AI agents shone on the HLTH stage; I got a different story behind the scenes

Alexis Kayser

By Alexis Kayser

Healthcare Editor

See more of our trusted coverage when you search.Prefer Newsweek on Googleto see more of our trusted coverage when you search.

This is a preview of the October 23 edition of Access Health—Tap here to get this newsletter delivered straight to your inbox.

Good morning. I’m writing today’s newsletter on the plane ride home to Chicago after spending the first half of the week in Las Vegas. I don’t know about you, but I have a HLTH hangover.

More than 12,000 of us gathered at The Venetian to discuss everything from health policy to health tech. Most of my conversations—and a number of the on-stage sessions—focused on innovation, with AI, unsurprisingly, at the epicenter.

In Other News

Major health care headlines from the week

Newsweek and Statista released our ranking of America’s Best In-State Hospitals for 2026. Our analysis recognizes 800 of the nation’s top hospitals across all 50 states and the District of Columbia.
- Click here to see how your hospital stacks up against peers in your state (and to read more about the methodology).
The U.S. is bracing for higher health care costs as the government shutdown drags through a third week. In this article from the Harvard Kennedy School, Mark Shepard, an associate professor of public policy, breaks down the health insurance “subsidy cliff” that 20 million Americans could fall off come January, if Congress fails to reach an agreement.
- Yesterday on NBC’s Meet the Press, CMS Administrator Dr. Mehmet Oz said that President Donald Trump has a plan to replace Obamacare. But he didn’t offer more when pressed on specifics. “I wouldn't extend the subsidies the way they were created for COVID,” Oz said.
- KFF released its 27th annual survey of employers that sponsor health insurance, finding that average annual premiums rose five percent for single coverage and six percent for family coverage, year over year. Newsweek's Workplaces Editor Aman Kidwai has the scoop.
- Meanwhile, a survey from the Associated Press found that 6 in 10 Americans are “extremely” or “very” concerned about health costs rising in the next year.
- And in a fireside chat with a former U.S. Secretary of State last Friday, Florida Governor Ron DeSantis suggested that people under the age of 50 only need a “catastrophic” health care plan. While this may help curb costs in the short term, it counters many health systems’ plans to play the long game by investing in primary and preventative care.
UnitedHealth Group has been served papers in a fresh set of health system lawsuits.
- In Tennessee, Ballad Health claims that UnitedHealthcare “systematically abused and manipulated the taxpayer-funded Medicare Advantage Program” and seeks $65 million in damages allegedly accrued over a five-year period.
- And in Wyoming, Star Valley Health accused Change Healthcare, owned by UnitedHealth Group and operated by its Optum division, of breaching contract by failing to submit claims to payers in a timely manner. The hospital district is seeking $2.5 million in damages from its former payment system.
- Change Healthcare denied Star Valley Health’s accusations in a statement to Newsweek’s Lauren Giella, but UnitedHealthcare hasn’t returned our request for comment on the Ballad Health suit.
Tens of thousands of Kaiser Permanente workers ended their five-day strike on Sunday morning. The California-based system resumed negotiations with bargaining units yesterday, promising to focus on unionized workers’ stated concerns: wages, patient care and staffing.
- But California can expect another strike soon—this time from nurses at six hospitals owned by Tenet Health. Yesterday, The California Nurses Association (CNA)/National Nurses United (NNU) announced a one-day strike later in the month. They’ll protest the hospitals’ alleged “refusal to address nurses’ deep concerns about patient care and safe staffing” throughout contract negotiations that have spanned the better part of this year. Read more at Newsweek.

Pulse Check

Executive Perspectives on key industry issues

Fawad Butt is the co-founder and CEO of Penguin Ai. (Penguin Ai)

...

At the end of September, I sat down with Fawad Butt, co-founder and CEO of Penguin Ai.

Disclaimer: I’m pretty choosy about the tech vendors that I’ll chat with, and even choosier about those that I give airtime to in this newsletter. The reason I wanted to take a “Pulse Check” with Butt is twofold. First of all, his current endeavor zeroes in on agentic AI, which leaders at HLTH widely described as the next frontier for their AI investments. And secondly, he’s not an outsider to the health care industry. He’s lived this stuff on some of the biggest stages, with some of the highest stakes.

It’s possible that you haven’t heard of Butt, but chances are, his decisions have affected you—or at least your PHI. Butt previously served as chief data officer at Kaiser Permanente, UnitedHealthcare and Optum, so throughout his career, he has overseen nearly 70 percent of U.S. health care claims data.

A little over a year ago, he launched Penguin Ai to tackle some of the administrative inefficiencies he encountered in corporate health care C-suites, which he alleges Google, Amazon and Microsoft “fell short of solving.”

