From Weeks to Seconds: AI-Powered Benefits Claims at WEX with Prashant Desale, SVP of Technology

Russ Fordyce 🏆 (00:01.569) Welcome back to the Winner's Circle. We are here with Parshant from Wex, and they're going to talk to us, or he's going to talk to us, about how as corporate citizens we can get our money back for health expenses just a little bit faster than we used to. Parshant, welcome to the Winner's Circle, and congratulations on the Big Innovation Award. Prashant Desale (00:19.753) Thank you very much, Russ. First of all, thank you for having me here. And really pleasure to be connecting with you and your listeners. Just to introduce myself, my name is Prashant Desle. I'm a senior vice president of technology at VEX. Predominantly, I lead the health and benefits engineering organization. Just for your listeners who may or may not know about VEX, let me set context. So VEX is an enterprise technology solutions company. We operate across multiple segments and our mission at VEX is simple but powerful. We simplify the business of running a business. And we operate across multiple segments like fleet, mobility management, corporate payments, and health and benefits. And we try to simplify the workflows. We work behind the scenes so the enterprises can focus on what really matters to them. So that's the intro about VEX and for the area that I work in, it's a health and benefit space that's where we build the top notch technology software solutions for the health and benefit space where, you know, in the context of conversation that we're going to have, my role is particularly about taking a high friction complex problem that our users face like... the claims, processing that you just referred, taking problem like that, aligning all of our teams, like our product team, technology team, operations team, compliance team, get everybody together to solve that problem for our university and meaningful measurable. That's what I try to focus on. At high level, our solutions are used by millions of people. to manage their HSA, their FSA, their, you know, benefits administration, COBRA, even investing the HSA dollars in a meaningful way for in different stages of people's careers and their life moments. I'm particularly motivated when technology disappears in background and creates a user experience where the users actually feel, that was really easy to use. Russ Fordyce 🏆 (02:39.721) You know, I'm so used to like corporations borrowing my money, right? We we lend it to them all the time, whether it's, you know, expense reports, and we're driving around in our cars using our gas. you know, it's funny, because, you know, I was at one large company who, you know, I think elevated a vice president once because he saved the company $35 million in interest by doing some some just changes to process. And so you're you it's We're talking about large dollar values here and I love what you've done with the FSA piece because you've shortened that cycle from kind of money out of pocket to money in hand and we're not floating banks money, the banks aren't floating us money and we're not floating the corporation money as much anymore. It's really kind of negligible with what you guys have done. So walk us through a little bit about kind of that customer experience once they're on the Wex platform and how you go from like two weeks to two days I think you guys are talking about. Prashant Desale (03:36.01) Well, it's better than two days, but let me walk you through, First of all, you already know the FSA claims process. On paper, it looks easy, but in practice, sometimes it's not. And FSA is just a starting point. We are going to expand that into different plant types, different account types, HSAs and all that. But at high level for listeners who may or may not know what claims is, what we are referring here. The claim that we are referring here is when an employee of any employer takes a health care service and pays out-of-pocket expense that that employee wants to now get reimbursed and they submit a something-or-other claim to get that reimbursement back. So that process, as I said, sounds simple on paper, but sometimes it's actually quite challenging and frustrating for the employees because the process is very... go and submit the claim while submitting the claim. You upload your pharmacy receipts, your medical receipts. And then the process is that once you submit the claim, it goes to a human review, meaning if somebody in HR administration or operations team is actually looking at that documentation, and if they are not satisfied with the documentation, they'll come back after days or weeks for you to provide more information. And then the process again repeats. So you see, right? It's a very complicated process. Russ Fordyce 🏆 (04:59.415) Yeah. And it's very iterative. And every time you step into it, seems to add a day or a week or something to it, just because it is complex and humans have cues and workflows happen. So you want to get it right the first time. That's the whole idea. How did you technically make that shift? mean, AI helps a lot. But how did you technically make that shift to days, to weeks, to days? I guess weeks to days to whatever you're going to tell us. mean, because I'm sorry, the accuracy level has to be extraordinary. Otherwise, it doesn't work. Prashant Desale (05:31.207) Yes, sir. Prashant Desale (05:38.154) Absolutely. And you hit the nail there. It's not about speed. It's about predictability in the