AI and the Revenue Cycle: What’s Actually Working Right Now

Welcome and Housekeeping 

Rosie Talaga: Hello everyone. This is Rosie Talaga with Becker’s Healthcare. Thank you for joining us for today’s webinar, AI and the Revenue Cycle: What’s Actually Working Right Now. 

Before we begin, I’ll walk us through a few quick housekeeping instructions. We will begin today’s webinar with a presentation and we will have time at the end of the hour for a question-and-answer session. You can submit any questions you have throughout the webinar by typing them into the Q&A box that you see on your screen. 

Today’s session is being recorded and it will be available after the event. You can use the same link that you used to log into today’s webinar to access the recording later. If at any time you have any issues with audio or visuals, please try refreshing your browser. You can also submit any technical questions into the Q&A box. We’re always here to help with that. 

Rosie Talaga: I’m pleased to introduce today’s speaker, Doug Marcy, Chief Technology Officer at Coronis Health. Thank you, Doug, for being here today, and I will now turn the floor over to you to get us started. 

Meet the Speaker 

Doug Marcy: Thank you so much, Rosie. Welcome everyone. Today is going to be sort of a continuing webinar in my little series on AI and RCM. I’m sure everyone has heard so much about AI recently that you’re getting pretty sick of it. 

The last webinar and the last time I spoke about this was really thinking about the possibilities, trying to see what was out there and all the new things that are available. This one is a little bit more about how to make that practical. How do we implement it? How do we make use of it, and what are some of the gotchas we have to watch out for? 

Doug Marcy: Just to introduce myself again: Doug Marcy, Chief Technology Officer here at Coronis. I’ve been in that role since 2023, been with the company for about 10 years now before that in various different roles in subsidiaries. 

Overall, I have 20-plus years experience implementing software and AI across a bunch of different verticals, and I studied machine learning at George Mason University. 

Today’s Agenda and the Core Message 

Doug Marcy: So when we really dig into it, what are we going to cover today? We’re going to be looking at the reality of what AI is actually producing, where value is coming from, where it’s not, and try to separate again the hype from the reality. 

We’re going to think about how to actually build a solution and what the moving parts are that go with that. I want to touch also on the economics of AI — this is a big piece in making something viable. And then what does that mean from a governance perspective, and what actions can we take? 

Doug Marcy: The main message I wanted people to take away from this is that while AI is super powerful and it can be very helpful, it’s not a replacement for human accountability. That’s the tag phrase I’ve been using over and over again: we don’t just want artificial intelligence, we want accountable intelligence. 

We still need to be able to be accountable to our customers, to their patients. The interesting thing that’s changed in the last year with AI, especially since about last November when Opus 4.5 and other things started coming out from Claude, and now we’ve had ChatGPT 5.5 — these are now super capable models. The reality is we’re no longer in the space of asking, can AI do something? It’s more around: can it do so in a way that works reliably, safely, and that we can scale without a huge impact on our profitability? 

Where AI Is Working Well Today 

Doug Marcy: The hype around AI is that it can do anything, and that’s probably getting close to true at this point, although it really can’t do everything. What we have to see is where it’s really good at something and where it maybe isn’t the best usage. 

Things we’ve seen really good use for in our claim life cycle are things like reading and classifying documents — this is a place where we’ve had a huge amount of advance in the last year. When we look at predicting and prioritizing work, combining AI with machine learning and simpler analytical tools can really give you a huge increase in productivity and efficiency. 

Generative AI and the language models are very good at drafting text, and probably too good at it in some cases. But when it comes to certain parts of our workflow, be they appeal letters or coding, AI has stepped up a lot. The important thing is you need a way to monitor the usage of AI, to monitor the quality and the capabilities coming out of it — we’re deploying AI to watch AI, to keep track of claim statuses, anomalies, and other things. 

The clearest wins we’ve seen in the last year have really been in the mid-cycle part of the workflow: data extraction, documentation validation, claims scrubbing and denial prevention, and the coding assistance piece. 

