HIMSSCast: Payers and providers gird for an escalating AI prior auth arms race (brought to you by SuperDial)

Speaker A: The good news for patients and doctors is that most of the time for these kind of simple cases where medical necessity isn't really at ah issue, the answer is going to be yes. And so AI is a valuable tool for speeding those decisions along.

Speaker B: Hello everyone and welcome to himscast. My name is Mike Milliard and I'm Executive Editor of Healthcare IT News, a HIMSS Media publication. With me this week is Michelle Mello. She's an empirical health law scholar who's a professor of law at Stanford Law School and a professor of health policy at Stanford School of Medicine. She's also the lead author of an article published in Health affairs earlier this year titled the AI Arms Race and Health Insurance Utilization Review. Promises of Efficiency and Risks of Supercharged Flaws. We're going to talk about that AI arms race today and discuss where that competition could lead and what it means for payers, providers, and ultimately patients and health plan members. Welcome Michelle, thanks for being here.

Speaker A: Thanks for having me. Kim's Cast is an editorial tutorial podcast from HIMSS Media. This episode is brought to you by Super Dial. Super Dial is building the AI native network that enables seamless AI to AI communication between payers and providers. Learn more at superdial.com.

Speaker B: We've seen in recent years, as you note, kind of doubling and redoubling of use of AI on both the payer and the provider side. The providers want to get paid, naturally, the payers don't want to pay if they can avoid it and they're using these tools to their advantage. So let's start with some basic questions. What were the goals behind this study? And maybe describe the state of affairs in the payer provider arms race as you see it today. What does your research show generally about the reimbursement space of 2026?

Speaker A: My mission as a researcher is to try to surface evidence to help organizations and public policymakers make good decisions about AI. And uh, recognizing that this was a really high stakes area for use of AI and things were happening very fast, we wanted to try to get a sense of the state of play and analyze in a balanced way what both the uh, promise and the potential drawbacks of the move towards AI might be.

Speaker B: What were some of the things that kind of maybe leapt out to you immediately as you started poking around and researching this specific topic?

Speaker A: It was important for us to ground our analysis in the counterfactual what would the world look like without the use of AI? It's really easy in conversations about AI to get critical really fast and forget how flawed human only Processes have been, and there is, uh, it's hard to find a place where that is more true than prior authorization or prior auth. That process we've all endured of having a drug or a procedure having to be pre approved by our insurer before the doctor can provide it. So we started from the premise that as uh is commonly said at conferences about this topic, prior auth sucks. And the question then is how might AI make it better? So that was the starting point and then we just explored what was going on and what some of the ups and downs of this transition might look like historically.

Speaker B: When did this escalation, this kind of tit for tat of use of AI on both the provider and the parasite begin in earnest? Obviously there must have been some automation and machine learning processes used on the margins. But when did this kind of arms race take off full speed, do you think?

Speaker A: Yeah, it's a great question because this war has been longstanding. I've been working with other colleagues at Stanford here, Elise Adams, and a, uh, med student, Stephen Gong, on understanding the history of prior authorization. And to understand that history is to understand that it's repeated cycles of locking horns, policymakers in the public getting very upset at insurers and physicians chiming in as well about the burden on them, and then a backing off of hostilities a little bit and a reformulation and then the cycle repeats it again. So that inference of AI is just the latest escalation of that war. But it's important to recognize again, the counterfactual is not that there were no algorithms involved. Insurers have been using simpler rule based algorithms for years to try to make decisions about insurance coverage. And they have to because of the volume of claims that they're dealing with. So one way to think about this is that these are potentially better algorithms than the ones they've already been using. But of course AI always introduces new risks.

Speaker B: Yeah, and before we talk about those risks, maybe let's talk about the other side. Obviously vendors are happy to sell AI tools to each player in this kind of skirmish. Payers are using it, but providers are also using it to optimize their claims and to make sure all the I's are dotted and the T's are crossed before they submit. How specifically are each side using AI most often?

