to Touch Points, Moments That Matter, the podcast about the practical side of patient experience, operational excellence, and the moments that shape how patients feel about the care they receive. Today we’re working inside the second pillar of the nine pillars of patient experience framework, inquiry and scheduling. This is the first live human, or in today’s conversation, potentially AI, interaction a patient has with your practice. And it sets the tone for everything that follows. My guest today comes at this problem from an angle I don’t think we fully explored on this show yet, the technology angle, not the EHR angle, not the portal angle, the AI voice agent angle. The idea that the first voice a patient hears when they call your practice might not be human at all, and that this might actually be better for patients and staff alike.

Matthew Rayette is the founder of Nobaflow, a company building AI voice agent specifically designed to streamline call center operations in healthcare. He studied applied mathematics and biomedical engineering at Emory University in Georgia Tech, which means he brings both a technical foundation to build intelligent systems in the domain knowledge to understand the clinical context they operate in. Matthew, welcome to Touch Points.

Mathieu Rihet: Thank you so much. I’m really excited to be here.

Melissa Gilkes-Smith: So before we get into the technology itself, tell me what brought you here. You have a background in applied mathematics and biomedical engineering. How does that path lead to building AI call center tools for healthcare practices?

Mathieu Rihet: Well, it all started my freshman year of college, uh actually. I was very much interested in healthcare from a young age. Uh, I received an appendectomy when I was around 14 years old and almost died on the operating table. Uh, my life was saved by a surgeon. And uh ever since then, I’ve been passionate by healthcare. I wanted to be a doctor, actually. So I came into Emory University uh wanting to be pre-med. Uh ultimately uh took multiple math courses, um, as it was also a side passion, and realized that my resources were best uh used to essentially uh create foundation models in healthcare, leveraging my mathematics experience. Um I really enjoyed the lab, uh really enjoyed conducting research.

Uh, I was looking into coronary artery disease uh research at Emory University’s uh lab uh and school of medicine. And it was very interesting to me. And as a result, uh I was working with multiple radiologists, uh, multiple physicians uh year-round. I was also a volunteer medical interpreter speaking French and Spanish. So I was really exposed to the healthcare space. And that ultimately led me to see the operational inefficiencies that existed uh within the system. And one of the most um striking examples of that that led me to that call center aspect and to building voice AI in healthcare was when I would get called at 6 a.m. by a variety of hospitals I volunteered for because there was no Spanish-speaking receptionist to translate um, you know, uh non-English speaking calls.

And I had to be there at 6 a.m. off-shift, taking those calls. I would get called at random hours in the day. And it was uh a really archaic, uh, an old uh system, as in they would pair the phone and put the phone next to the uh hospital phone, and I would interact with the patient. And I felt like there was such a better way to streamline this, um, which ultimately uh started the first idea of you know me building uh a medical translator for those um, you know, health systems and uh hospitals that I was working at. And ultimately the idea grew and grew into um me exploring more and more inefficiencies that could be uh addressed by Voice AI. And uh that led us to where I am today. We currently automate a variety of workflows with uh voice AI that involve appointment scheduling, rescheduling, uh cancellation recovery, patient discharge follow-ups uh calls, uh referral management, outbound referral management, and so on.

Melissa Gilkes-Smith: Very cool. I love how your personal experience in a call center and the inefficiencies that you saw, it it created your solution. You’re very solution-oriented person, I can tell by what you just said. So before we get to the solution, though, I want to start with the problem because I think it’s important for our listeners to sit with how significant the gap is. So, in your work building technology for healthcare call centers, what does the current state of patient scheduling actually look like? What are you seeing?

Mathieu Rihet: I think we can divide the problem into two uh criteria. The first is a volume problem, uh, and the second is a revenue leaking problem. Uh, and I’ll start with the first. I think uh staff are ultimately overwhelmed uh by the amount of calls that they’re getting. You have a tailwind of an aging population, uh, right, and that aging population really enjoys calling. Um, and more and more people uh start caring about their care when it comes to longevity. Uh so you have a variety of groups of clinics um that were once, you know, staffing two people uh, right, to answer the calls that are now seeing hundreds of thousands of patients per month, if not, and those two staff members remain. Um, there’s a huge problem there where the staff are overwhelmed, they’re not able to take every call, and most calls can go to voicemail.

