Your payer can underpay you without a single denial | Ep 327

Context

I spent two and a half days at MGMA 2026 in San Antonio, the annual conference of the Medical Group Management Association, as an AI builder with a medical background who went there to learn. Practice leaders are under pressure to do more with AI, and most of what’s on offer arrives one tool at a time. This episode lays out the blind-spot triangle I came home with: whether a practice is paid what it’s owed, whether its AI’s work is right, and whether anyone owns the whole picture.

Key Takeaways

• One tool at a time rarely adds up to a plan. A phone vendor here and an ambient scribe there, and nobody checks whether they overlap, talk to each other or reach the right people.
• Payers already run AI against your claims. A claim billed as a 99214 can come back paid as a 99213 with no denial attached, so it never raises a flag. Catching it means comparing what you were paid against what your contract says.
• AI output needs someone checking it. A practice in one session said its ambient scribe kept ignoring the documentation rules they set. In my own underwriting system, rules enforced in code held on every run where written instructions to the model slipped.
• Coding is a verification task, and much of it is still done by hand. Some groups run large teams of human coders working through spreadsheets. The practices handling this well are letting natural attrition shrink those teams rather than cutting jobs.
• Name the number before you renew. Pick one AI tool you pay for and write down the single number that would prove it’s working. If you can’t name it, that’s your finding.

Prefer audio? Listen to Serious AI with Chris Duff on Spotify, or wherever you get your podcasts.

(Podcast transcript below)

Buying AI one tool at a time is how health care practices overspend. So welcome to Serious AI with Chris Duff discussing enterprise AI implementation. So I just returned earlier this week from the annual MGMA Summit in San Antonio. I went over in last week’s podcast about how this was kind of spur of the moment, making healthcare my primary vertical to go after and leveraging my MD and also healthcare venture capital background, plus a lot of strong macro-related forces behind the healthcare industry overall, even in the face of AI. So you can check out that conversation previously here.

But this was an absolute whirlwind conference. So this MGMA organization, it focuses or is targeted primarily for or medical groups around the U.S., so a lot of independent physician practices, so really diverse group of folks. I hadn’t been to a conference this large in many years. I think it was like 1,700 attendees in their most well-attended summit since before COVID and this was their hundredth annual summit here. So this organization has been around for quite some time. So my whole approach for coming into it, again, abiding by their non-solicitation strict requirements here, which not everybody was following, and I was on the receiving end of at FYI, not the best feeling there.

So I see why those policies are often in place. And really they’re just to learn again. And like that first day I showed up on mid-Sunday attending a revenue cycle management session, just all this jargon and these acronyms thrown at me. And it’s like, oh man, this is a whirlwind to get back into this groove. And even when I was in healthcare to a large degree several years ago, this wasn’t like the day-to-day that I was going after and how insurance companies interact with medical practices and things like clearing houses and so forth too. So that took a lot of up to speed and I’ll talk more about that next week in terms of how I oriented my meeting strategy here.

But in regard to key takeaways from this conference and what I learned here, which was just tremendous, absolute fire hose, is all these practice leaders and operators within these rooms here, like AI is just, again, it’s the topic du jour. Like pretty much every session that I attended had AI mentioned in some form or fashion. And so there’s a whole bunch of noise and also vendors and sponsors at this conference, like spouting AI tools and some of them very well funded and so forth. And a lot of these operators that are all already super busy, it’s like they already feel the pressure, should we be doing more with AI?

If they have AI, are they using it properly? And potentially some doubts yeah precisely unlike we don’t even know how to fully utilize what we’ve bought anyway and we’re so busy that you know some things are probably slipping through the cracks and we’re not fully optimized here so you know that was my role coming into this conference is just figuring out okay where’s the possible value at here, let me just familiarize myself, come in as a enterprise AI builder with a medical background who’s here to learn and see where there might be some holes to fill here.

And why a lot of this matters too is that a lot of these practices might be buying AI one tool at a time, particular phone vendor here, a what they call ambient AI, ambient scribes, as another tool where if a clinic provider’s having a patient encounter, they have some recording device that’s basically taking in everything that’s said. You can almost think of when you attend Zoom, a Zoom meeting or you have Notion or Granola or whatever running in the background, logging everything so that’s what we would call ambient scribe technology which is just rapid implementation as you can imagine you know just like any recording software and I use it all the time as well too but you’re getting different vendors here and you might have a bit more piecemeal strategy in place for putting this all together you know does some of it overlap Does it all talk to each other?

Does it get to the right people? That’s an open question. And critically, within this too, pretty eye-opening, but not necessarily surprising, as I was in a room that was specifically focused on, hey, how do you implement AI within these practices? And it was incredibly basic not to disparage the presenters at all. It was like, OK, they’re meeting their audience for where they’re at here versus, you know, me just going, you know, deep Claude Code setups, terminal CLI is running all the time here. Like it’s a very different type of approach. And, you know, out of that room of a hundred plus practice medical practice leaders in there, only 6% said they were running a mature AI program.