system, meaning user knowing upfront that what I'm expected to submit and validate that information in seconds right there before user actually submits the claim. That's first thing that we try to solve with AI. Second thing that we try to solve is actually making that adjudication decision. So just to bring a little bit at high level. The way it works in the analogy that I can use is eyes, brain, and judgment. So we use AI technology, something called OCR, and OCR stands for Optical Characteristic Mission. That is just to extract the text from the medical receipts that somebody is submitting along with the claim. When we extract that text, right there we validate it. We already have the accurate information. So user knows upfront if they have submitted. information or not. Then the second is the brain, where we take that extracted information and to your point about, hey, it needs to be accurate, right? So there should be no hallucinations in the game. So what we do is we take that extracted text, we combine that with the image of the receipt, and we submit that to live memory model to actually build the context around it, meaning what are the different line items on the receipt. what differentiating between what is medical expense versus what is non-medical expense. Because when you go to pharmacy, you may be buying candy and the prescription at the same time, same receipt, right? So it does that. And then we run some very advanced AI agents to do adjudication. They run in parallel paths to determine the eligibility against the IRS codes, or eligibility against the FSC, HSC. what items are eligible or not, the service stage, the amounts, all that. And then it comes to a decision. The most important factor, as I would say, we have trained system, or we have purposely trained the system to be unsure, more than actually trying to make decision. And the reason for that is, know, speed is good, but accuracy matters a lot more. So if AI system is not 100 % sure that... Prashant Desale (08:00.294) it should approve or deny claim, it actually puts in a human right. And that's a human centered design that we focus more on because trust is lot more important than speed in this space. Russ Fordyce 🏆 (08:13.091) Well, what you guys have done is really cool. It's going from kind of that, I guess, transactional model and really building something to solve a problem and solve that problem for a customer. What's astounding to me, I think I read in your nomination is you actually, you guys built this tool, this entire end-to-end application in something like under a quarter. And that's crazy to me now in the AI world, I guess that's a little bit easier, but I'd like to hear a little bit about that. And how did you move that quickly from kind of that idea into production? Give us a little sense of how that happened in under a quarter. Prashant Desale (08:53.353) Absolutely. I think it was discovery started a little bit before quarter. So end to end, I would say we had a team of around eight team members that took it from literally starting with idea to production in around six to six and a half months. And what ended up happening is we first spent a lot of time on discovery of the problem that we tried to solve. Because for us, technology is not that hard to get. More than focusing on, what AI model should we use? What data option we use? What technology should we use? We spend a lot of time on figuring it out. What is the most important friction that we want to remove from users? And we found this one to be the most important friction that is going to help both the employees and HR partners. So we focused on that. And from there, we quickly moved on our data centers, our team members are pretty good at figuring out the solution. We aligned different teams together and we value, we spend a lot of time in more than building, we spend a lot of time in knowledge. We wanted to make sure that it's creating accurate solutions. I have to be honest, actually, even after deploying in production for one month or more than a month, we were taking every decision it made even automatically. to a human. We wanted to make sure that is making accurate and right decisions. So hopefully I'm answering your question that it was more about aligning everybody together and moving quickly by shortening the distance. Russ Fordyce 🏆 (10:30.827) No, I love it. And AI just accelerates that. I'm finding that in our work with this business is you're still limited by human capacity, by the brain powers, the fingers, the time that we have to spend to this. AI gives you this limitless engine, but you have to narrow it down sometimes just to get stuff done. And one of the things that I I was really intrigued by was you're doing this in a very heavily regulated environment, right? It's you're in healthcare, you're in insurance, you're spanning, you know, I guess you get a foot on both docs and, you know, you're straddling like Jean-Claude Van Damme in the finance world and the healthcare world. Where does that, you know, kind of risk management and compliance compatibility fit into all this, you know, all this application? Prashant Desale (11:23.919) You're absolutely right, Russ. We operate in a highly regulated environment. We have different different compliance that we need to stay true to. And HIPAA is one of them. we operate in an environment where it's not just compliance. It's more about the users' trust on us that we are doing the right thing and making sure that our HR