AI That Pays: Lessons from Revenue Cycle — Nathan Wan, Ensemble Health

So great to be here today. Excited to talk to you about a little bit of the healthcare system that often gets overlooked. It's part of a healthcare system that actually continues to grow in multiple dimensions. Over the past couple of decades, its size, cost, and complexity outpace many other benchmarks.

And that's becauseright now, 40% of hospitals operate at a negative margin. Let me put it another way: almost half the hospitals in the country are losing money. And it's not because of the clinical costs. It's because of the broken and manual processes around the revenue cycle.

There's delays, denials, a lot of rework, a lot of lost revenue. My name is Nathan. I'm the head of AI at Ensemble Health Partners, and we work with hundreds of hospitals and health systems in the U.S. to manage the revenue cycle.

We're about 14,000 people, and we've been a leader on the quality side within the industry.

As an end-to-end solution, that means we support the entire process, every stage, and it gives us a really unique lens into all the problems and inefficiencies that occur throughout the entire process, and also an opportunity to stop them before they happen.

Revenue Cycle Management, or RCM, refers to the financial process of the patient's journey within the healthcare system. And it's traditionally thought of as a series of steps that goes from one to the next.

And just a little bit about how I got here. I started my career in tech, working at Google, building software for operational teams, and then working on speech recognition and language modeling. Back in my day, which is now like over 10 years ago, we were comparing language models that took traditional models with Google scale data and compared them to deep learning models, trying to make them compete with each other and see which one worked better.

But one of the really interesting projects that we worked on was called what is now called Ambient. This is one of the things where people are trying to use this technology to improve the administrative burden

for doctors. That's because oftentimes doctors will spend hours after they see a patient writing, documenting, and creating notes for themselves.

While we weren't successful back then, today there's multiple projects and multiple groups launching this and making it commercially viable. And so it's been really successful and really exciting to see that change.

Then I spent a little time in the world of startups, in the world of biotech. I changed both the scale and the domain that I worked in, and it was very exciting. I was really excited to work on a very strong mission, a really exciting mission where we had a big opportunity to make a big impact.

I started first in diagnostics. This is where I built models and built teams to detect cancer from blood. The goal was to give early insight into whether or not a patient had cancer or not. We used machine learning to look at the blood and look for biomarkers and look across multiple data sets and patients to identify where might be the signal for cancer.

After spending some time there and seeing the company grow from 30 people to over 300 people, I ended up at an even smaller therapeutic startup. We worked on novel data sets, looking for unique interactions in complex microbiome communities to try to identify compounds that could be unique and really valuable in drug discovery.

And that's the thing. When most people think about AI in healthcare, these are the things that most people think of,right? There's diagnostics, there's imaging, there's other ways to improve documentation and decision-making for clinicians and drug discovery. And these are really impactful and really important problems.

I really enjoyed working on them, but they are some of the toughest and hardest to solve. But that's also because some of the benefits are going to be massive as we continue to see groups and organizations crack parts of it and make headway.

We're seeing a lot of change already. But there is another part of the system, the healthcare system, that's not nearly as flashy, doesn't nearly have the same attention on research or in media

that also has a huge measurable real impact and is also ripe for AI disruption. And that's the healthcare, the financial side of healthcare. Right now, we estimate that 20% of the GDP is attributed to the healthcare system. A large proportion of that is the administration of healthcare.

This is billing, insurance-related things, but it starts from the very beginning through to the very end of the patient's journey with any healthcare provider, starting with eligibility checks, registration. There's documentation, medical coding, denial management, so on and so forth.

Oftentimes, these aren't things that you'll see when you visit a doctor's office, not unless you encounter a really challenging situation or you lack coverage or you face some complication with your care. But this process is very complex because it is very manual, it is very rules-driven, and very inconsistent.

And over the past three decades, this healthcare administration in general, the number of people working in healthcare administration has increased 30-fold. But in the same time period, the number of clinicians has barely doubled. So it just goes to show how much faster, how much more complex, how much more quickly this area grows compared to the clinical side of healthcare.

And just another note about terminology. I'll keep using these words: patients, providers, and payers. Patients, I think we all can understand those are the people who receive care. Providers are the ones who deliver it, people like the hospitals, the specialty offices, the specialists, nurses, and doctors that actually conduct and provide care, medical care to the patients.

And then payers are those who provide the funding. So largely insurance companies, which would be private payers, but also other government institutions like Medicare and Medicaid.

And just to help make that more concrete, the cost and complexity of healthcare is really correlated. We estimate a large amount of the cost associated with healthcare is actually related to friction. And in this case, friction means the inefficiencies around communicating back and forth between payers and providers and patients.

And often, a lot of that actually results

in the same outcome,right? Either the claim gets paid or it doesn't get paid. One of the things we'll talk about is denials, because that's one of the biggest components of friction. And that's because it's both time and money for the provider to manage, appeal, and work through that process.

And then with, again, very slim margins for these hospital systems, any impact or any change in the denial rate can have a huge impact. And ultimately, AI has a big opportunity to shift resources from this bureaucracy, this friction, towards hopefully something else,right?

Something more productive, we might agree, would be like the clinical care or anything else that we've described. So just to make it concrete for you guys, this is an example of what a claim might look like and also how much conversation occurs between the patient, well, largely the providers and the payers, before, during, and after a patient visit or patient encounter.

In this case, this claim was denied four separate times and appealed four separate times as well. The provider had to send documentation multiple times through multiple interfaces and probably through multiple different manual processes. And for the provider, they didn't receive payment for 200 days until after the procedure occurred.