Speaker A: Sure. We should start by understanding that there are three components of insurer's utilization management processes that algorithms are used in. The one that I'll focus on is prior authorization. That's that pre approval step before you get a service or A drug. But there is also concurrent review. That's where you're getting the thing, but you need more of the thing according to your doctor. And the insurer is going to decide whether you get to keep getting it. And then there's retrospective review. You got the thing, and then your insurance company is receiving a request for reimbursement after the fact in prior auth. There are a number of challenges that insurers confront in trying to adjudicate these requests with the limited information that they have in hand. So from their perspective, AI might be really helpful in first of all, finding requests for prior authorization that are incomplete or missing information, because that's a major reason why they have to initially say no and send it back to the doctor. And then secondly, a lot of these requests are relatively straightforward in that they, um, involve simple matching tasks, like did Mike already have his 20 physical therapy visits for the year? Or was Michelle actually on our insurance plan on May 27th of this year? So for that kind of stuff, it's only natural that they would think about algorithmic determinations. And the good news for patients and doctors is that most of the time for these kind of simple cases where medical necessity isn't really at, uh, issue, the answer is going to be yes. And so AI is a, uh, valuable tool for speeding those decisions along. Now, on the other side, providers and patients are often confronted with denials in cases where medical necessity is at issue. So from their perspective, not only do they want, want to make sure that their requests are complete and persuasive, as they possibly could be the first time they get sent in, but also that they contain all of the relevant clinical information that the doctor has noted in the ehr. And it turns out that extraction of that information can actually be pretty tough to do. And then it's often done by relatively low level administrators in healthcare organizations, not by the doctor, but by herself. And so the prospect of AI here from the provider side is that it can do actually a better job of finding relevant clinical information to support medical necessity determinations in a voluminous electronic health record, and then it can compose a nice persuasive request.

Speaker B: Clearly, we're at a moment here where the horse has left the barn. To borrow a cliche, this is not going to be done without AI ever again. Probably. But the question is, where is the human in the loop? Obviously the headlines are the risk of AI itself and AI alone making decisions like this. How close are we to that becoming something that is commonplace?

Speaker A: Yeah, that's the rub. That's what all the fighting is about and litigation is about. By law, insurance companies are required to have medical professional review of medical necessity decisions before they issue denials. And by all accounts they do that, there is a human in the loop. The question is, what exactly is that human doing? How long are they spend doing it, and what incentives do they have to reverse an algorithmic recommendation that the answer should be no. I want to just press the pause button again. And note that almost always in prior auth, the answer is actually yes. Being the kind of the lowest scoring plans, we might have no's 20% of the time, but more typically we're talking 85, 90, 93% of the time it's a yes. And so arguably we don't need a human in the loop because these are again, relatively straightforward yeses. And so why not just auto approve it? Just pluck the human out together. They're not serving a useful purpose. Even if they might occasionally find a false yes, it's not worth the insurance company's time to do that. And let's save the humans for the denials. And then the difficult part is making sure that when they're processing those putative denials, those human reviewers are conducting that review fairly and meaningfully and transparently.

Speaker B: Was there anything in your research that maybe took you by surprise a little bit?

Speaker A: I really don't think anything comes to mind that I find surprising. No.

Speaker B: I guess the logical next question is where do you see this going? Typically in an arms race, when we use terminology like that, it implies that there's going to be a continuing stacking up on either side. Is that where you see this going? Or do you think there's eventually going to be a balance of sorts where each side of the claims processing equation uses AI more or less in equilibrium?

Speaker A: I think there will be an acceleration of this trend in the sense of more companies, both on the insurance side and on the provider side, using AI to help make coverage decisions. And an important reason for that is that providers hate doing this. They hate it. They cite it as a major contributor to burnout. It takes them hours every week or they have to pay a staff person to do it. It's terrible. So they should be, with good reason, interested in tools that can help them do that, if those tools do it well. And on the insurer side, they're under actually a lot of regulatory pressure to make these prior auth decisions quickly, sometimes in just a matter of a couple of days. And given the volume that they're dealing with and the fact that A lot of this is just relatively straightforward. Again, it's enormously appealing to use these tools. So I think we'll see an acceleration in that sense. We're also though, seeing a little bit of movement in the opposite direction in another sense, which is that, as I mentioned, there's been this historical contestation of how often and in what ways insurers are using prior auth in general, apart from whether they're using AI or not. How often are they using prior auth? People do not like this process. It might be the best bad idea we've had for keeping healthcare costs down, but everybody hates it. And one thing that I worried about in doing this research is that AI might be used as a reason to expand the scope of prior auth. In other words, if you're an insurance company now, to do a prior auth determination doesn't cost you the 40 or 50 bucks you used to have to pay a human. You can do it practically for free. So why not do it for more services? Policymakers and the public have gotten very upset about the scope of prior auth, about the entrance of AI. So we're actually beginning to see large companies make commitments to reduce the scope, meaning that this process would be applied to fewer kinds of services. And that seems like a very positive de escalation of this conflict.

Speaker B: As we've been talking, I'm realizing we've confirming this mostly about private sector insurers and providers. What about the biggest payer of all, Medicare and this wiser model? Uh, do you have any thoughts on that? About that? That's been getting some pushback Capitol Hill from Democrats. And uh, they're saying it's delaying and denying care unnecessarily to seniors. Where do you see that perhaps going and how does that impact this conversation? With its own gravitational pull, perhaps?