When I spoke with Butt, he described Penguin Ai as “the agentic platform for health care,” enabling both payers and providers to “build, deploy and manage AI agents” that “are supposedly going to replace your current workflows.” He declined to share the current number of customers but spoke of “rapid adoption” and a quicker-than-anticipated buying cycle. As of our conversation on September 29, Penguin Ai was preparing to launch with four new customers over the ensuing 30 days.

Read on for a portion of our interview, including Butt's take on the risks of AI agents, the failures of large language models (LLMs) and the AI misconceptions he’d like to clear up.

Editor’s Note: Responses have been lightly edited for length and clarity.

Fawad, give me the rundown on Penguin Ai. What is the platform’s functionality, exactly?

Look, if you think about any enterprise, especially a health care enterprise, there's a spaghetti mess of information and data that lives across many, many systems. The first thing any platform, agentic or otherwise, has to do is to be able to connect to that data, to be able to understand that data, to be able to interpret that data, to be able to preserve the privacy of that data, to be able to govern that data. So that's really the first thing we do, is we provide what we call the foundation layer, where you have out-of-the-box connectivity to all the ecosystems where your data lives.

Today, we have a private engine that preserves the privacy of that data, so PII and pHI, the personally identifiable information and health information, are sort of redacted unless there's a specific use for it. And then we provide a governance engine on top of that, because a lot of enterprises are using this data to fine-tune or build their models, and my experience, having been in health care data for a long time, has been that there tends to be a lot of bias in this data, so you have to first govern it and identify the bias and actually remove it before you can start to use it. So those are some things we provide foundationally in the first layer.

The second is the model layer. This is where we provide custom-built models for each one of the use cases that we're going after, both on the payer side and the provider side. Examples of this would be a small language model for prior authorization, a small language model for risk adjustment, another one for claims adjudication, another one for payment integrity. So think about eight or 10 key processes that underpin health care—we're building small language models for each one of those and making those available on the platform so the customers can adopt them.

The third layer for us is the agent layer. This is where we build agents and provide those out of the box so that the providers can start to use them immediately. And these agents use the small language model that we have on the platform. They use the data connectivity and privacy that is in the foundation layer, and then they do the work that humans currently do in many formats.

Got it. Why build small language models as opposed to large language models, or LLMs, which have been at the forefront of the AI market so far?

I would actually invert the question and say, why not build a small language model? Why use the large language models?

Because think about it this way, right? If you look at the world today, we have generalists and we have specialists in every industry and in every area. That's kind of how I think about the models: the large language models are the generalists. They are designed and trained on everything these research labs can get their hands on, which is primarily Wikipedia, Reddit, all of the internet and very low proprietary content. They're an inch deep and a mile wide. They're designed to solve all problems, for all industries, for all people at all times, and when you build something like that, it's hard to make it a specialist in any area.

So when you apply these large language models to specific industries such as health care, they pretend to know what's happening or understand the language, but when push comes to shove, their accuracy starts to decline very, very quickly. They start to fall down. So if you're doing prior authorization and you want to use a general purpose OpenAI model, you can do it. But what you're going to start to see is that after 65 percent accuracy, you're going to flatten out. And if you're an insurance company, 65 percent accuracy is not enough, right? You need to be in the 90+ percent accuracy range.

And the only way you get there, Alexis, is to be able to build a small language model that is a specialist in prior authorization. You have given it reams of data that are very specific to that task. It has seen all dimensions of [that task-specific data], and only then does it become a specialist that can be deployed and can be used in enterprises.

That's why we've taken this approach. When [Penguin Ai] went on the market a year ago, we started talking about it. I don't think small language models were even a concept that people knew, we had to educate them on it. But now the market seems to be pivoting a little bit more, and there's a better understanding that specialist models are the way to go when you're trying to solve very, very specific problems.

If organizations have been focusing on LLMs, do they need to change anything about their datasets to begin feeding them to small language models?

I think that the first thing to recognize is that the onus changes from OpenAI or Meta or Llama models [for LLMs] down to the organization [for small language models], right? The beauty of small language models is that they’re organization-specific. It's trained to own your data as such. It has your own proprietary information and ultimately becomes part of your IP stack. It becomes your intellectual property that is unique because your data is unique.

The second thing is, there are lots of nuances for enterprises to work through. A lot of their data is very disorganized. Some of it is in the EHR, some of it is in the claim system, some of it is in utilization management, some of it is in a Salesforce or custom CRM type of tool. Classically speaking, you only needed a slice of this data because you were trying to solve very, very application-specific problems. But large or small language models may need a variety of data, meaning they may need some CRM data, they might need some EHR data. The more variety you give [these models], the better they get. I noticed most organizations don't have the pipeline built to be able to support that.