And

with AI happening across the field, providers aren't the only one looking to AI to make and improve their process. Payers are also thinking about it as well. They've also increased their denial rates. They're leveraging AI as well for increasing the volume that they're able to adjudicate and to identify issues for denials, making this entire process

more strenuous and creating a much larger backlog for all these providers. And the thing is that most of these denials don't necessarily require a better appeal system. It's not like they need a smarter appeal system. They just need a way to avoid errors that cause these in the first place.

That is, most of the time, they're not necessarily medical agreements. They're just technical errors in registration or missing data that if we were to put them together in theright way the first time around, we could avoid a lot of this friction.

So how, at Ensemble, are we hoping to be able to solve this problem? We think because we are an end-to-end RCM organization, full-service provider, we have an opportunity to see the longitudinal data, connect the dots from the very beginning of the process to the very end of the process, and really make a change before the error occurs.

One of the first situations I'll talk about is prior auth. This is an issue that affects the entire industry. Prior auths occur because the payers have required providers to ask for permission for certain procedures. But it's truly challenging because it's often not clear when a prior auth is required.

You sometimes have to go to the payer portal and say, does this procedure require prior authorization? And even when you do, sometimes they still might deny it because it was incorrect or

just the policy had changed. And I think this is where we really think we have an opportunity to change, to basically correct the error before it happens, because we can see that data from, see all the historical data from the beginning part where prior auths are requested and acquired, to the end of the process where we see the final denials.

Where AI can help, not only can we try to predict denial, we can also try to identify and correct the denial. So an example is if we see certain procedures and we know that often another denial reason is that the procedure was missing from the original document.

And so we can try to flag that early and say, if you're looking at these procedures, you might actually want this other one. And finally, even the actual process of acquiring prior authorization is a big opportunity. It's a manual process.

It requires documentation from different parts of the system to be put together by someone and sent off to the payers to make that request for prior authorization.

But sometimes we're not as successful. Sometimes it still may be the case that

denials may still occur. We can't always avoid denials,

but we're excited because GenAI really has an opportunity to help us accelerate and improve that process as well.

The case study I'll talk aboutright now is called clinical denials. And this occurs when the payer and the provider disagree about what was medically necessary to care for the patient.

And when this happens, the provider has, in order for them to appeal, they have to go through a process where they have to build the entire appeal packet. They'll need to look through the patient record. They'll look through guidelines or standard guidelines of care to identify

what care should have been provided to the patient. They look through payer policies to see what would or would not be covered. And this is all a very time-consuming process. Some of these EMRs or electronic medical records are hundreds of pages long.

They have all types of data in them: text, images, labs, notes, tables.

The clinical guidelines themselves have hundreds of clinical guidelines. And for different situations, you'll need different guidelines. And all this is done under tight deadlines to make sure that you respond in time to the payer after the denial. And all this means that there's a very real and limiting factor of how many expert clinicians can we get into the process to help us build and generate these denials, appeals, excuse me.

As you might expect, GenAI can actually generate an appeal letter for you. An off-the-shelf one can, if you prompt it, will give you some appeal letter. But unfortunately, that alone wasn't sufficient. We found that when we worked directly with our clinical experts, that off-the-shelf models alone wasn't sufficient.

And so we really worked hand in hand to develop a model and a pipeline that allows not only the letter to meet the standard, the quality standard that we have as an organization, but also to allow the clinical expert to make the final decision on whether or not the letter meets the standard of quality before it gets submitted to the payer.

And this is important because there is also complexity around the clinical appeal process. There's different service lines, different clients, and all that gets put together in our GenAI system to make sure that we can deliver these appeals more quickly and more consistently.

We've seen already that we're increasing the speed of the process, a 40% reduction in time, sometimes even more. We've also seen higher quality. We've been able to measure quality in terms of the overturn rate. How often are we seeing

denials being overturned? And we have also seen, as a result, the volume grow. But one thing I really want to point out here is that as part of this operational and service team, we're able to really measure the ROI directly.

It's not just like hand-wavy, this is value. We're tracking it very specifically and measuring it very concretely. And that's one of the really exciting things about bringing AI to this RCM process.

So I know I won't purport to say that AI will be able to solve all of the problems overnight. This is an industry that's been reliant on a lot of processes for a long time. As I said, there's lots of rules.

They're inconsistent. It's unstructured. You'll see data scattered across a wide range of systems. And this is one of the things that makes it really challenging to bring together as a cohesive and unified process. Ensemble has spent and invested

a lot already in building up a single consistent infrastructure to be able to do this. And one of the reasons we have been successful has been because of the platform that we call EIQ, where we bring together multiple formats, multiple data formats within a single platform.

But obviously, there's still great opportunity to be able to do that. You'll see that EMRs have many different format types, and it will challenge any multimodal LLM to parse correctly.

We're excited because we already see AI deliver value, as you saw with the clinical appeal letter case, but also in the prior authorization case. We're continuing to build agents for all aspects of the revenue cycle process. But we know automation alone isn't going to be enough.

There's complexity in revenue cycle that clicking buttons and pushing

pieces through faster might not be the only way to do it,right? There's really an aspect of reasoning and connectivity that we think about when we think about how to take errors that occur at the end of the process, like the appeal process or denial process, and try to fix them upstream and early on.

And what we're really hoping for and we're really excited about is being able to not just build better tooling, but also a smarter, more coordinated system that allows us to reduce waste in the overall revenue cycle process.

So that's why I'm really excited to be at Ensemble. I think we have a unique position to lead this transformation. We have theright data set. We've been building theright team as well to bring all the experts from multiple disciplines to achieve this goal.

And we have the full scope of the RCM process to not only collect the data, but also intervene and act on behalf of our clients. I thank you all so much for your attention. I hope this gives you a new way to think about AI in healthcare.

And if you have a chance, please find me and connect with me afterwards. Thanks so much.