And my general read was that was probably the case here. Most folks are like in more of a pilot stage or kind of earlier implementation for a lot of AI. And again, we have to consider like these AI systems, these agentic softwares have only really gotten good during, you know, 2026. And I would say like after the first quarter of 2026. So it’s all like very recent, we’re super powerful processes can actually handle a lot of what’s needed on an enterprise AI side. And critical now because these massive players in the healthcare industry, these insurance companies that can just throw unlimited money into these AI developments, like they are already automating denials, which means that they’re denying claims for health care services or procedures that were done to you know, various patients here, the insurance are denying claims that they didn’t think were medically appropriate.

And so like, it’s basically withholding pay to the to the practices. And they’re just throwing like AI just like crazy to not deny it as much as possible. And there’s some macro reasons behind that too, because healthcare utilization has risen significantly since the start of COVID. So insurance companies, even though they’re extremely rich by like any measure here, they are getting squeezed, their margins are getting squeezed more because of increasing claims. So any way they can deny more, it’s gonna be their incentive. And, you know, also coinciding with, yeah, some of these healthcare startups in the AI space have raised hundreds of millions of dollars this year.

There’s a company, Assort Health, that’s, you know, raised over another hundred million dollars within the last couple months here. You know, they’re a unicorn worth over a billion dollars now. Oracle, one of the largest companies in the world, is directly targeting healthcare, trying to fold AI into the EHR itself. They had bought Cerner a few years ago. I’m familiar with that software back in my med school days. So all of these buying and implementation decisions are happening like now, this quarter here. Like it’s really a turning point, I could feel that.

A lot of these AI sessions were some of the most well-attended, standing room only, during the conference because of, you know, again, some of these, there’s so much extra fat that can be chopped off within healthcare site. Like they’re still one of the only industries that probably uses faxes routinely. Some folks are still doing like written paperwork. So all of that can just be removed which some of these practices have implemented here. So I attended 18 sessions over two and a half days, nine separate conversations. And yeah, I ran the whole conference through a Claude Code system I built.

I don’t think anybody else had something like this. And I’ll comment more on that next week here. But yeah, really to me, this kind of pointed out a what I’ll call a blind spot triangle for healthcare practices. And so in regard to healthcare practices, relationships with payers, which are insurance companies, our practice is actually being paid what they’re owed. That is very hard to track for healthcare practices, or maybe that’s not even the right word. It’s just it’s not necessarily prioritized as well as it should be and can be run by very archaic systems.

And again, you have payers who are oftentimes adversarial in their relationships with practices. So you really have to fight to get what is owed to you here, again, because the incentive kind of runs in the opposite direction on the other side. So that’s an open spot. There’s a whole bunch of conversations about that. Regarding the AI output, this is what I talk about all the time. Is your AI’s work correct? Who is verifying that things are actually working correctly? Because you can always take your vendor’s word for it, and some of these vendors, I think, were some of the smartest folks I’ve met in the AI and software space thus far here with similar approaches to how I would consider doing tooling for various businesses.

But nevertheless, if they’re spread thin across a bunch of clients or don’t necessarily have medical background, who’s doing that verification? A ton of these healthcare practices, they don’t have spare capacity for someone to really understand what is the AI’s output, how does it relate back to the underlying metric and the data that they’re trying to get, and even if something looks correct, again, the AI seems so smart week over week that plenty of times you can be misled. So that’s another opening too for a blind spot. And then also with all of these various vendors, I would say like most of them are doing like 80%, even 90% of the same thing, maybe with slightly different approaches is, you know, does anyone own the whole picture when it comes to healthcare practices?

And I know some of these vendors certainly have plans to tackle that. Nevertheless, I think every, I’ll always be of the opinion that if you can have an in-house build that knows your practice and operations, in and out is like always going to be better than a, you know, vendor tool, no matter how like customizable it can be here. That’s why I think like, you know, again, like the base LLMs and the underlying coding models. So like, you know, obviously Claude Code and ChatGPT Codex are effectively universal tools. I continue to learn more week over week on their own capabilities here.

And I think if you fully customize for specific businesses, operations, that’s going to be the best in class approach. Especially if you can layer on vendors for particular parts of the business that would just be unnecessary to tackle with your own solo build. So yeah, like some of these AI voice technologies and so forth. Like there’s no reason to rebuild some of that because that’s already very well done. But I would say like a lot of the guts of the practice in determining what are our core constraints to go after, even if some of these vendors can really tackle it effectively, in one way or another, it might not be the most cost-effective approach to handle longer term again, especially as these AI models continue to get better and better.

What are you really paying for? Do you know that you’re actually getting the best deal here? So I would say that’s another blind spot here, that unless you’re in this at the frontier on a day-by-day basis, you’re not going to have the capacity to think about that. And so, you know, regarding, you know, the how for this blind spot triangle, too, is like, okay, for payers, I mentioned, you know, they’re being incentivized to send out denials, but they’ll also do what’s called silent down coding. So you as a practice might bill a, you know, for example, let’s throw out this number, these are CPT codes that, You know, basically define certain procedures or services that medical practice might run that’s supposed to be tied to a certain payment coming back to them.