Uh, there’s also the huge problem of labor turnover among those staff. Um, it’s very complicated to, you know, manage staff, hire, fire, and uh, you know, manage that whole system. Uh, and then there’s the revenue leaking problem. There’s the voice AI that is a little bit uh, you know, uh underrepresented. We typically think of voice AI for inbound appointment scheduling, but where the revenue can really leak uh comes in outbound voice AI. So um hospitals conducting the calls themselves, calling the patient. And why that’s significant is say a patient uh doesn’t show up to their appointment. Um, it would be very important to call them and understand if they’re on the way or why they didn’t show up.

Uh, it would also be very important to then call a next patient in the list uh to schedule them so that you’re you don’t lose money on a visit and the provider’s time is spent uh with the patient. Uh that ultimately ties into the key performance indicator that uh many physicians track and many hospital groups track, which is provider utilization.

Melissa Gilkes-Smith: You really bring together like the whole operational impact. You know, it’s I constantly hear this from the patient experience side. The phone call is one of the first live impressions a patient has of your organization. And when it’s a bad experience, hold times, transfers, having to explain yourself twice, being told to call back. The patient is already building a story about what care at this practice is going to feel like. From the patient experience perspective, you know, what does, you know, what do AI call centers bring to the table?

Mathieu Rihet: Um, I think in terms of the patient experience, first, uh, I think uh everyone can agree. Uh I’m a patient myself, right? Uh, Melissa, you’re also a patient, right? When you ever call a physician group, uh, how long do you get put on hold? Right. Uh it’s a question that I asked myself. And so I timed my median. I had recently uh broken my knee uh and dislocated it too. So I did a median timing analysis of how long it took uh for me to be on hold to contact uh my surgery center. Uh it was 50 minutes. Um my gosh. Yeah. So it was a huge, uh, huge uh you know shock to me, reinforced the idea that I was building the right solution. Um and I think ultimately when it comes to uh what AI call centers can bring at the inbound scheduling level, uh it’s a simple uh performance indicator.

It’s a zero call abandonment rate, a zero percent call abandonment rate from the patient, right? The patient immediately gets their call answered, is able to schedule their appointment, get their question answered, reschedule their appointment if necessary. And right now, this is where the state of AI is at. This is what AI is able to do. But in the future, I see voice AI call centers being able to address absolutely every concern a patient has, whether it comes to their bill that they need to be that they need addressed, uh, whether it comes to uh, you know, finding the referring physician and like getting to know their availability for this specific time slot uh and booking multiple appointments at once on the same phone call.

Um, just a variety of use cases that AI cannot currently address that it will be able to address in the future. Prescription refills it already addresses as well, which is really exciting.

Melissa Gilkes-Smith: So it this is why I frame inquiry and scheduling as a patient experience pillar, not just an operations function. Every decision you make about how patients reach you is a patient experience decision. The whole music is a patient experience decision, the hours you answer phones, and increasingly whether a human or an AI answers that call is a patient experience decision. So let’s talk about what that decision actually involves. So I want to spend some time here just making sure our listeners understand what we’re actually talking about because there’s a lot of confusion sometimes. When I say AI voice agent, I’m not talking about a phone tree that makes you press one for appointments. I’m not talking about a chat bot on a website.

What is an AI voice agent and what makes Nova Flow’s approach specific to healthcare?

Mathieu Rihet: Uh yeah, absolutely. So an AI voice agent is speaking and operating like a human. That’s the way I see it. So it answers the phone. Uh, you don’t have to press a keypad to be transferred to a receptionist. Uh, it speaks to you, asks you what you’re in what you’re calling for, and is able to coordinate uh based on its LLM model reasoning. That’s the way I see a voice agent. Now, what makes Novaflow’s approach different and specific to my company is that typically those uh voice AI agents are connected through APIs to the electronic health record system. And from a non-technical standpoint, that just means uh it’s connected through data uh that is at the back end of the electronic health record system.