Choosing the Right AI Projects 

Doug Marcy: As we dive into AI, we have to figure out how to narrow the field. There’s always a ton of ideas like we could use AI for anything. How do we make sure that when we’re deploying something, we’re doing so in a way where putting it in production really does make sense? 

Is it high enough volume? Does it have the kinds of repeatable inputs? Does it have a very clear and traceable outcome? Are we in a place where if it does make a mistake, we have built-in processes to recover? And do we already have an established baseline so we can make sure we’re improving on our capabilities, either in volume or quality or both? 

Hallucination, Context, and Human-in-the-Loop 

Doug Marcy: The reality is that AI still does make things up. It has to deal with situations where it doesn’t have enough context, it doesn’t know what’s going on, and so it has to figure out how to fill in the blanks. 

What we found is that’s much more of a symptom of a lack of context than it is an inherent problem in the system — if it’s pushed, AI will come up with something. The trick is making sure it has the necessary information and context to understand what you’re asking it to do. We also need to make sure that whenever we’re putting an AI in place, we have some kind of escape valve, some way to hand off the work to someone with the expertise to make the call or track down the missing information. 

Doug Marcy: What I like to say is: when AI fills in the gaps, we call it hallucination. When our people fill in the gaps, we often call that expertise or experience. Those differences are critical when we talk about accountability to our clients or to our patients. 

How AI and humans interact is as important as the actual AI itself in the agentic workflow. If you don’t put in those backstops, if you don’t have some way for the AI to escalate, then you end up with an AI generating hallucinated, bad output, and you don’t have a good way to even track that it’s happening. 

The Confidence Spectrum and Accountability 

Doug Marcy: So what we’ve done across the board is really start to think of this in terms of a maturity model. Look at what we can automate, and let’s look at confidence as a concept — can we say with high confidence the AI is doing a good job? 

The confidence spectrum is where we’re really learning how to classify the work the AI is doing, make sure the work is consistently high quality, and then decide whether we can hand it off without a person, or whether it really needs a human to come in and review, or even step in and do the job themselves. 

The goal across the board is that our humans are doing the highest-value work — handling ambiguous, exceptional, or complex cases — while AI handles the mundane with oversight. We still hold our leaders accountable for the performance of their teams, whether that team is all human, or a mix of AI and human. 

Doug Marcy: I’d like to end this section with a quote I’ve used for a long time now. It comes from an IBM training manual all the way back in 1979: a computer can never be held accountable, therefore a computer must never make a management decision. 

This is probably starting to blur around the edges, and a lot more people are letting AI make judgment calls. I will say that in the testing we’ve done, we’re not there yet — we don’t really trust AI to make the critical decisions. Instead, we make sure we have solid ways for the AI and the humans responsible for everything to interact, whether that’s humans monitoring the AI’s progress, human-in-the-loop verification, or handing things off to a human to be done entirely. 

The Landscape: From Machine Learning to Frontier Models 

Doug Marcy: So how do we actually go about doing this? When we look at the opportunities we have, there’s classical machine learning — a lot of people, when they think about AI now, tend to go right to the biggest, baddest thing, but there’s still quite a lot of tooling at the lower end of the spectrum: predictive models, statistical models, regressions, cluster analysis, anomaly detection. These are still really useful. 

Deep learning is sort of the baby form of the modern AI — neural networks used for specific tasks like image recognition or classification. And then we have the big guys, the generative AI, the large language models. At the high end we have what we often call frontier LLMs — Claude Opus and ChatGPT — the ones that require huge amounts of infrastructure, hundreds of millions of dollars to run and train, trying to be super intelligence. 

Doug Marcy: We also have this world now of open-weight models. This is labs or teams making available the underlying structure of the model that you can run yourself or that can be hosted by a third party. These are often smaller and not quite as capable, though some recent ones are quite large. What that allows you to do is tailor and choose the right tool for the job — by task, by risk, by economics. 

