Payers say AI doesn't deny care. The truth is more complex.
Payers say AI is not clicking the deny button on prior authorizations and provider claims, but the technology does play a key role in deciding which requests need further review.
Artificial intelligence does not determine whether care is approved or denied, UnitedHealth Group's Chief Financial Officer Wayne DeVeydt told reporters at the company's second quarter earnings call in July.
This is a sentiment echoed by other payers, including Cigna and Aetna, which have maintained that AI is never used to deny claims or care. Still, insurers continue to shovel billions of dollars into AI investments to make their operations more efficient, particularly for highly administrative tasks, like prior authorizations and claims processing.
The situation is more nuanced than payer executives make it seem, according to Stacie Sutter, vice president of payer strategy at Ensemble Health Partners, a revenue cycle management company that handles medical billing and claims management for providers.
What payers are saying about AI and care denials is technically true, Sutter explained. AI is not clicking the deny button, she clarified, but the tools are routing which cases are seen by an actual clinician and the condition those claims are in when they get there.
"What I'm saying is that technology determines which cases enter the denial risk pathway," she stated.
A tale of two requests
Sutter explained that prior authorization requests fall into two categories: those that can be auto-approved by the AI and those that are actually denied by the AI.
"Auto-approved requests are cases in the payer system that are considered straightforward and low risk. Those are going through the system, which means the physician is not reviewing a random sample of requests," she explained. "What's happening is that the other cases are being routed for additional scrutiny."
AI doesn't deny the requests in the traditional sense, rather it determines that it doesn't fall within its technical parameters for approval and must be reviewed by a human. The workflow changes from a subject matter expert reviewing random samples of auto-approved requests to ensure accuracy to digging into a specific case.
This changes the dynamics of the review process, Sutter maintained.
AI has transformed from more basic, rigid software used by insurers to flag or deny claims to more sophisticated generative models and EHR integration that actively draft submissions, handle multistep appeals and comply with federal transparency rules.
"I want to go on record that I'm not suggesting physicians are biased with this. What I'm saying is that technology determines which cases enter the denial risk pathway," she said. "The computer may not write the denial itself or the denial letter, but it often helps determine which of these cases flows through as not high risk versus which ones pause for provider review."
A dilemma for providers
Providers don't necessarily know which of their prior authorization requests will trigger the AI denial. This creates a dilemma for providers seeking to treat their patients and be paid for that care: They can invest more time and resources into adding clinical detail upfront to try and route more of their requests toward auto-approval, with no guarantee it'll work, or submit standard documentation and face a higher likelihood of denial or delay.
There is no winning option in this case, Sutter suggested, because of the information asymmetry associated with payers' use of AI for claim processing. Providers are making decisions without knowing what the AI is looking for, as payers frequently claim that knowledge is proprietary.
"We're pitching to a black box," Sutter stated.
Even when providers think they have figured it out, criteria can change, Sutter continued. She likened it to chasing ghosts, leaving providers constantly having to reverse-engineer to keep up with AI-enabled decisions.
"That's why providers are spending the time and resources trying very hard to get it right on the front end," she said, referring specifically to recent provider investments in improving clinical documentation.
The use of ambient technology and AI scribes has exploded in the last year, giving providers back facetime with their patients while better capturing billing codes. But this has created a so-called AI arms race between payers and providers, especially with claims processing. The tools are also not necessarily alleviating the issue at hand; rather, they are shifting it.
AI tools for clinical documentation improvement, prior authorizations and claims processing are designed to relieve the administrative burden of completing the tasks. Without visibility into the black box, though, payers' AI is placing a greater burden on providers as they try to figure out how to submit clean requests.
Opening AI's black box
Transparency is the key to understanding payer AI, according to Sutter.
"Transparency starts with understanding where and how AI is being used in the decision-making process, such as authorizations. Providers also need to know the algorithms influencing the routing of cases," she said.
The question is whether AI is being used to make healthcare more efficient or whether it's being used to make denials more efficient.
Stacie Sutter, vice president, payer strategy, Ensemble Health Partners
More specifically, Sutter wants payers to reveal what factors the AI uses to evaluate cases, especially those routed to clinicians and how determination outcomes differ between AI-screened requests and those that aren't. She also urged payers to share approval rates, denial rates, overturn rates, turnaround times and utilization trends.
"All of those for the [requests] routed versus the non-routed, and the results should be broken down by service type and specialty," she stated. "This is all feedback that the provider needs and deserves to understand the use of AI for transparency, to make sure that we are checking all those boxes on the front end."
However, Ensemble has had some success implementing technology to automate the process. Sutter shared that the company is working with a national payer on touchless prior authorizations -- meaning neither side touches the request; an automated solution does the heavy lifting.
"We know what fields need to be filled in. Those are the fields that are going to go to the payer. Within an hour, it comes back. We've had a 92% success rate with this," Sutter said.
This success has shown Sutter that payers want to partner to automate prior authorization, one of healthcare's most notorious processes. But she clarified that this applied only to radiology cases, a high-volume specialty. The company has yet to move beyond that immediate scope, she said.
Technology systems have to be robust enough to handle scaling prior authorization automation. This is why Ensemble has worked with Epic, the leading provider of EHR systems in the U.S., to automate more requests.
"But the payers also have to be engaged," she noted.
Still, AI is a promising tool for improving claims processing. The technology can significantly accelerate a transaction like prior authorization by automating data extraction, matching medical records to payer rules and enabling real-time approvals. More revenue cycle management vendors and health IT companies are bolstering AI for prior authorizations to deliver more value to customers.
The Centers for Medicare and Medicaid Services has even tapped AI for Medicare prior authorization requests in certain services. The AI-driven workflow is currently being tested in six states, although it has been highly contested for its low approval rates and potential incentives to deny care.
"AI absolutely does belong in healthcare. The question is whether AI is being used to make healthcare more efficient or whether it's being used to make denials more efficient," Sutter said. "Without transparency, it's really impossible to know the difference."
Jacqueline LaPointe is a graduate of Brandeis University and King's College London. She has been writing about healthcare finance and revenue cycle management since 2016.