Privacy also has to be a foundational component when you're building and training small language models. You can't just take this data as is and feed it to the model. You have to run it through an engine and make sure that it is de-identified. If there's a record for Alexis, then it shouldn't say that this is Alexis’s record as it's being fed into the model, because models don't forget. Once they've been trained on something, they will somehow retain that information.

The [next] thing that really has to happen is around governance. [If you’re] Kaiser Permanente, for example, you operate in Northern California and Southern California and some parts of Oregon, some parts of DC, some parts of Georgia, some parts of Colorado and so forth. So one of the things that is critical as you're building these models is to make sure that you have a representative sample from these different populations, because the population in Northern California is different than Southern California, right? Northern California will have more highly educated, higher net worth people in that population, so to speak. But if you go to Southern California, you'll have a much more diverse population there. If you just take the data as is—for example, you say, I'm going to train this model on the Northern California data—that model will be very good at identifying those people that represent Northern California and be able to make good decisions for them, but not so much when it comes to Southern California.

You have to understand that there's inherent bias in your information, because it's a function of where you operate, who you serve, what kind of products you sell, what kind of services you deliver. I think that is something that a lot of companies hadn't thought about. And I think in the world of generative AI, it becomes critical to review your data with that lens. That's an uphill battle for some organizations.

What kind of structure do you recommend that organizations have in place to make sure that they can tackle all of this, from the governance side to the bias-vetting side?

If you're a CIO or a CDO or a CEO or a CTO, one of these “Cs,” and you're thinking about how to deploy generative AI in health care, one of the decision points you have today is, do you want to go buy point solutions, or do you want to buy a platform?

The challenge that I see is the quick wins. People come from point solutions, or they've gone down that path, but very quickly they realize, hey, I need to build one thousand agents. I can't go out and buy one thousand point solutions. That's the first realization that's coming, is that a platform approach is more robust versus going with point solutions.

But then when you're talking about platforms, well, platforms come in all shapes and sizes. So you could say, I'm going to go buy AWS, or I'm going to buy a [Google Cloud Platform] GCP stack, or I'm going to buy an Azure stack, and that's my platform.

The next set of challenges that you as a CEO have to deal with [is similar to buying] a bunch of tools from Home Depot, because that's what it is. If you go to AWS or GCP or Azure and then use those tools to build a health care specific capability set—which could be done, but it's a very new technology—there's not enough talented people that are available in a lot of these organizations to be able to manage and do this work. So I think the decision point really is, are there health care specific platforms that are available? Should I buy those versus should I buy and bet on like an Azure or an AWS that has a more generic stack?

Those are kind of the decision points that people are thinking about, and there are pros and cons to all of them. But if you're looking for something that gets you ROI within 120 days, something that doesn't require you to hire $10 million worth of new talent (which is really hard to find), something that you know doesn't require you to spend 18 months buying these tools from these generic platform vendors and then trying to build a solution on top of it, the answers are pretty obvious. You have to go for a health care specific platform, or what we call a health care native platform: something that's not lipstick on a financial services platform. It has connectors to the [whole] health care ecosystem. It understands governance in the context of health care. It has small language models that are health care specific. You have agents that are health care specific.

That's the journey that I see a lot of the [health care C-suite leaders] going through. And some are early on this journey, and some are maturing quickly.

What new risks do AI agents introduce to health care? They’ll be operating autonomously, but with generative AI models, we’ve typically had humans reviewing every output.

I think it's a false dichotomy that the industry has created about agents versus humans. And I think the right answer is agents and [humans].

The way we build agents [at Penguin Ai] is that we allow the agents to be able to do the heavy lifting that's necessary, but then the final results should be reviewed by [people]. The way we think about it is like, if you think about a complex prior authorization request, it might come with 150 pages of documentation or information. If a human is doing it today, they would have to read that information, they would have to process that information, they would have to look up the guidelines, and they would match the criteria. Then they would have to make a decision, then they would have to summarize it, and then they would have to approve or deny.

In an agentic world, what we're suggesting and what we're doing is that the agent consumed that information, the agent looks up the requirements, the agent sort of matches the evidence against the requirements and then provides the recommendation that then the human can review. But instead of spending 25 or 30 minutes like in the classical process, in an agentic-plus-human process, a human would spend probably a minute or less reviewing that information.

So I think the guardrails have to be built in, at least for the foreseeable future, and we see that the outcomes are better in the long term when there's an agent plus a human doing the work, versus if you just outsource everything to [an agent].

