Host Intro (Gene Kim)

Gene Kim: I'm delighted to introduce our first speakers of the morning, who come to us from WEX, a global technology leader in financial technology. They have revenue last year of nearly $2.6 billion. They operate payment platforms and software solutions that help businesses manage fleet fuel cards, employee benefits administration, corporate expense management, and more.

I'm so delighted that Prashant Desale, their SVP of Health and Benefits Technology, and Jennifer Whitmire, head of Product, Health and Benefits, are going to be presenting an experience report. There are two reasons why I love this talk. One is that, just like the Grainger talk yesterday, it's a joint presentation from an engineering leader and a product leader describing how they span that organizational boundary. Another reason is that they're presenting on their attempts to use AI to solve an especially thorny problem around their benefits claims processing process.

I suspect after hearing the story, you may find certain approval processes that you run or have to use in your organization just a little less tolerable and be inspired to do something about it. Here's Prashant and Jennifer.

Prashant Desale

Prashant Desale: Good morning, everyone. That was quite an enthusiastic response. Hopefully everybody had coffee.

We are absolutely delighted to be here. Last year, there was one statistic that I recently came across: all the companies around the globe spent $250 billion on AI initiatives. The kicker was that more than 80% of those AI initiatives never made it beyond the pilot phase. If you think about it, billions of dollars spent, thousands of teams working on it, and most of it goes nowhere.

I could have made up these stats, which I did, but all of this is not far from reality. At WEX, when we were thinking about leveraging AI, we wanted to beat these odds. How we successfully did that is the story that we are going to share with you today.

Jennifer Whitmire

Jennifer Whitmire: Hi, everyone. Just as important, this is a story of partnership. I'm Jennifer Whitmire. I'm Senior Vice President of Product, head of Benefits for WEX. We're here to tell the story of how we turned AI from a set of models into a foundation of business value, and more importantly, how we did it in a way that puts people at the center. Our employees, our HR teams, our consumers are at the center of this experience.

Prashant Desale

Prashant Desale: I'm Prashant Desale, Senior Vice President of Technology. Jennifer and I are amazing partners. You can see we are color-coordinated on our outfits as well. But this is not about technology, nor is it about product. It is about collaboration. It's about how we brought our teams together to solve one of the most foundational problems that our customers face, and that's what we are going to share with you today.

Jennifer Whitmire

Jennifer Whitmire: For those that might not know who WEX is, our mission is simple but very powerful: we simplify the business of running a business. We are about 6,500 employees. We are in 16 countries and have about 40 years of proven success.

We operate at the complex intersection of data and payments in three areas: mobility, corporate payments, and health and benefits, so employee benefits essentially. In mobility, we help fleets manage fueling and logistics for thousands of vehicles. In corporate payments, we streamline vendor payments globally. For employee benefits, which we'll focus on here today, we support millions and millions of employees managing their financial and healthcare wellness.

Prashant Desale

Prashant Desale: This mission to simplify the business of running a business is what led us to build an ecosystem that's both vast and powerful, from fleet management to corporate payments to health and benefits. The area that we're going to focus on today is employee benefits.

When you think about employee benefits, employee benefits are complex and personal. The reason they are complex and personal is because when you're touching employee benefits, you're not just touching transactions, you're touching people's lives. Employees at different companies need certain benefits. Their family members need certain benefits. When you are touching something, you are actually touching people's lives. When we were thinking about leveraging AI, we thought this would be a perfect proving ground for us to build responsible and human-centered AI.

Jennifer Whitmire

Jennifer Whitmire: For those who might wonder where Prashant and Jennifer sit, one of our key goals at WEX was pioneering for the future. Prashant reports up to the chief technology officer. I report to the chief digital officer. Together, this partnership was super critical as we looked in the organization at how we're going to innovate and how we're going to bring new technology utilizing AI and pioneering for what we're doing.

Prashant Desale

Prashant Desale: Every leader knows the promise of AI, but we also know all of the barriers that we all face. Data is either siloed or very inconsistent. Delivery cycles are super fragile, and our feedback loops are usually missing. We all know that AI is 20% technology and 80% about organizational change. It's behavioral economics and anything you ever look at. Does this sound familiar?