partners, other partners, our employees of the employees who are our customers trust us. by making sure that the data is protected at all costs. So to answer your question, we have a very mature governance framework in the organization. And that framework essentially tiers every AI system in three different tiers. And the tiering is done based on the intent of the user, the complexity of the model, the data that it tries to use, also the management of the line saw. And so there are different parameters that it uses to tier the AI systems. If any AI system is a severe risk, this is a risk tiering by the way, if it's a severe risk, then that's just not allowed. We just don't deploy that. don't pursue the building of that or productionizing that. If it's a high risk, then generally we put really a of guardrails around documentation requirement, the independent review requirement, the very frequent assessment requirements from independent parties, the model validation. Even before deployment, models go through a thorough validation. After deployment, there is a very advanced monitoring of the models. We also make sure models are designed and developed in such a way that there is a traceability with auditability. So when a human reviews any decision that is made by AI, a human can actually get to the bottom of why AI made that decision in the way it did. And that is to make sure that there is we need to serve the data within. Same thing with data. We operate in a highly regulated environment, so we have to comply with very high data privacy and protection laws. So there is authorization, authentication, authorized access to specific data. So all in all, there is a very mature framework around governance and data protection. Russ Fordyce 🏆 (13:42.211) Well, yeah, as a consumer, I'm certainly glad that those protections are in there, right? And you guys are following the compliance rules. Because I think what's interesting is, I guess, as a citizen, I'm proud of that HIPAA regulation. As much as I don't like signing all those forms that we have to in doctor's offices, it seems to be doing its job fairly well. And it seems to be flexible enough to allow you guys to do things within the bounds of those of those broad kind of constructs to move in there. when that legislation was written, they weren't thinking about AI. So it's really impressive that it's lasted this long and it's been nimble enough to kind of maneuver the new world that we have. What I'd like to hear from you is a little bit about these outcomes. I know we keep touching on it a little bit, but it was a two week process at least, I would say. At least I haven't submitted FSA in a long time. few years, it was longer than two weeks. Walk us through kind of what it looked like before and what your customers kind of typically see as outcomes. Prashant Desale (14:50.633) Absolutely. So in terms of outcomes, there are two major outcomes that I would walk you through. First is for the employees. In the past, your employee was submitting a claim, was days to weeks process. Employees did not know what to submit and when they submitted, there were questions and more processes. So now, after we rolled out this, that days or weeks of wait has come down to seconds. So literally, experience now is very trustable and predictable for any user because as a user now, you submit the information while submitting. AI tells you if you are submitting the right information or not. So there is upfront validation right there. And then second is once you submit the claim, within 10 to 15 seconds, you will have an answer whether your claim is approved. or it's declined. But there are also cases where if it's not approved or declined, AI is not certain, then it puts in the human review for the queue. But since now we have removed the cognitive workload on the human reviewers, that turnaround time also has increased dramatically. So long story short, for employees, the experience has become more predictable, more trustworthy, and that helps them now that build their confidence to use the system. And our goal is to encourage the users to use the system because people should be using their benefits to make their and their family's life better. So that's the really important outcome for us. Russ Fordyce 🏆 (16:28.688) Yeah, and as an employee, I have seen plenty of corporate systems where we were software houses effectively. then you look behind the scenes at internal systems, and you're like, we're really running this into inside here. So a well-designed and employee-friendly system is super, super important. You as a leader have kind of moved this team of eight into this new world. What did you learn about leadership in this transition as you guys kind of accelerated deployment? Prashant Desale (17:05.257) Yeah, and this team is a small team, but they made a huge difference here. As a leader, what I learned is that solving problems like this is only 20 % technology. 