C-Suite Shuffles

Where health care leaders are coming and going

University of Vermont (UVM) Health will lay off its COO, chief medical officer and chief of staff by the end of the year, according to an email CEO Steve Leffler sent staff on Monday, obtained by VTDigger.
- The Burlington-based system’s chief external relations officer also decided to resign amidst the transition.
- By phasing out these leadership roles, UVM Health aims to become leaner and more efficient, according to Leffler’s email.
USA Health, the academic health system affiliated with the University of South Alabama, has named Dr. Natalie Fox its CEO.
- Fox has been with the health system since 2011, previously holding leadership roles including chief physician enterprise officer. She has been serving as interim CEO for four months following the planned retirement of Owen Bailey.
Dr. Jordan Orange has returned to Children’s Hospital of Philadelphia (CHOP) as pediatrician-in-chief and chair of the department of pediatrics.
- Orange completed his residency at the historic children’s hospital and used to serve as its chair of immunology research—but most recently, he chaired the pediatrics department at Columbia University in New York City, and was physician-in-chief at NewYork-Presbyterian Morgan Stanley Children’s Hospital.

Executive Edge

How health care execs are managing their own health

Dr. Ricky Bloomfield is the inaugural chief medical officer at Oura. (Oura)

...

Wearable devices were a hot topic at this year’s HLTH conference in Las Vegas. Nearly one in three Americans use a wearable device like a Fitbit, Apple Watch, Whoop band or Oura ring on a regular basis. Now, health systems and plans are exploring ways to bring that data to care providers—allowing them to keep an eye on patients’ health metrics beyond the doctor’s office.

At HLTH, I sat down with Dr. Ricky Bloomfield, the chief medical officer of Oura. We discussed the company’s new cumulative stress feature (which launched Monday) and runs data from users’ wearable devices through an algorithm that has been validated against classic stress assessments like the Perceived Stress Scale and the Copenhagen Burnout Inventory.

Bloomfield hopes that this new metric can serve as a “window into how [users are] doing overall for stress, not just in the moment, but over time,” he told me. And as a doctor himself, he sees potential for clinicians to combat burnout with wearable devices like Oura’s.

“As a physician, I know burnout is real,” Bloomfield said. “There's an epidemic of burnout going on in the clinical community, but also across the country and the world. Many times, just seeing where you're at from a stress perspective is the catalyst that you need to make a change in your behavior. And at the end of the day, that's what we want to do: how do we give people information that helps them change their behavior in positive ways?”

I asked Bloomfield how the ring could prompt users to change their behavior as it migrates deeper into the health care space. For example, Oura will run a study to measure its ring’s ability to detect high blood pressure; Bloomfield hopes that data can be submitted to the FDA, and ultimately clear Oura as a medical device for hypertension.

He emphasized that Oura will not serve as clinical counsel, but can recommend simple lifestyle changes, like getting more sleep. The company intends to integrate new features into its AI-powered “Oura Advisor,” which can give personalized advice to users based on their history and immediate needs.

Here’s what Bloomfield told me about how he uses wearable tech to keep himself in check:

Editor’s Note: Responses have been lightly edited for length and clarity.

“I was very skeptical when I first started using the Oura ring and the app last year, when I was sort of exploring everything. And what was interesting to me is that so many wearables are all about motivating you to do more. Like, ‘more, more, more, take more steps, close your rings,’ just do more.

“The reality is, there are times when we actually need to back off and take a break and generally wearables don't do that. I started using Oura and it actually does do that. It tells me, ‘Hey, you have minor signs of illness,’ or, ‘Your readiness score is down because you just did a lot of exercise yesterday. Why don't you take it easy?’

“That's been so helpful to give myself permission in a way to take it easy, to rest. Maybe I skip some exercise, or I try to get a nap in, which is hard to do. I've seen that have a direct impact on what I do, so it doesn't need to be clinical advice all the time in order to really make a difference in someone's day-to-day.”

CEO Circle

Insights from health care thought leaders around the world

There’s no single guidebook for exemplary leadership in health care. But if you’ve ever wished that there was, check out this thought leadership article from Dr. Mazda Farshad: medical director and chair of orthopedics and spine surgery at Balgrist University Hospital in Zurich, Switzerland, and a member of Newsweek’s CEO Circle. He worked up a recipe for “excellence in specialized health care,” complete with four actionable “ingredients” that might come in handy at your own health system.

This is a preview of the October 23 edition of Access Health—Tap here to get this newsletter delivered straight to your inbox.

View 0 Comments

Request Reprint & Licensing

Submit Correction

View Editorial & AI Guidelines

Read More on Health Care