At Spanish Peaks Regional Health Center in rural Huerfano County, medical claims submitted by the nonprofit health facility are denied daily by insurance companies.
Sometimes, someone wrote in the wrong code or forgot to check a box. Other times, the insurer requests more information. Whatever it is, it takes human employees to dig into denial codes, research the cases and submit an appeal. Administrative staff just can’t get to all of them.
“We go after the high-priority or the high-dollar ones first,” said Daron Hashir, chief financial officer of Spanish Peaks Regional Health Center. Unfortunately, she said, “if appeals aren’t resubmitted in a certain amount of time, you automatically lose it.”
That led Hashir, who joined Spanish Peaks in February, to start a trial with Denver-based Iterate AI in hopes that modern artificial intelligence could recover missing insurance payments. Iterate AI’s system created software tools, or “agents,” that learned how to identify various delays, billing errors and missing payments. Agents also analyze pages and pages of the denial documents. It then gins up an appeals letter pulling in language from the insurer’s contractual obligations — in minutes.
The Iterate trial, which analyzed a limited amount of the hospital’s historic data, proved its worth.
“I mean, easily, we’re working with losses up to hundreds of thousands of dollars,” Hashir said. “That could be the difference between us being able to continue certain service lines like, for example, our acute services. … When you start picking off some of these denials, that adds up really quickly, especially when they average anywhere between five to $25,000 per claim. Even getting half of that back is going to be instrumental.”
Denied medical claims is a big issue for patients. The Commonwealth Fund, an organization that promotes equitable healthcare systems, found that 1 in 5 patients with private insurance had experienced a coverage denial for themselves or a family member. That cost patients more out of pocket, created worry and anxiety and delayed medical care. Only about half appealed, according to the report.
It’s also a challenge on the business side, where a growing number of claims to health insurers are also getting lost. In an analysis of healthcare revenue, health-tech consultancy Kodiak found that the rate of initial denials by insurance companies increased to 11.6% last year, from 10.2% in 2021. It also contributed to the $48.4 billion in lost revenue for billable services last year, a 25% increase from the prior year.
The allure of modern AI has healthcare companies and other industries eager but cautiously moving forward to reclaim lost revenue. The technology has propelled the investor market to new heights even as it has created legitimate concerns about accuracy, privacy and employment. But for many companies, adapting and adopting modern AI as a tool is seen as almost a necessity. And often regulations are in place to protect private data, like personal medical information, as well as other technology to secure and protect data.
“AI is different than other technologies in the way that it’s being used, the speed at which it can produce outputs and the way it uses data that may be trained on different data,” said Alex Kimata, an attorney at the Boulder office of Holland & Hart who advises clients on data, privacy and AI governance. “But a lot of the guidance is that it’s not different from (cloud services). You need basic building blocks for how you protect the data. Encryption, access controls and making sure you vet your vendors. That is all the basic stuff we keep recommending companies contemplate when they’re using AI systems.”
What is private AI?
Private AI systems, like Iterate’s Generate Health, are finding their way inside companies where protecting sensitive data was a priority long before the new generation of AI emerged.
Private systems are very similar to the well-known public large language models, or LLMs, built by OpenAI, Anthropic, Meta and others. But where large public models are gulping down as much public and private data as possible to better understand how humans communicate, private systems are cut it off from the outside world.
They operate within corporate firewalls and specific devices and servers, also known as staying “on premise.” The private data doesn’t leave, though the LLM base can be updated with new learnings from the mothership.
Private systems are seen as a way to rely less on Big Tech, which don’t always prioritize privacy, said Rory Mir, director of open access and tech community engagement at the Electronic Frontier Foundation, which advocates for privacy and free speech.
“We are seeing that these smaller models are improving quite rapidly as well, and this sort of reliance on Microsoft or Google or the handful of big players isn’t that necessary,” Mir said.
But there are protections companies should take as part of good data-security hygiene.
“When you’re sending information off premises to a different company, you are really relying on the precautions they take to protect your data, not just what they say they will do or won’t do, but also on a technical level. Are they securing it? Are they making sure it’s not accessing things that it’s not supposed to access,” Mir said. “That is where the ambiguity of what private AI comes in because most AI, unless it’s on premises, is going to someone on a server and sharing information in a way that the person hosting it can access.”