We've all been there. We all know that data is messy. Data is problematic. A lot of delivery cycles break. Users get frustrated. If you have bad data or messy data and you feed that bad data to AI systems in a hope that AI systems are going to turn that into insights, it doesn't. If you feed bad data to AI systems, it turns that into really, really confident nonsense. It's like feeding your child Red Bull for dinner. If you do that, the outcome is going to be loud, messy, unpredictable. That's what happens when you feed bad data to AI systems.

When we were thinking about building something leveraging AI at WEX, we knew that if we focused only on technology or model, we were going to fail. Instead, what we thought of focusing on was building a system for the flow that we believe would make it successful. The flow that we went with takes highly governed data, and I'm emphasizing highly governed because we wanted to build trust from the get-go, feed that to the platform that we are building, have that platform leverage the secure processes that we have in place, and then drive highly trusted outcomes for our customers.

In order to do that, we grounded ourselves on three key principles. Number one, we actually prioritized fast feedback over perfection. Instead of taking a model and trying to perfect that model in isolation, we prioritized getting it out there quickly: deploy it, learn, iterate, deploy, learn, iterate. That's what we prioritized over perfecting something in isolation.

Second, we decided to focus on simplifying the human workflows. We strongly believe that technology should reduce the cognitive load of humans as opposed to adding to it.

Third, instead of making it a technology problem or product team problem, we aligned all the teams toward a common solution or common goal. Teams like Jennifer's team, my technology team, our sales team, product team, marketing team, everybody came together to think about what we want to do and how we want to solve it.

Jennifer Whitmire

Jennifer Whitmire: With these principles, we asked ourselves: where do we start? We looked at a process that was manual, painful, super frustrating for everyone, not only internal in the organization but also for our consumers in our business. It was really, really clear. It came down to our flexible spending accounts.

Most of you out in the audience all have benefits. You probably have some type of tax-deferred accounts, and if you've ever filed one of these, you totally know it can feel like applying for a mortgage 20 years ago when it was tons of paper. You're doing this manual process today to try to get back money you've put out of pocket, like $40 for Band-Aids.

Here's what used to happen. I'll call each of you the consumer. You would go out, spend the money, and need to file a claim. Then you file that claim. It's going to sit in a manual queue for someone to review. If that document isn't perfect, it's going to get denied. Then you're going to get sent a letter, and most of you get them in the mail. That still happens today. We have lots of mail going out around this. You might be a lucky one that gets an email or a text, very few of those.

Then you're sent that letter, you receive it, you're going to call into the call center, and you're going to talk about why you're frustrated and what you need to do differently. You're going to file that claim again. It's going to go back through this same process. You're going to submit and wait and wait, sometimes weeks.

Very frustrating. This isn't efficient. It was a burden not only for our internal teams, but also for HR teams and employers and employees that are waiting, as a consumer, on your money. You just want to get reimbursed for an item you've spent out of pocket and that you already have ready to get reimbursed.

Prashant Desale

Prashant Desale: To solve the problem that Jennifer just described, what we launched is a first AI-powered tool, what we call intelligent claims automation for FSA. The goal of this tool is to eliminate every human interaction or every human effort from claims processing. The target customers of this tool were the overburdened HR employees at different companies, the teams that actually spend a lot of energy manually processing the claims that are submitted by employees, and employees themselves who are looking for instant and accurate reimbursements for the claims that they submit.

This tool does three things. One, it automatically verifies the documents that are submitted for every claim. Second, it actually auto-populates the forms for employees to submit using the document that they have just uploaded. Third and most important, it actually automatically adjudicates the decision whether to approve or deny the claim.

This entire tool was built in seven months, from conception to production, by a team of eight team members from different parts of our organization.

Jennifer Whitmire

Jennifer Whitmire: I want to let you see this in action. It's going to extract the data, the relevant information. It's going to check against any plan rules and the plan designs that are loaded into the system. It's going to determine eligibility instantly, and it's going to complete this claim in seconds. Just keep in mind, this went from days, sometimes weeks, to seconds.

Demo Narration

Demo narration: For many of us, filing a health claim can be a frustrating experience. Let's see how Crystal's journey is being transformed from a tedious chore into a simple, fast, and predictable process.

The old way was a multi-step hassle. Crystal could snap a picture of the documentation with her phone, but then she had to manually fill out a long, tedious form and attach the photo. Then she had to wait for days to see if it was approved.

Now the experience is transformed by intelligent automation. Crystal starts a new claim by clicking "reimburse myself," ready to get it done in just a few easy steps. First, she confirms which account to use and where the reimbursement should be sent. The simple, clear options let her direct her money with ease.