Doug Marcy: Just taking one step back: the model is really only one part of it. It represents the most generic part of the intelligence, the engine you’re going to use to ingest and process data. But what data you feed it, how you integrate it into your workflow, what controls and monitoring you put in place, how people interact with it — all of these pieces become really important, and all of it has to be built on a strong foundation of security and privacy. 

Building the Agentic Harness 

Doug Marcy: The goal is to try to dial in the complexity of the solution to meet the needs of the task. On the low end, we still have simple business rule engines — those still make great sense in a lot of places. Predictive models, classifications, clustering, simple regressions have been working for years. 

Language models with trusted context is a really key piece of this. Making sure the model has the right information, and then giving those models legs and hands in the form of agents — letting them actually do the work, make changes, use other applications. An agent can call out to an RPA bot when it needs information from a website, or use data from a predictive model to decide what work to do next. 

Doug Marcy: Ultimately, for us, people still are the key to maintaining strong judgment and accountability for our clients. We never want to just throw technology at a problem and tell our clients, well, sorry, the AI made that up. No, we still own that. The more autonomy you give it, the better guardrails and monitoring you need. 

Doug Marcy: We have a really strong model around human-agent collaboration. The best model we’ve found is to think of these agents as really, really productive junior employees. They need guidance, they need to be monitored, they can work really fast, but they don’t always know how to handle things or exercise the best judgment. 

Case Study: Autonomous Coding 

Doug Marcy: I want to dive in a little bit on one use case where we really invested a lot in the last year, which is autonomous coding. Coronis Health is a nationwide RCM solution provider, and we work with specialties across the spectrum, from general practitioners all the way up through level 1 trauma centers, surgery, hospital facility billing, and anesthesia billing, which is a very big part of our business. 

So when we set out on this path, the sort of hypey idea was: what if we just ask Claude to do it? We went down that path, got an enterprise agreement with Claude, got a BAA set up, all of that. And frankly, we weren’t terribly impressed with the results out of the box. What we realized is that’s not a replacement for all the expertise we have — we have hundreds of years of expert experience across our different needs and workers in this space. 

The Knowledge Graph Approach 

Doug Marcy: So we ended up doing a couple of different things. One, we created for ourselves a really good data set from scratch, in agreement with our clients, to train and set precedent for a model. Then we pulled in all of the policies and subject-matter expertise, all of the business rules we apply, all of our customer-specific and payer-specific policies. 

The question was: how do we make that available so that when a model is looking at a case, it’s able to benefit from all of this accumulated experience? And the result ended up being a knowledge base — we actually call it a knowledge graph — where we put all those policies, codes, and examples, and linked them together so the model could fetch the relevant information into its context. 

Doug Marcy: This ended up being a huge breakthrough for us. Interestingly, this also became a really great tool that we were able to give to our human coders — the same information, the same context. So did the auditors, so did the more experienced coders. By creating a tool for AI, we actually created a tool for any kind of agent doing that kind of work, be it human or AI. 

Doug Marcy: We actually added another step to this process, which we call the reviewer or the adversary. Now that we’ve recommended codes and got supporting evidence, how do we say if this is good enough or not? So we created a separate evaluation that looks at the original document, looks at the recommended codes, and tries to validate the evidence it’s been given — and it gives back a confidence score. 

One interesting thing that fell out of that: when we talk about clinical documentation improvement, we started finding places where we could see improvements to the clinical documentation itself. That creates another route for us to help clients improve what they document for certain procedures, even earlier in the process than our denial prevention piece. 

Doug Marcy: As an example, we’re processing around 60,000 cases a week through our base coders right now. We have another 100,000 or so going through our computer-assisted coding platform. When we looked at bringing in a frontier model to do this, it actually cost more than our people and didn’t do better, significantly. That really got us into looking at the economics overall. 