So for example, this number might be like a 99214. But the payer, the insurance company might silently on their end downgrade that number to a 99213, which actually pays less. So it doesn’t count as a denial. It’s not going to show up as like a red flag in your system. But you might just not be catching that, Hey, you’re not getting paid for what you were actually logging. And if you’re not double checking the contracts or making sure everything matches, you know, going to the payers and back a lot of folks aren’t going to catch that. And there were a whole bunch of examples where you know, for example, some what one practice thought he was getting paid 150% of Medicare’s value.

Again, there’s a lot of jargon in here, but you know, Medicare kind of sets the standard for what healthcare services should be, should be paid back to whoever’s providing the service. So if you have like 50% over what Medicare’s paying, it sounds pretty good. But then when you’re actually diving through with all the silent down-coding that happens and the insurance company might have its own fee schedule that doesn’t come back to Medicare, again, these complex contracts, that 150% that is on paper might only turn into 89%. So like that, that was a real example thrown out there.

And in one of these sessions when a room was asked who checks what they’re paid against what they’re actually contracted for. Fewer than 10% of the hands went up in the room. Again, there’s so much going on. Healthcare practices are some of the busiest companies in the world, in the country, certainly. And so it’s just, it’s hard to be maintaining something like this. It’s like a perfect job for AI. And also for that Scribe technology that’s recording and making sure, okay, things are getting logged appropriately, documented. You know, there was someone in a room that said, hey, yeah, we’re using some of this scribe technology, but 20 to 25% of the time, the rules that they’re setting are just getting ignored.

And, you know, this comes into play, like when you were building the SLC Deal Engine, is you can have prose-based rules, but those are not necessarily as effective as doing code-based approaches where it’s much firmer. It’s easier to break if you don’t follow the procedure exactly, but the rules are much more firm in getting the exact response every single time. And it doesn’t cost quite as much. You don’t need judgment models for running that. So that’s another piece that could come into play here. So like a documented review is really the only defense.

And you have these large groups of employees called coders that are making sure that whatever the medical practices determining the services that they’re meaning to provide that these coders have to go make sure that they are sent off correctly to the insurance companies. And they’re held to certain levels of accuracy here too. But a lot of these folks aren’t utilizing AI assisted technology and are just relying on their own Excel skills or risking human error. But I was just blown away where you have some of these groups that have like 200, some odd medical providers and they have 50 coders, human coders you know, still verifying by hand, like pretty much Excel files here which, which to me, like, just does not make any sense. [Correction: that group has 250+ providers and 50+ coders.]

That’s like overhead that can be cut. And that’s something too, like there’s the human element and a lot of these practices were like, yeah, how do we deal with a folks are scared for their jobs here, they think they’re just gonna be fired. So a lot of them have taken what I think is a pretty intelligent approach, where it’s like, hey, we know our labor force in this particular area faces a certain attrition over time, so instead of just saying, hey, yeah, we’re gonna cut a bunch of jobs, it’s like when people naturally attrition out of certain areas, then they just don’t rehire instead, which I think is both smart, good for morale and a more humane approach too.

But nevertheless, if some of these practices are really facing cash flow pressure here, it might force some tougher decision making, especially if you have some less or more cutthroat competitors who are just firing off the significant amounts of their workforce and you might just need to compete accordingly. So that’s just something to, to note here. And also again, like regarding who’s owning that whole picture of the AI sphere is like a lot of these vendors, again, like I mentioned, they’re smart, a lot of them are well funded, tens of millions, hundreds of millions of dollars several measured ROI.

Well, like using practices own numbers, own reported numbers and then seeing, okay, are we tracking a real ROI here in terms of, you saved or money gained in terms of like more patients booked, which tends to be a better number that they’re looking for. Folks would rather, hey, build my throughput up higher versus like cutting costs is going to be their preference. But again, a lot of these like own one slice or they might kind of overlap with each other. I saw that a whole bunch in my opinion. I attended a keynote with a higher up at Oracle and they were warning that buying point tools one at a time is certainly risky.

Obviously Oracle’s incentivized because they have their own enterprise healthcare AI software that’s meant to tackle the entire healthcare operation. But nevertheless, if you look at Oracle’s offerings too, they’re going point side one by one. Right now they’ve been mainly focused on clinical documentation. I know they’ve recently opened up some other tools here so you have to take that with a grain of salt as well too for who’s mentioning everything. So you know to wrap this one up here there’s certainly more obviously this is two and a half days worth of conferences that I’m trying to boil down but I want to keep this concise for your own memory and what you can apply especially if you’re running a healthcare practice or somebody who did, is if you were to pick one AI tool that you currently pay for or that you’re considering paying for, what would be that single number that would prove whether it’s working or not?

But if you can’t name that number, then that’s your finding to go back and figure out that constraint. And that’s exactly how we approach our implementation at Serious AI Partners, which you can check out at seriousai.partners and if you’re in the healthcare field, throw a slash healthcare on the end there and you can find exactly how we work within the healthcare sphere. So super, super interesting conference. I’ll be reporting more on this. Honestly, I could do like three or four podcasts at least on everything from here. Subscribe and share everybody looking forward to next time.

And if you would like to work with us again, you can fill out that contact form on our website, seriousai.partners looking forward to next time. Take care everybody, bye.