Speaker A: Yeah, great question. For context, it's important to understand that Medicare coverage is really taking place in two different worlds. One is the world of traditional Medicare plans, which are fee for service models. But the other and increasingly more dominant world is what's called Medicare Advantage. It's managed care plans. Most Medicare beneficiaries these days choose to go into managed care plans because they are less expensive out of pocket. And in that world, prior auth has been very dominant. It's been widely used, and it's just used more and more over time. In the other world, the traditional world of Medicare coverage, it has not been very common. Less than 12% of services in that world historically have been subject to prior auth. Now for the first time we are seeing proposals to start expanding the scope of prior authorization in those traditional plans. They're not coming from the government in the sense of the payer who's moving this. They're coming from the Trump administration wanting to root out what it characterizes as fraud and waste. So it has created this pilot program called Wiser to go after at this point, a very small set of services that it believes are over prescribed by physicians, arguably for financial gain, that aren't evidence based. And importantly, it is using AI to do that. And not just using AI, it is using third party contractors that are tech companies to run this program and most controversially of all, to share back more money with those companies, the more denials they make that are not subsequently overturned.

Speaker B: As we wind down here, reframe this conversation on the patient and the health plan member, the person receiving care. You talk in your research about trust, both AI itself and insurers, it could be fair to say, have a bit of a PR problem these days when it comes to the public. Uh, what do you see when it comes to what this all means for the patient as we look to the years ahead?

Speaker A: I think you're absolutely right that there is a crisis of trust here because this experiment with AI enabled Prior Auth brings together two things that patients hate, which are Prior Auth and AI. They're very suspicious of both those processes. So the intersection, of course it's going to be doubly suspicious to Pat. I do think patients should also see this as an opportunity to get decisions more quickly and if the tools are good, more accurately, particularly if the tools are good on their doctor side. So their doctors are submitting higher quality requests that are getting turfed out more. But it is entirely reasonable for the public and for public policymakers to want more transparency about how these tools are being used. Because the fact is it is almost impossible to tell how they're being used from the outside. Insurance companies will simply repeat the adage that the there's always a human in the loop before they make a denial. They don't really own up to the fact that human might be looking at a file that's already been curated by an AI tool, give them a nice little pithy summary that is teed up to support a no decision, and they don't share data about whether AI is making a process that many people found objectionable before, better or worse. And truly either of those things is very possible. So I think what's going to have to happen in order to improve trust is more sharing of information with regulators at the state and federal level and with researchers that can help the public understand whether this development is good for them or bad.

Speaker B: Yeah. In your research paper you offer some recommendations in institutional governance, monitoring, models for underperformance, staff training and basically transparency. What do you have to say as far as how we can make this stuff better?

Speaker A: So I do think there are some structural interventions that need to happen here. For example, my belief is that human reviewers at insurance companies should look at a file fresh without any pre curation by an AI tool so that they're not subject to the decisional biases that we all encounter when information is neatly summarized and packaged and framed for us in a particular way. There are also, as you said, needs to think about who is using this technology and what do they understand about it, even just on the provider. The folks at our hospital here at Stanford who do this work work very hard at it, but they're not clinically trained. And so for them to be good overseers of a tool whose job it is to find clinical stuff in the medical record, that's tough. They have to know whether it's right or wrong. And importantly, if they can't be good correctors of errors, the algorithm thinks it's right when it's not. And so through reinforcement learning, it might just get worse over time. So these kind of structural things need to be addressed. But beyond that, as I said, information sharing is really critical if the goal here is not just to get it right, but also to show people that you're getting it right. And because the health insurance industry is in such a crisis point from um, an optics or politics perspective, I think insurance companies should be interested in doing that.

Speaker B: Any closing thoughts or I suppose maybe I could just ask, are you optimistic that we will find a balance in this arms ra Obviously AI is now a fact of life. It's woven into processes of every shape and size in all of our lives. But this is a particularly sensitive issue, as we've discussed. Are, ah, you ultimately optimistic that this could find its proper balance and ultimately work better for patients and those receiving care?

Speaker A: I do have some optimism about that. Again, because we should all remember how bad it has been historically. This is not like improving physicians notes where they've always done a really good job of taking notes for the most part, but it's just really hard on them. No, this is an area where they have not done a good job. Insurance companies have not done a good job of delivering outcomes that the public feels are acceptable. So we should be welcoming of earnest attempts to improve, but I think it deserves more scrutiny than it's getting and more oversight, again through information sharing.

Speaker B: All right, Michelle, this has been a very informative discussion. I appreciate you taking the time to be here today. So thank you. And thanks, of course, the Hymns cast audience, for joining us today. If you like what you hear, please subscribe to himscast on, um, Apple podcasts, Spotify or Amazon Music. Thanks again, Michelle. It was great to talk to you.

Speaker A: Thanks very much.