Instead of typing, Crystal simply uploads the photo of the documentation. Our AI-powered technology instantly validates that the receipt has the information needed to approve that claim, creating the claim and populating all the necessary details, boosting her confidence through this real-time feedback that ensures her claim is complete and accurate the first time.

Crystal quickly reviews the auto-filled claim, provider, service date, and amount and sees that everything is perfect. A task that used to take minutes of typing is now just a quick glance and a click. With a final click, the claim is submitted. Instead of waiting days for approval, she'll have an answer in minutes.

Crystal feels a wave of relief, happy that a once long and uncertain process is now fast and predictable. This is the power of a smarter system. By proactively identifying the need for a claim and using AI-powered technology to handle the details, we give Crystal back her time and remove the mental burden of managing her benefits. It's not just a faster process. It's a seamless experience that lets her move on with her day, certain that everything is already being taken care of.

Jennifer Whitmire

Jennifer Whitmire: These results have been so transformative, both for internal and our external customers. We've had a 90% reduction in processing time, high accuracy in our reimbursement determination, and most importantly, we've had happy consumers, our external consumers and our internal ones. This isn't just about efficiency. It's about trust. Consumers that feel cared for and get their money fast, without a hassle, learn to trust this and learn to trust the process.

Prashant Desale

Prashant Desale: Here is a testimonial from somebody who uses this day in and day out. This is a testimonial from Chantel Hollen, who is a director of claims processing organization. She and her team predominantly process claims manually. Until this tool was launched.

Ever since the launch, they have been very successful using this tool. According to Chantel, they have processed more than 25,000 claims that otherwise they would have processed manually using this tool with 98%-plus accuracy. What that means is that 98% of the 25,000-plus claims were adjudicated automatically by this tool. Only 2% actually went into the human review for her team to handle. According to her, it's not just an upgrade for her team, it's a strategic investment.

All this doesn't happen by magic. Just to give you a little glimpse behind the scenes how it works: our architecture has three core principles. One is governed data. We focused a lot on governing the data that we feed to the system, just to make sure that trust is built in from the get-go. The second key pillar of our architecture was making sure that it's API-driven to make sure that we are ready for flexibility and scale in the future. The third, one of the most important pillars, is that we kept the human in the design. If AI for whatever reason is not confident enough to make a decision on a certain claim, then it's going to put that in a queue for a human to review. Those are the 2% that I was referring to earlier.

A little bit more detail on the technology architecture behind the scenes. On the left side, the red box shows where the claims submitter submits the documents to our claims processing microservice. These documents are generally submitted as images or PDF documents, and they are generally pharmacy receipts or hospital bills. They are submitted to the microservice. That microservice submits that to Microsoft Azure OCR. For people who don't know what Microsoft Azure OCR is, it's an optical character recognition service that Microsoft provides that allows you to extract text from images or documents.

We are also testing Gemini right now to see if there is higher accuracy. What OCR does is extract all the text from the documents, and then it allows us to reduce hallucinations. From that point, we take that text combined with the documents and send it to what we call an extraction module. That extraction module is an AI agent that uses GPT-4 to decipher the line items on every receipt. A pharmacy receipt can have five line items, 10 line items, 200 line items. Same thing with hospital bills.

Once it deciphers every line item, every line item is structured as a JSON object that goes to different agents. We use LangGraph for orchestrating all of our agents today. Think of LangGraph as an air traffic controller, so all the AI agents don't collide with each other.

All these AI agents take those line items and run multiple parallel adjudication paths. One agent could be validating the FSA eligibility of a particular line item. We store all the FSA-eligible items as vectors in our vector database, so one agent might be checking that. A second agent might be checking the totals against the billing from the hospital bill. A third agent might be checking the presence of an FSA tag on the receipt. There are multiple agents that work in parallel, and they provide a score.

At the end, there is one aggregated score. That score determines two things. First, it determines whether AI should make the adjudication decision or not. If the answer is yes, it makes the decision to either approve or deny the claim. If the answer is no, it simply puts that in the review queue. At the end, it sends through an audit process where it compares with past outcomes, makes sure that the current result is in compliance with past outcomes, and also the system keeps learning.