Understanding Token Economics 

Doug Marcy: So what is the basis for the economics of models? You’ve probably heard the term tokens. Basically, the model converts words or parts of words into numbers, and those numbers become the token. Pricing is often set up as a million tokens ingested and a million tokens output cost this much. 

What’s interesting is tokens aren’t simply about input and output. In the modern world where models are reasoning or thinking models, they talk to themselves in intermediate stages — the more effort you tell the model to put in, the more tokens it generates talking to itself about your problem. Your cost can go up very quickly even if you’re only giving it a small amount of data. 

Doug Marcy: That gets compounded by things outside our control in RCM. Maybe a simple ambulatory surgery is just a few pages of op notes, or maybe it happened as a result of a car accident and the person was in the trauma ICU — a Cerner hospital sends us the entire inpatient record, but we only want the surgical piece. We still have to look at this 200-page record, and asking Claude or ChatGPT to look at a 200-page document costs 100 times more than a 2-page one. 

Doug Marcy: The other big piece is that right now there’s more demand for token processing than there is processing power in the world to manage it. The reason costs are where they are today is that these companies have been subsidized by a huge amount of investment — they’re all basically running at a loss. At some point that bill is going to come due, and that kind of uncertainty is very dangerous if you’re making AI integral to how you service your customers. 

Open-Weight Models and Cost Control 

Doug Marcy: So that leaves us wanting to say: OK, we’ve got AI, we’ve got a strategy for context, but how do we deploy it safely, reliably, and cost-effectively where we can model and reason about how much this is going to cost us? We ended up turning a lot to open-weight models. 

With Claude and Anthropic, it’s much easier now to get to that healthcare-safe place — they even have products like Claude for Health. But it’s still that pricing piece — you’re not in control of your infrastructure or your supply chain. So we turned to open-weight models. We’ve had a lot of luck with both the Qwen and DeepSeek models, and looked at the Gemma models from Google. 

Doug Marcy: The actual thing that tipped us down this path originally was security — being able to guarantee no data is leaving our environment, none of our PHI, none of our customers’ PHI, and importantly, making sure we don’t co-mingle data where one customer’s data and another’s are in some way sharing space. 

The trade-off was that it was more difficult to deploy — it can be hard to find hardware or rent a VM with the capacity you need. But these open-weight models are often much smaller than the frontier models. 

Model Size and the Sweet Spot 

Doug Marcy: When we talk about size, we think about parameters — basically every connection in the model has a number. A frontier model of Claude or ChatGPT is estimated by experts to be in the 3 to 4 trillion parameter range. Running an instance of Claude Opus requires a whole rack of equipment costing about $5 million just to run one instance. 

On the open-weight side, there’s a lot of smaller models, from 8 billion parameters meant to run on phones, up through 30-40 billion parameter models, and medium models in the 300-400 billion range. Since July, a couple of labs have released open frontier-class models — DeepSeek version 4 at about 1.4 trillion parameters, GLM 5.3 at 750 billion. 

Doug Marcy: We have found the sweet spot isn’t at that huge scale — it’s the small-to-medium range. When we take a smaller model and add in that curated coding guidance, that knowledge base we talked about, we’re able to get as good or better outcomes from a small model than from a large one, tuned to a very specific set of use cases. We can deploy multiple models like Legos, snapping them together to make a solution that’s better than the sum of its parts. 

These smaller ones can be run on something like a Mac Mini, or NVIDIA has a $5,000 device called the DGX Spark. The barrier to entry is much lower if you want to own your own hardware, and there are also services, including Microsoft, that will host these small models for you with set costs instead of usage-based pricing. 

Governance and the Maturity Model 

Doug Marcy: The big takeaways on economics: having a repeatable governance process is key. We’ve defined set steps — do we have the right data and access? Do we have a clear and defined process for making decisions? Do we have the ability to test effectively before release? And once we’re in production, do we have good ways to monitor, clear escalation, and a rollback policy? 

As we start with things, we often look at an initial deployment that’s more of an assist mode, where we’re not handing the reins over to the model. As we feel more strongly about the controls in place, that’s when we start letting the model drive a little more. 

