The traditional FSA claims process has long been a source of frustration for employees everywhere. You pay out of pocket for a healthcare expense, submit your receipt, wait days or even weeks for a human reviewer to examine the documentation, and then often receive a request for additional information that restarts the entire cycle. Prashant Desale, Senior Vice President of Technology at WEX, recognized this friction as an opportunity to fundamentally reimagine how benefits administration actually works.
WEX, an enterprise technology solutions company that simplifies the business of running a business across fleet management, corporate payments, and health benefits, has effectively developed an AI-powered claims adjudication system that delivers results in seconds rather than days or weeks. The innovation earned WEX recognition with a BIG Innovation Award, and the story behind its development offers valuable lessons for any organization looking to deploy AI in meaningful and measurable ways.
The Real Problem with Traditional Claims Processing
On paper, FSA claims processing seems straightforward enough. An employee receives healthcare services, pays out of pocket, submits a claim with documentation, and waits for reimbursement. In practice, the experience has historically been anything but simple. According to research from the Employee Benefit Research Institute, Americans hold over $116 billion in health savings and flexible spending accounts, yet many users find the reimbursement process so cumbersome that they leave money on the table rather than deal with the hassle.
The traditional workflow typically involves manual document review by HR administrators or operations teams. When documentation is incomplete or unclear, reviewers request additional information, adding days or weeks to the process. This iterative back and forth creates frustration for employees who are essentially waiting to receive their own money back. As Prashant explains, the key insight was recognizing that speed alone would not solve the problem. The real goal was predictability, meaning that users should know upfront exactly what they need to submit and receive validation, and claim decision in seconds.
Eyes, Brain, and Judgment: The Technical Architecture
WEX designed their AI claims system around three interconnected components that Prashant describes as eyes, brain, and judgment. The eyes function uses Optical Character Recognition technology to extract text from medical receipts and documentation. It’s important because receipts are sometimes crumbled, faded ink, or with handwritten notes on them. This extraction happens immediately with a high degree of accuracy when users submit their claims, providing instant validation that they have included the necessary information.
The brain component takes the extracted text and combines it with the actual receipt image, submitting both to a large language model that builds context around the submission. This is particularly important when dealing with pharmacy receipts that might include both eligible medical expenses and ineligible items like candy purchased at the same time. The AI actually differentiates between what qualifies as a medical expense and what does not.
The judgment layer runs multiple AI agents in parallel to determine eligibility against IRS codes and specific account types like FSA or HSA. These agents evaluate line items, service dates, amounts, and eligibility criteria simultaneously. Most importantly, Prashant notes that the team purposely trained the system to be unsure rather than forcing decisions. When the AI is not 100% confident in an approval or denial, it routes the claim to a human reviewer. This human-centered design philosophy prioritizes trust over speed, which is especially critical in the regulated healthcare benefits space.
Building Production Software in Six Months
Perhaps the most remarkable aspect of the WEX innovation is the development timeline. The team took the project from initial idea to full production deployment in approximately six months with just eight team members. While AI tools certainly accelerated some aspects of development, Prasahnt credits the success primarily to thorough problem discovery and cross-functional alignment. The real breakthrough was how an aligned team worked together to solve the one of the largest friction.
The team spent considerable time at the outset truly understanding the problem they wanted to solve. Rather than starting with technology and looking for AI Models, they identified the highest-friction user experience in their portfolio and worked backward. This approach meant that product managers, engineers, data scientists, operations specialists, and compliance experts all aligned around a common goal from the beginning.
Prashant emphasizes that solving problems like this is perhaps only 20% technology. The remaining 80% involves understanding the problem deeply, aligning different functions and team members toward a shared objective, and creating genuine accountability across the organization. The shared joy of transforming an experience that positively impacts human lives became a powerful motivating force for the small team.
Building Trust Through Human-Centered AI Design
The WEX approach reflects a broader philosophy about how AI should function in regulated industries and sensitive applications. Rather than deploying AI to replace human judgment entirely, the system augments human decision-making while maintaining appropriate guardrails. Trust is non-negotiable in this space, and building that trust requires designing systems where humans remain at the center of consequential decisions.
Industry data from Gartner suggests that 45% of organizations are now testing generative AI in customer-facing functions, yet many projects stall in pilot phases because they start with the wrong question. Prashant advises organizations to avoid asking how to use AI and instead ask what the topmost friction is that users experience. Starting from that user-centered question leads to more meaningful implementations.
The WEX team is now looking to replicate this pattern across other workflows in their benefits ecosystem. They are building what Prasahnt calls an Agentic-AI mesh where AI agents move beyond answering questions to actually completing tasks for users. A simple example involves payment processing, where the AI would not just tell users when their next payment is due but offer to complete the payment right there if they have a card on file.
Lessons for Leaders Deploying AI in Regulated Environments
For technology leaders considering similar AI deployments, Prasahnt offers three key recommendations. First, start with friction rather than technology. Identify the user pain point that would have the most meaningful positive impact on people's lives and work backward from there. Second, align teams around shared goals and accountability from the beginning. Cross-functional collaboration is essential when projects touch product, engineering, operations, and compliance simultaneously.
Third, make trust a non-negotiable requirement rather than a nice-to-have feature. In regulated environments especially, but really in any user-facing application, building confidence in the system requires demonstrating that accuracy matters more than speed. When users know they can rely on the system to get things right, adoption follows naturally.
The WEX innovation ultimately reflects a broader shift in how enterprises are thinking about AI deployment. Rather than chasing headline-grabbing capabilities, the most impactful applications often focus on removing friction from everyday experiences that affect millions of people. The goal, as Prashant puts it, is to help employees use their benefits to make their lives and their families' lives better. When technology disappears into the background and users simply feel that something was really easy to use, that is when AI has truly succeeded.
Enjoying insights from industry leaders? Subscribe to The Winners' Circle podcast on your favorite podcast player and never miss an episode. Listen and subscribe at bintelligence.com/podcast.