Essentially, it is a combination of OCR, GPT-4, our validation logic, our RAG pipeline, our vector database, all that has created this system that is fast and accurate. On average, we spend somewhere between 10 to 15 seconds to make a decision on one claim. We're trying to bring that down, but that's where currently we are. That's the technology architecture at a high level.

Jennifer Whitmire

Jennifer Whitmire: We deliberately built this around Gene Kim's five ideals. We achieved locality and simplicity in our design. We created flow, focus, and joy for our end users and consumers. We drove real improvement of daily work for our HR teams. Most importantly, we fostered psychological safety and a system built on trust while maintaining a relentless focus on our customers.

When you're building an AI tool, what is the single biggest lesson we can all take away from this? It's trust. Governance is the foundation of trust. We tier models by risk. We validate independently. We monitor for bias and drift, which is super critical, and we bake in privacy and security from the start.

I like to point this out because it was a joke in a couple of other sessions, I believe yesterday, when we talked about there not being many compliance or security people in the room. Typically when we're developing new products, what do you do? You forget to leave those individuals out until the very end, and you have problems with adoption. We brought security to the table at the very, very beginning, especially in this process. We wanted to make sure that everything we did was built.

Yes, we talked about fewer checklists, and we did that having security at the front, because doing that, no one has ever innovated from a 200-slide deck on compliance. We've all been in there. This isn't compliance theater. This is just how we have earned our trust and adoption for this tool.

Ultimately, once trust is unlocked, users need three things. They need explainability; they want to understand why. They want reliability; they want to know that it works every time they go to it. And agency; they want to have control over it and when they use it. When you deliver on these things, adoption is never a struggle. It is natural and they trust it.

Prashant Desale

Prashant Desale: Here is a playbook that worked for us and we would recommend to everybody who is thinking about leveraging AI in your product lines. It's a very simple three-step playbook that we strongly believe will work across the board.

First, find a friction that is meaningful for you to solve for your customers. This friction could be something that brings a very high value or high-value return to your customers. Your customer could be an internal developer, your partner, or your end consumer. Whoever it is, find a friction that is meaningful to be solved.

Second, measure the human impact of your solution. AI is not just about efficiency, accuracy, and cost reduction. We strongly believe AI should be leveraged to bring a positive impact on human life. Measure the impact of your solution on human life.

Third, make sure trust is built in as a non-negotiable KPI from day one on the solution that you're thinking about.

We are not experts and we're not done. We are still learning. That's why we are also here to learn from all of you. There are certain things that we are still trying to figure out. Number one, everything that we have built now is on more modern technology, but we have a lot of legacy systems that are built 20 years ago, 15 years ago. We're trying to figure out how to embed AI into our legacy systems, and most importantly, how do we embed the trust into the legacy system?

Second, we are building a lot of AI agents now, not just to make decisions for the users, but actually to take actions on behalf of the users. What we are trying to figure out is how do we govern all these different AI agents to make sure that trust is never broken?

Jennifer Whitmire

Jennifer Whitmire: Leadership itself must evolve. What does it mean to manage trust at large scale for organizations? There have been sessions, and I really hope that we all can learn from each other about how we start to lead an organization into the AI age, because we have to lead it at scale and it's got to have the trust of the individuals. We don't have all the answers, but we really hope that we can help and learn from each other while we're here.

Prashant Desale

Prashant Desale: With that, we'll leave you with one final thought. AI is not going to transform enterprises because you have a complex model or complex technology. It's going to transform enterprises if you focus on solving problems by designing a system for the flow that you have. Keep the human in the center of your design and make trust a non-negotiable KPI, or the foundational element for your design.

Remember at the beginning of the presentation I shared one stat: $250 billion plus 80% of those projects fail. Now I'm happy to report that our project is part of the remaining 20% that succeeded, and this time I'm not making it up. That's something that we have tested success now with what we have done.

Jennifer Whitmire

Jennifer Whitmire: Absolutely. Thank you. If I think about our AI, it's always available. Our solution doesn't need a coffee break. It doesn't need a motivational speaker or a poster. It's just there, and it always works, and it's trust. Both our internal and external customers greatly appreciate that.

Everyone can breathe a sigh of relief. Thank you so much, and thanks for coming.

Prashant Desale

Prashant Desale: Thank you again. There's a link to our WEX technology blog. We publish a lot of technical information on it. Please follow, and if you have any questions, please ask ChatGPT. We'll do the same thing. All right, thank you.