Doug Marcy: The most important thing we’ve found is that data question and that expertise question. I don’t just mean the records — I’m talking about the intellectual property that your company has that allows you to do the job. That is gold. You’ve probably seen dozens, if not hundreds, of startups claiming AI has solved RCM, but in a lot of cases those folks really need your data and your expertise. 

So you have to be careful when partnering — what is that relationship doing for you versus for them? The risk is people coming to us saying they want to simplify our process, but really wanting to learn from us to make their tool better so they can sell it directly to our customers. You need to understand how valuable your intellectual property is — that’s the thing almost any startup in the space doesn’t have and will kill for. 

Doug Marcy: That maturity model is really defined around a gated process. We wanted to define early on what the goals and thresholds are to move something from a measurable place to an assistance place, to partially autonomous, to more fully autonomous. Sometimes we guess wrong — when we originally started auto coding, we had to develop a more nuanced definition of accuracy: accuracy versus CMS guidelines, accuracy versus our auditors, and clean claim submission as another quality marker. 

The worst possible situations for us have come from places where people threw AI at something without a plan ahead of time to tell if it’s actually working. Especially when it comes to the cost-benefit situation for a specific solution, you need to be able to measure the value — that’s the first question we start with these days. 

Looking Ahead: World Models and Generative Interfaces 

Doug Marcy: I like to end with at least one forward-looking slide. One is that the whole world of AI right now is dominated by large language models — but there’s an emerging area addressing a problem with language models, which is that they’re just language. They don’t have a capability that’s basic to most mammals, which is the ability to simulate the world around them and make predictions. 

A self-driving car right now has a hard time if a pedestrian walks behind something — our brains automatically predict they’ll come out the other side; that’s not a capability models have right now. Think about a model that can detect when something doesn’t make sense on a medical record because it’s physically impossible, or that it’s missing a piece because there’s a gap. Right now, language models don’t have that capability unless you explicitly train them into it, and even then they’re often wrong. World models have the potential to really understand the world at a level language models don’t. 

Doug Marcy: The other one: a lot of people are still making up the conversation as we go, using chat. There’s an emerging discipline where the model itself generates user interfaces for the human being — rather than just dumping a description to a person, it dynamically creates a view that’s most useful for that person to interact with and make a decision. This is going to make human-model interaction that much more powerful over the next couple of years. 

Closing Thoughts 

Doug Marcy: So, last notes: AI is useful. It’s no longer a question of can it do it — yes, AI is going to be useful. The question is, can we do it in a way that maintains our accountability and our quality to our clients? That’s the most important thing for us. We need to manage that risk, manage the compliance. 

The economics of it is complicated, and if you don’t understand the cost structures and how you’re going to get return on your investment, you need to dig into that. And lastly, the big win we’ve had this year is realizing those smaller models, those lower-cost models, can really do more for us than we thought at a much lower cost. That’s unlocked what we’ve ended up calling our revenue intelligence — taking that accountability, taking all our expertise, and building it into a platform where humans, agents, and our clients collaborate effectively using these tools. But they are just tools in the end. We still maintain our accountability to our clients. 

Audience Q&A 

Doug Marcy: Right, Q&A, and I see some have been coming in. Rosie, did you want to MC the Q&A portion here? 

Rosie Talaga: Yeah, that sounds perfect. Thank you so much, Doug, for a truly great, engaging discussion. We’re going to start the Q&A discussion. We’ve seen a lot of great questions come in, so audience members, feel free to continue submitting your questions via the box you see on your screen. 

Q: Have you worked on AI coding for specific specialties, such as colorectal surgery? 

Doug Marcy: Yes, we have actually worked on a number of different specialties: anesthesia, pathology, orthopedic, and colorectal is one we’ve looked at. The interesting piece is figuring out how much information or knowledge we need to give the solution to make it useful. 