And this can limit uh operations uh for a multitude of reasons. There is a decision tree uh that is involved when it comes to API calls, and there is only so much that data can do. Uh, our approach uh revolves around giving uh our AI uh a mouth and a keyboard. That means operating like a human, as in logging in via the electronic health record system’s front door, like a human would, looking at the electronic health record pages, understanding what they mean and interacting with them in real time while the patient is on the call.

Melissa Gilkes-Smith: So let’s let’s dive into natural language processing. So the natural language piece is, I think, what surprises people the most. These systems understand what a patient is trying to accomplish, not just words that they said. So, and that’s a meaningful distinction. Walk me through a real patient interaction. Someone calls at 9 p.m. because they just got a test result, they need to schedule a follow-up. What does that call actually sound like with an AI voice agent handling it?

Mathieu Rihet: Uh it would sound just like a human would. Uh at the end of the day, the patient calls in at 9 p.m., which by the way, mind you, is after hours. So typically a human wouldn’t answer this call. Um, the idea here is that the AI answers the call, uh, and if the patient say has a chart that they need to reschedule or follow up with, uh, the our AI agent would essentially pull that chart, understand the context uh before the patient even mentions it, ask if the information is correct for what they need to schedule a follow-up for, and then go ahead and proceed and schedule that follow-up. Uh so in fact, we’ve seen that the AI can operate faster than a human on those electronic health record pages by just clicking and improving inference.

Melissa Gilkes-Smith: And the 24-7 availability of this is huge. And so I assume, of course, there is there is some logic that is it’s logic based also. So if the AI agent is not able to handle the specific um requirements, specific requests, or based on what the practice wants the AI agent to allow it to handle, um, what does that handoff look like, you know, to a person during after hours?

Mathieu Rihet: Absolutely. So when it comes, it really depends on the practice, uh, as a matter of fact. When it’s a surgery center, typically uh we can route to the physician’s direct cell. It’s a question of transferring to the right person, right? This is also what AI can do seamlessly, is it’s able to leverage uh the information on the electronic health record system and use its own reasoning to know where to escalate and to who. All right, right. If it’s an emergency, the AI will naturally forward to a real person. If they’re in office, if they’re not out of office, it’ll encourage you to uh dial 911 or directly call the physician for specifically uh surgery use cases, um, which we’ve seen be common among uh ASEs we work with.

Melissa Gilkes-Smith: Okay. So this is fully customizable, of course, based on you know the practice.

Mathieu Rihet: Yeah, absolutely. And this is a part of our point of differentiation is that we’re really able to customize our workflow seamlessly uh by prompting the AI correctly uh to interact with the right pages on the electronic health record system. It’s ultimately a learning process. What I I like to visualize it as in once the AI is initially deployed, um, it creates uh use cases, creates volume, it uh conducts tasks. And the more tasks this AI conducts, the more feedback it we can give it. And it creates an essential feedback loop that trains it to be smarter and smarter and more operationally efficient to the point where it can reach 99% accuracy on some of these workflows. Uh and it’s truly brilliant uh at the speed at which uh it’s able to operate with even say uh 10 calls, uh 10 points of data.

Melissa Gilkes-Smith: Very cool. So I want to shift gears here a little bit because I promise our listeners a balanced conversation, and that means we have to talk about the concerns. And we all have them. I think patient experience who the practitioner who cares about what healthcare feels like to a vulnerable person should have them. Let’s start with the patient side. Who is this not right for? Who is going to have a worse experience because of an AI agent rather than a better one?

Mathieu Rihet: I think uh ultimately it comes down to escalation, right? Um, I don’t believe that uh AI uh voice agents should be operated in behavioral health facilities, uh, for example. Suppose a patient is calling uh with a crisis. Uh, I don’t believe that AI is at the uh pace or at the understanding level to fully understand when a patient is having a crisis. And I’ve seen this across a variety of case studies, the variety of research that I’ve done myself as we were looking to expand into behavioral health. Um there are many things that AI can do in behavioral health that involve back office operations, prior off, insurance. Uh, but where AI is not suited is for any patient uh facing uh a major uh emergency as well.

