Fawad is CEO and co-founder of Penguin Ai, a healthcare AI startup that sells to payers, providers and revenue cycle management companies.
At one point in my career, I had a fair amount of oversight of U.S. healthcare claims. Billions of transactions a year moved through systems at Kaiser Permanente, UnitedHealthcare and Optum. What always alarmed me was how much of the work existed only because the systems underneath couldn’t talk to each other. Thousands of people spent their days stitching together information that lived in five different places, interpreting rules that changed depending on which system they were in.
The U.S. spends roughly $1 trillion a year on healthcare administration. We tend to treat it as a spending problem. We think it’s something to solve with better tools or faster workflows. But after two decades managing the data infrastructure underneath those workflows, I see it differently. The operating model itself produces the spending, and it always has.
The Model Was Built This Way
Claims move across multiple disconnected systems before anyone can act on them. Prior authorizations make the problem worse, since they require manual handoffs between clinical and administrative teams who often work from different information. Underneath all of it, coders spend hours reconciling documentation that was captured in one format but needed in another. Each of these workflows was designed independently, and each one assumes a person will connect what the systems can’t.
I managed teams that did this kind of work. You’d automate one handoff, and two more would surface downstream. You’d consolidate a data feed and discover three other feeds still running on legacy logic nobody had mapped. The system runs on fragmentation and human interpretation, and that’s what makes the $1 trillion figure so persistent. The industry has improved efficiency for decades. It just improved efficiency inside a model that structurally produces more administrative overhead.
AI Exposes The System
AI depends on consistent inputs and clearly defined paths. Healthcare operations offer neither. Data arrives in different formats depending on which system generated it. Context that a human staffer would carry in their head from one step to the next doesn’t exist in any structured form for an AI agent to pick up.
Look at the denial and appeals process. Providers are deploying AI to auto-generate appeals for denied claims. Payers are deploying AI to process the resulting volume. Both sides are automating their half of a broken handoff, and the administrative burden between them is only growing. The underlying workflow stays exactly the same. Documentation is still fragmented, and the criteria for approval still shift depending on who reviews it. AI just makes both sides move faster through a process that was flawed to begin with.
At the scale I was operating, long before the current wave of AI tools, we’d encounter the same issues during even basic analytics projects. You’d try to build a unified view of claims performance across business units and immediately discover that the definitions didn’t match or the rules governing the same process varied by region. AI amplifies the problem by orders of magnitude, because it doesn’t sidestep inconsistencies the way people do. It just stops. Or worse, it confidently produces output built on the wrong inputs, and someone downstream has to catch it.
The System Has To Do the Work
Most healthcare technology was built to record what happened. It stored the data and tracked the status, moving work from one queue to the next. Execution still lived with people, and that made sense when the systems themselves were the constraint.
But with AI, the goal now is end-to-end execution. For that to work, context has to carry through from one step to the next without resetting every time it crosses a system boundary. The workflow itself has to hold together. That means consolidation of vendors and a unified execution layer where work can actually be completed, start to finish.
For years, the industry has focused on improving parts of the process. But improving parts doesn’t create a system. The trillion-dollar cost structure has always pointed to a design problem, and AI has made it undeniable.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
