Updated March 5, 2026
0:00 Welcome to Colaberry AI podcast brought to you by Colaberry AI Research Labs and Carle Foundation. Today, we're doing a deep dive into a Wall Street Journal article, titled UnitedHealth now has 1,000 AI use cases, including in claims. You share this one with us and, yeah, our mission today is really to dissect how a major player like UnitedHealth Group is weaving artificial intelligence into its operations. We want to, drill down into the technical side and the results they're actually seeing. Absolutely. 0:30 And what really jumps out right off the bat is just the scale we're talking about here. Over a thousand AI applications. I mean, this isn't just dipping a toe in the water. It really signals a sort of fundamental change in how they're operating across the board. A thousand. 0:43 That's it. Yeah. That's staggering. It makes you immediately think about the, the sheer infrastructure needed. Right? 0:48 The data pipelines, everything behind that. Now the article highlights claims processing is a major focus. Okay. Let's get technical here. What kind of AI techniques, what sort of architectures would they likely be using for something as as notoriously complex as claims? 1:03 Well, for claims, you'd expect a pretty sophisticated mix for just getting the data in, probably optical character recognition, OCR paired with, natural language processing models, NLP is key there because claim forms are well, they're all over the place, structured, unstructured data. Then for checking eligibility, looking for potential fraud, you'd likely see rule based systems working alongside machine learning classifiers. These classifiers, they'd be trained on, you know, tons of historical claims data to spot patterns that look suspicious or ineligible. Anomaly detection algorithms would be vital to flagging claims that just look odd compared to the norm based on service codes, patient history, provider patterns, that kind of stuff. Okay. 1:46 So it's not just about making it faster. It's applying some serious computational muscle to find these, like, subtle signals and massive amounts of data. And the aim, I assume, is hitting both efficiency and accuracy gains. Exactly. Automating the routine stuff, that's where you cut down on manual work, speed up processing, lower admin costs. 2:03 That's the efficiency part. But the machine learning models also have the potential at least to be more accurate catching inconsistencies or errors a human might miss, especially working at scale. And the article mentioned AI isn't just for claims. It's in health services data analysis too. Can you give us a feel for the technical approaches there? 2:20 What might that look like? Like? Sure. In health services, think predictive analytics. Maybe using time series forecasting models to guess patient demand or, predicting health risks using patient histories over time. 2:33 You could see things like, say, recurrent neural networks, RNNs, or maybe even transformer models analyzing sequences of medical events to flag patients at high risk for, like, hospital readmission, for data analysis, maybe clustering algorithms to group patients with similar characteristics, you know, for more tailored care plans. And causal inference techniques could be really powerful trying to figure out the real impact of different treatments or interventions. Right. And applying AI this broadly across so many different areas, it must come from the top strategically. The chief digital and technology officer mentioned this responsibility, did they? 3:09 Ensuring AI actually leads to more affordable, more accessible health care. What are the technical underpinnings of achieving that kind of goal? Yeah. That's a crucial point. Technically, making AI affordable and accessible means focusing on building systems that are, well, scalable and efficient. 3:27 That involves optimizing the algorithms themselves, using cloud infrastructure smartly to handle the data and compute load, and making sure these new AI systems smartly to handle the data and compute load, and making sure these new AI systems can actually talk to the existing hospital IT systems interoperability is key. And accessibility isn't just about cost. It's also about designing interfaces that doctors and nurses who might not be AI experts can actually use effectively. Plus, you absolutely need rigorous monitoring for performance and critically for bias to make sure you're not accidentally making health care less fair. Okay. 3:55 But the article also raises this red flag about past problems. Right? Flawed AI algorithms and claims reviews leading to lawsuits even. What kind of technical issues could cause those sorts of failures? Oh, absolutely. 4:09 Those past issues likely came from a few potential culprits. Bad data is a classic one, insufficient training data, or data that already has biases baked in. That can lead the model to make skewed decisions, maybe denying claims unfairly for certain groups. Overfitting is another possibility. The algorithm learns the training data too well like memorizing it but then fails when it sees new real world claims. 4:32 Or maybe the features weren't right. The input variables they fed the algorithm just didn't capture the important factors for making a good decision. And, of course, maybe just not enough testing, not enough validation on diverse data before rolling it out. That's a big risk. And the article mentions they're shifting now to newer algorithms that evaluate claims and then send automated letters. 4:51 What's the tech behind that likely look like? How's it different? That shift sounds like maybe a move towards a more hybrid system, perhaps combining some stricter rules based logic with the machine learning models. The evaluation part might still be AI models flagging claims based on criteria, but maybe not making the final denial decision outright. Then the automated letter generation kicks in based on that evaluation. 5:16 Technically, you could use natural language generation NLG models to write those letters explaining the outcome, what needs to happen next. It seems different because it potentially keeps a human more involved even if the initial flag and the communication are automated. It's not a fully black box denial, maybe. Okay. But then there's that statistic about one of these newer algorithms, the small test, where something like 90% of the claims it denied. 5:40 Well, they still weren't paid even after a human reviewed them. That sounds like a pretty high failure rate still. What did that tell us technically? Yeah. A 90% nonpayment rate even after review is significant. 5:51 Technically, it suggests a deep mismatch. Either the AI is flagging claims based on features that genuinely do lead to denials upon human review almost all the time, or there's a problem in the review process itself. It could point back to persistent data issues, or maybe the AI's internal logic is flawed in how it weighs issues or maybe the AI's internal logic is flawed in how it weighs factors. From a validation perspective, it screams that you need better metrics than just simple accuracy. You need to look at why things are being denied, the precision and recall for specific denial types, and really dig into the errors what kinds of claims are being flagged. 6:23 Why are humans agreeing with the AI denial so often in this test case? It also puts the spotlight on that human review process. Does it have the power and information to actually overturn the AI effectively? It's complex stuff. And finally, just thinking about the scale, we got UnitedHealth having 20,000 people working on AI and related tech. 6:40 That's a massive investment. What does that signal for AI and health care more broadly, do you think? It signals a huge bet on AI being transformative, plain, and simple. When a giant like UnitedHealth invests that heavily, it tends to pull the rest of the industry along. We'll likely see faster development and adoption of AI tools across health care because of initiatives like this, their successes, and, yeah, their failures too, will be learning experiences for everyone else trying to implement AI. 7:06 Expect more advances in personalized medicine, maybe AI helping with diagnostics, speeding up drug discovery, making hospital operations smoother. AI is going to be increasingly central it seems, and that large workforce. It tells you there's a massive demand growing for people with skills in machine learning, data science, AI ethics, specifically tailored for health care. Well, that wraps up our deep dive into UnitedHealth Group's pretty extensive use of artificial intelligence. We've covered this year's scale, especially in claims, look at the potential advancements, but also some of the really serious technical hurdles and challenges, like getting those algorithms right. 7:41 It really highlights just how complex and fast moving this whole field of AI in health care is. Thank you for listening in. Subscribe and follow Colaberry on social media links in the description, and check out our website, www.colaberry.a I backslash podcast for more insights like this.