For example, uh say a patient says that they’re struggling to breathe, uh, that they have chest pain, uh, our AI immediately escalates to a real human. Um it’s a part of HIPAA policies, it’s the law. But it’s also a question of safety for the patient. Uh, you know, there’s still the opportunity and the um ability for an AI agent to hallucinate. Uh and this what hallucinations mean is for the AI to interpret something that’s not there or to make something up or to think incorrectly. Uh, though this happens uh very, very rarely, it can happen. And those are the moments where it is always best to uh escalate to a real person if there’s an emergency.

Melissa Gilkes-Smith: Okay. So let’s talk about data and privacy because this is healthcare. And anytime uh patient information is moving through a system, HIPAA is in the room. And what does HIPAA compliance look like for an AI voice agent? What questions should a practice be asking before they sign a contract with any vendor in this space?

Mathieu Rihet: Uh I think it’s really about security posture. Um, it’s really exploring whether the company that you’re working with uh is doing the right amount of uh pen tests every uh every so often, right? We have a policy here at Nova Flow. Uh we signed with a third-party vendor that essentially conducts a monthly pen test for us. Ash, I was just reviewing one this morning. Uh, so it’s good that you’re mentioning that. Um and uh essentially it’s just about you know maintaining uh HIPAA and SOC2 type 2 compliance. And it’s over uh 200 controls that you have to maintain to ensure that PHI is uh stored correctly or processed correctly and that there’s no revenue leak. So naturally building uh around HIPAA and SOC2 type 2 compliance is a challenge.

Uh, but we also have to understand the industry that we’re in, right? It’s healthcare, and you would rather be safe than sorry when you’re working uh with a third-party vendor, especially in such a new field, um, which is why um, you know, I think if you’re a medical practice and you’re looking to work with a vendor, you should really ask if they’re HIPAA compliant. You should ask to see their security controls, uh, you should ask to hear their disaster recovery protocol. Uh, and you should most importantly sign a business associates agreement with them uh before provisioning them any kind of access uh to patient data. And therefore, in our case, uh electronic health record systems.

Melissa Gilkes-Smith: Those are a lot of great points. So let’s say a practice leader is listening right now and they’re genuinely interested, not just like curious, but they actually are ready to move forward. What does implementation actually look like? Where does it start?

Mathieu Rihet: Well, that’s a part of what makes us unique, uh right? Implementation looks like simply provisioning us a username and password uh and read and write access on that user, like you would a typical human. Uh, and from there, what we do is we create templates across the workflows that you want automated. Um, and this can be across scheduling, uh, this can be across uh patient discharge, uh RCM, uh anything that has to do with uh operations within the electronic health record system. And those templates that we create uh allow us to virtually let the AI access the right electronic health record pages and automate a variety of tasks that can that are upon requests. And as I mentioned, this can be anything that’s RCM patient access related and it can range from the clinical setting as well.

Melissa Gilkes-Smith: So speaking of workflows, the readiness piece is something I really want to emphasize for our listeners because in my work and patient experience consulting, I see technology implemented into broken workflows constantly. The technology does not fix the workflow, it amplifies it. And if your scheduling process is already chaotic before you add AI, it will be chaotically automated after. So what does workflow clarity need to look like before you go live?

Mathieu Rihet: Um honestly, it involves having a background agent run and it involves a lot of uh, you know, uplift. What I’ve noticed is specifically is that the larger your physician group, the more uh your call center operations uh will be decentralized. And that’s normal, right? Um, the more the practices you have by definition, the more staff you have. And by definition, um decentralization uh may occur. Uh and the situation here that we’re doing uh across uh our enterprise groups and the physician groups that we’re working with is that we have um, first of all, we have if it’s a large contract, you have members in our team that directly go visit an office and look at the call center operations. We look at how they’re operating, what’s working, what’s not working.

And we then create templates uh for our AI to work around those workflows uh to essentially enhance uh the call center team rather than limiting them or uh automating a chaotic process, right? I think when it comes to how we’ve been able to develop our AI agents, um, having a background agent run and understand the processes that are currently done by staff and using our own expertise, working with other medical groups, having all that data to understand what can work across call center operations makes it truly um a solution that can improve upon itself and help staff if there’s issues with decentralization, if there are issues with uh lack of ability to schedule or uh poor operations, uh, rather than simply automating a complex process uh that becomes even more complex and uh decentralized as a result.