Kimata, with Holland & Hart, advises companies to address such concerns in the vendor contract, and make sure private data cannot be used to train public AI models. More importantly though, he said, AI is a tool that should be used for a specific purpose.
“A lot of our clients are like, we just want to use AI for everything. And we’ll say, well, what does that actually mean? And they’re like, we don’t really know yet, but we’re kind of building toward that. My response often is this should be intentional and it should be customized to how you want to use AI, including what vendor you use, what program you use and what controls you put on it,” he said. “This idea of I just want to be an AI company quote-unquote, doesn’t really make sense.”
The Spanish Peaks AI trial
The nonprofit Spanish Peaks operates a 20-bed hospital, a 90-bed nursing home and several clinics near Walsenburg. Patients who can’t afford to pay still get the care they need. But when insurers pay the claims filed, “it prevents us from having to write it all off,” Hashir said.
About 90% of its insurance claims are paid after being submitted the first time. That’s a typical rate for a rural hospital of its size, she said.
That other 10%, though, is what she’s trying to make sure the nonprofit hospital doesn’t miss out on.
“I’m trying to break even every year,” she said. “That’s what I’m supposed to do.”
Iterate has its origin in retail, having launched in 2013 by eBags cofounder Jon Nordmark. The technology company helped consumer brands and retailers like Ulta Beauty and Circle K increase revenues to better compete with retail giants like Amazon. They built software tools to track orders, provide recommendations and more using its own artificial intelligence system. Targeting healthcare companies with private AI is just the latest tool, and it’s benefited from advances in agentic AI, which goes beyond generating responses in a conversation. Agentic AI can take action.
A hospital’s finance department is “chasing bills all day” because they’re submitting invoices to insurance companies that return them with obscure responses, said Kevin Homer, Iterate’s vice president of sales. That’s where AI excels — it processes large amounts of data, figures out what to do next and then does it. Healthcare billing seemed ripe for this service.
“It’s not as much about coding as it is volume. They can’t get to the data. They can’t get as many claims out the door that are coming in,” Homer said. “We’re working with another large healthcare provider, they’re not a client yet, but we’re working with a large provider in the Southeast. They sent us $27 billion worth of claims. It was 10 million rows of data. They have 20 to 25 people in the revenue cycle department, which is a big revenue cycle department. How are 25 people going to review 10 million lines of code?”
Iterate’s system uses AI to match up claims, accounts, the insurer’s responses and all the supporting data and then tries to figure out what’s not making sense. The technology also types up appeal letters with supporting data so everything is ready to go and just needs an administrator’s review, instead of staff having to hunt down the data to write it up.
“They’re having to go into the claims, which may be 20 pages long. And they’re trying to find why the payers are denying it,” Homer said. “Insurance companies have stopped using the words denied or denial because they’re too easy of keywords.”
When the insurance company doesn’t pay the claim or delays payment, the hospital may not even notice because without a denial code, it doesn’t show up in a standard tracking system.
Hashir calls these “ghost denials.” The insurer didn’t deny the claim outright so it doesn’t show up as denied. Sometimes, she said, a payer can pay the claim but then can take all or part of the payment back without issuing a denial code.
“The common thread is that the dollars are gone but the denial doesn’t surface in the standard denial reports a revenue cycle team would routinely watch. That is what makes them hard to chase,” she said. “Quantifying it cleanly across thousands of claims is exactly what a tool like Iterate is designed to do, which is why we engaged them.”
Iterate only had access to partial historic data so calculating the potential value for Spanish Peaks was not available. But the results helped Hashir better understand how the technology could identify issues quickly and help speed up the claims process. She’s also working with Iterate to develop a new AI agent that’ll identify doctors and medical professionals who need more training on filling in accurate codes so the hospital gets paid.
“There’s a lot of money that gets left on the table,” she said. “This is a tool to really help us recover that and start having the conversations with insurance companies and have more data to back up the findings when we’re getting either chronically underpaid or not paid at all.”