Yes, we’ve worked with colorectal, ophthalmology, orthopedic, cardiology. The next big one we want to take on is emergency department, and we’re starting that work now. 

Q: As providers begin using AI to improve coding, are we seeing or can we expect to see payers use AI to limit or reduce payments to providers? 

Doug Marcy: Oh yes, and this is where I can get myself in a little bit of trouble. But yes, definitely, especially the larger payers have already probably deployed some of this and are working hard to use AI in lieu of some of the reviewer expertise they’re required to do right now. 

Make no mistake, payers are already deploying AI, and they’re going to continue to do so. As an RCM vendor, we see it as incumbent upon ourselves to bring the best technology to bear for our clients, to counter or interface with that technology, however you want to put it. 

Q: Medical coding is complex. How could AI interact with different insurance vendors who have different and even dynamic contracts with their clients for revenue management, often for the same procedures? 

Doug Marcy: That’s the big win we got from that knowledge graph I was talking about earlier. We didn’t just pull in the code books and crosswalks — we also brought in all the documentation we’ve built out and a way to understand and distill the contracts down. 

Even when it comes to processing denials, we’ve crafted prompts and context so that when we’re looking at a denial, we’re also looking at the contract, because sometimes payers put stuff in the denial letter that’s not actually supported by their contract. When we pull in a code, we also identify the client and the payer and walk that graph to pull in those specific policies, SOPs, and standard operating procedures for that payer. 

Q: E&M coding is not just time-based, it’s also based on complexity of care. How does AI integrate complexity into decision-making? 

Doug Marcy: That’s been a bit of a catch-22 in the space — you’d love to route cases based on complexity, but you don’t get to know that until you look at the record. So what we’ve done is break down the problem: we first extract what we need to know using a model trained to summarize a case, and part of that summary gives us the key indicators and data points we need to evaluate complexity. 

That becomes part of the input to the model that makes the coding decision. I want to be clear: the model doesn’t just say here’s a code — it says here’s a code and here’s why I think this is right, and if it’s not certain, it says I think it’s one of these three codes but I’m not sure, which is a trigger for a human to do a quick review. We did something similar with anesthesia records, where we ended up doing two summaries — one pre-op and one intra-op — because getting the right pre-op diagnosis and comorbidities to support the physical status score was much more about the pre-op eval than the actual work done in the OR. 

Q: Do you ever have to overcome handwritten records? Certain rural hospitals still use handwritten records, and we struggle with our systems analyzing those images. 

Doug Marcy: Short answer, yes. We’ve evaluated a number of different models for reading handwritten records. We’re getting to a point where we have about 98% accuracy on typed, scanned records, but for EMR-generated handwritten ones, we’re still getting only around 60 to 70%, depending on the type of record, especially for small cards that have been sent to us. 

The biggest determiner right now is actually the quality of the scan or how we’re receiving it — we still get quite a few handwritten records by fax, and that black-and-white, low-DPI quality is where we have problems. Where we’ve had pretty good success is with mixed records, common at rural or critical access hospitals, where part of the record is handwritten and another part comes from an EMR. We use a couple of different visual models — DeepSeek OCR version 2 for understanding document structure, and Qwen 3-VL, a visual model that does more reasoning — and compare results between them. My heart’s hope would be to get people off handwritten records entirely; it’s something we’ve done in our anesthesia business, where we bundle our own EMR with our billing services for locations that don’t already have one. 

Closing 

Rosie Talaga: Thank you so much. That is all the time we have for today. Thank you so much, Doug, for such an excellent, insightful presentation, as well as Coronis Health for sponsoring today’s webinar. And thank you to our audience for being so engaged, submitting great questions, and for joining us. We hope everyone has a wonderful rest of your day. 

Doug Marcy: And I’ll make sure that some contact info is in the recording as it goes out — I’ll work with the team there. But I’m on LinkedIn, just at Doug Marcy on LinkedIn, so feel free to reach out if you have anything more you’d like to talk about. Thanks so much.