Melissa Gilkes-Smith: I love that. I love that. So let’s talk about staff because this is the implementation conversation I think it’s avoided most. The one about where your team thinks is happening to their jobs. How do you introduce AI to a call center team in a way that builds trust rather than fear?

Mathieu Rihet: Uh I think uh we bring in the aspect that we’re there to help them uh hold conversations that actually matter rather than uh, you know, essentially um working overtime for conversations that can easily be automated by AI today. Um I don’t believe that AI will replace uh call center teams. Uh if anything, I think it will allow AI to uh allow call center teams to operate uh in a much more efficient manner, and most importantly, actually maintain meaningful conversations with patients where call center teams would be required. And that is uh in emergency situations. That is in complex situations that require looking at billing charts, that require looking at uh physician schedules and requiring them to pair them all up to, you know, synthesize uh and uh get the expected outcome.

Melissa Gilkes-Smith: I love that reframe. AI handles the transactional so humans can own the relational. That’s exactly what the culture of care argument is for AI because the staff member who spends their day answering the same five questions and confirming appointments is not using the capacity they have for genuine human connection. And so now we see what’s possible for that individual when AI takes over like the volume of calls.

Mathieu Rihet: Yeah. It’s a volume bottle. Like I said before, there’s really two types of problems. Uh the first is revenue leaking, and the second is a question of volume, right? Uh really just calls going to voicemail. It’s an opportunity that’s lost. Uh, that’s a potential human connection or emergency situation that could have been taken care of by a real person that hasn’t been.

Melissa Gilkes-Smith: So technology in healthcare, it doesn’t quite exist in a vacuum. And this is where I think a lot of AI implementations run into trouble. You’re not dropping this into a clean slate. You’ve got an EHR, you’ve got a scheduling platform, you’ve got a patient portal, you might have a billing system that doesn’t talk to any of them. How does NovaFlow integrate into what already exists?

Mathieu Rihet: We do not believe in creating a fifth or sixth platform for a physician group to use. Uh, as in, we don’t believe that there should be, you know, a complex dashboard for call centers to operate and then feed information into the electronic health record system, the billing portal, their scheduling secondary system, et cetera, et cetera. We believe in integrating into all of those systems to make that process seamless for the call center team and the operations team. Uh, we simply uh have a dashboard, uh, an analytics dashboard uh on our page that allows uh physician groups to run the fully automated workflow and look at statistics and key performance indicators tied to that workflow from what we see on the electronic health record system.

So that is provider utilization, call abandonment rates, um conversions, no-show rates, and how they’ve improved leveraging our AI system. And to answer, to go back to your question on how we integrate with those systems is the brilliance of mimicking human behavior, right? Being provisioned a username and password to all those systems and operating on them.

Melissa Gilkes-Smith: So the EHR piece is where I do want to press a little because in my experience, EHR integration is where most healthcare technology promises go to die. Vendors say that they can integrate. What does that actually mean in practice? And what should a practice ask to find out whether the integration is real or just a checkbox?

Mathieu Rihet: Uh I think it’s really about abiding the term to the terms and conditions of uh the electronic health record system itself and making sure that you don’t uh get uh your user revoked. Uh ultimately, electronic health record systems like Epic, Athena, Cerner welcome RPA style uh integration. They welcome uh AI mimicking human behavior, um, but they need to understand that that is going on in their system uh and why that’s going on because of that security aspect that we just talked about, right? Because of maintaining hypercompliance, uh, because of adhering to their terms of service. I think it’s really important to be extremely transparent about the methods and what you do to integrate and how you go about it, because that not only allows you to reinforce the workflow you’re building with the medical practice, it also allows for complete oversight for everyone to understand what you’re doing, how you’re doing it, and how it’s benefiting the medical practice.