80 % is about understanding the problem that you want to solve and aligning different functions and team members towards a common goal and making sure everybody understands that, hey, there's a shared accountability here. There's a shared goal here. And most importantly, There's a shared joy of transforming experience that actually makes a positive impact on human life. So that's what I learned from this particular experience. Russ Fordyce 🏆 (17:47.812) Well, I think you're hitting on something that's think near and dear to everybody's heart right now, which is we're going to go through this time when AI can do certain tasks and certain jobs and everybody's nervous about what's next and how do I play a role. And you've uncovered something really important. And I know it's such an important thing for anybody building out tools, which is we have to identify the problem. And I know. Salespeople, the good ones are really good at this. What's the problem that business is having? And they sell into that because they're solving a problem. They're solving a challenge for that business. That's going to be around with AI or not. Somebody's still going to need to uncover those problems and then go look for solutions, whether it's AI or something else. How are you guys evolving in that kind of area of really looking at that kind of customer solution type of angle? Prashant Desale (18:42.985) I'll probably tell you what our true north is. Our true north that we're trying to chase is removing or eliminating all the friction from benefits solutions. And the reason for doing that, we inherently believe that benefits are designed for employees to use to make sure that they and their family members live happy and happy lives. And in order to do that, they need to have confidence in the system and that confidence will come only if we reduce the friction and we make the system very easy to use for them. So that's where we're trying to leverage AI. But AI by technology itself is advancing it. So many new advancements are happening. And our goal is to leverage AI in a human centered design. And when I say human centered, meaning we always keep human as the center of decision making. and leverage AI to accelerate and bring more accuracy to decision making. So the way we are evolving will evolve the technology and the pattern that we have learned here about extracting information, validating of data, know, deterministic verification of information and then relying on humans to still make sure that everything is controlled properly. That pattern we are going to replicate on all the workflows now. There are a lot of other frictions in the system that we need to solve for. We are going to use this pattern to solve. Russ Fordyce 🏆 (20:14.871) Well, and without revealing too much, give us a little preview of kind of that broader AI roadmap that you guys are looking at inside of WAC. Prashant Desale (20:18.153) Thanks. Prashant Desale (20:25.129) So many things, we are building what we call agent-ai mesh, essentially now, see, agents are moving from just answering questions to actually doing things for you. So we believe that software should not just guide people, it should do things for people. And for example, I'll give you a simple example. If somebody's payment is due, then if somebody logs in and just asks question that, hey, when is my next payment due? It should not just tell when my next payment is due. It should actually ask counter question, hey, do have a card on file. Do you want me to go and process the payment for you? And that's a very simple example. There are a lot of other examples like that. But idea is to take now AI from just answering the questions to actually doing the things and removing that cognitive workload from employees as well as the job administrators and partners. Russ Fordyce 🏆 (21:18.103) Well, that's an interesting example you give because I always gave that advice to employees was if when somebody emails you a question, you should consider the question of why they're asking it. Because it's not a ping pong. Let's read into it a little bit, get some context. And now AI is great at that. You know context about an event, know context about a receipt or an expense. And so you can ask the right questions. I'm glad you guys are getting there. That's really, really interesting. As a leader in this kind of space, in AI and responsibility in this regulated industry, if you were kind of to give somebody some advice on how to accelerate kind of that rollout and get something moving this quickly, what would it be? Prashant Desale (22:04.361) I would say, my advice, a lot of times I feel that lot of AI projects are stuck in the pilot phase. That's because you start with the question, how do I use AI? And I feel that that's a wrong question to start with. The better question to start with is, what is the topmost friction that my users have that I should be solving for? You start with that question, find the friction that is more meaningful. to solve for your users and have positive impact on human life, then what you do is you align your teams around that object, that goal of solving that. And you make sure that trust, especially in a regulated environment, but not beyond regulation, it's also important for building the trust with users, is that build the trust from get go. Meaning trust in the system should be non-negotiable requirement. I think if you do those three things, things will fall in place and you will use AI for really innovative solutions in your organization. That would be my recommendation. Russ Fordyce 🏆 (23:14.999) Well, that's great advice, Parshon. I really appreciate you taking the time to walk us through. am so impressed that you're doing this. I love these niche applications that really go at the heart of really kind of pain in our existence a little bit. you're making it a little bit better for everybody. So I appreciate that. And congratulations on all your success. Prashant Desale (23:34.612) Thank you so much, Russ, and thank you for having me.