It can even inspire some electronic health record systems to build uh AI workflows around what the third-party vendors are already building, which I think is uh a plus. I think it’s a bonus uh to keep systems uh and some workflows electronic health record system native, as in the EHR builds it for you. I think that’s a brilliant opportunity for companies like us to find more and more workflows and to integrate with a variety of other uh systems to feed back to the electronic health record system where those automations can’t happen.

Melissa Gilkes-Smith: So once it’s live, what does the ongoing relationship with the technology look like? Because I think organizations sometimes treat implementation as an event rather than a relationship. And what does version updates, system changes, and ongoing maintenance actually require from the practice?

Mathieu Rihet: Um actually the limited overhead uh if you leverage our system, which I feel like I’m elevator pitching here, but I think the computer use uh piece of our product allows us to immediately see changes to electronic health record uh page changes, uh, essentially, which allows us to um seamlessly adapt to what the EHR changes. Uh from there, what we do have is with our growing customer base, is that we now leverage a ticketing system. So there’s a concern uh right from a physician in uh the enterprise, in the physician group, uh, right. What the physician can do is that they can open a ticket uh on our uh platform where they explain their concern or they explain that what they’re seeing and what they don’t like.

And what we do is we have our engineering team look at that ticket and fix the problem. Um most importantly, I think that uh an observability page for the um, you know, operation leaders is extremely important. I think uh again, transparency is extremely important when you’re building an AI. And for uh I want uh operations leaders, clinic operation leaders to see uh what’s going on with our AI system, what it’s doing right and what it may be not doing right, and how we can improve it and allow them to use their expertise and experience in the space to give us feedback.

Melissa Gilkes-Smith: I think this is all going to be really useful to all of our listeners, to the practices listening to this. Here’s what I want our listeners to take from this conversation. AI in the healthcare call center is not a threat to patient experience. Bad AI, carelessly implemented, is a threat to patient experience. Thoughtfully designed, properly integrated, and honestly evaluated AI is an opportunity to close one of the most persistent and frustrating gaps in the patient journey. The patient who calls at 9 p.m. and gets helped, a staff member who stops answering the same question for the 50th time and starts having real conversations, the practice that finally has capacity to do something, scheduling right because it’s not doing it all manually.

Before I let you go, what do you want healthcare leaders, practice managers, and patient experienced professionals to walk away with from this conversation?

Mathieu Rihet: Yeah, that’s an excellent question. And I think uh whether you choose to go with NovaFlow or not when it comes to automating various tasks uh in your practice, I think there are two things that you should really look after. One is the safety behind the AI agent, and two is starting small, uh letting the vendor showcase that they’re able to automate uh the process that’s uh the hair on fire problem for your medical practice, and then expanding from there. Uh the approach we take uh with medical practices that we work with is that we always like to start with a seamless workflow we can automate and then go from there, add on and develop that relationship. I think relationship building in healthcare is uh the most important uh part uh in the industry, especially in the tech and health tech industry.

I think it’s really about developing trust and establishing it and then going from there. I think the dream of every medical practice uh and most importantly, every enterprise physician group that gets approached by hundreds, if not thousands, of uh third-party vendors is to only leverage one vendor that they fully trust and go with them and ask them if they can uh automate an X and X as they grow, which is why we’re super thankful for our design partners uh that we worked with uh over the last uh almost two years now uh to really establish the scope and establish the use cases that we were going to automate first. And uh yeah.

Melissa Gilkes-Smith: That’s great. Thank you. So for our listeners, your challenge this week is simple. Call your own practice as a new patient. What happens? Is there an answer? Is it helpful? Is it warm? Does it make the next step obvious? What you experience is the gap your AI strategy either needs to close or your human strategy needs to fix. Either way, now you know. Matthew, thank you for your time, your honesty, and for building something that takes the patient experience seriously from the first line of code. I’m glad you’re here.

Mathieu Rihet: Thank you so much, Melissa. Very glad I’ve done my first podcast with you. This was great.

Melissa Gilkes-Smith: So Matthew’s information and Nova Flow’s website are in the show notes. Every interaction is a moment that matters. I’m Melissa Jilk Smith, and this is Touchpoints. If this conversation moved you, subscribe wherever you listen and connect with me on LinkedIn. I’ll see you at the next Touchpoints.