Picture Alliance | Picture Alliance | Getty Images
AI is penetrating deeper into the already expensive and complex U.S. healthcare system, where it is being used not only to help doctors diagnose and treat patients, but also to determine how much hospitals will bill, what insurance will pay, and which claims will be denied.
And even though AI is being talked about as a productivity tool that will lead to efficiency and drive deflation, some of the early evidence raises questions about whether these tools will make already expensive systems even more expensive.
The Blue Cross Blue Shield Association recently estimated that hospitals’ use of AI-assisted medical coding contributed nearly $1 billion ($942 million) in additional costs to health plans from 2023 to 2025. BCBSA says much of that increase is due to secondary diagnoses that move patients into higher reimbursement categories, types of diagnoses that “may be derived from a single test value and are particularly amenable to detection by AI tools.”
The findings come as hospitals, insurers, and other parts of the health care system increasingly use AI to bill, code, and review claims, raising questions about whether reducing administrative burdens can also strengthen the financial incentives already built into the system.
Christopher Whaley, a health economist at Brown University who studies hospital coding, said AI appears to be “accelerating and in some ways making the existing underlying billing incentives in the system easier to obtain.”
Additional diagnoses identified by AI are not necessarily inappropriate. “In many cases, the diagnosis is legitimate and goes undetected,” he says.
However, he said there are conditions that are “less clinically important and do not impact patient care” that still allow a different billing code to be applied and increase payments.
Blue Cross Blue Shield AI findings
BCBSA wrote in its analysis that the rise in what it calls “complex coding” occurred at a time when 60% of hospital systems began using AI coding tools. “There is a clear disconnect between coding and treatment,” the report states.
Luke Chalker, BCBSA’s senior vice president of product and data science, told CNBC that about 70% of claims identified by insurers, or $653 million, were related to additional diagnoses that did not result in a change in care.
Chalker could not attribute the entire increase to AI. “Multiple factors contribute to coding strength, and our findings suggest that AI-powered coding and documentation tools play a role,” he said.
He added that consumers have reason to be concerned. More complex coding can lead to higher refunds if proper care is not taken. “These costs can ultimately come in the form of higher premiums and out-of-pocket costs,” he said.
Health insurance costs per employee are expected to rise by an average of 8.2% in 2027, the largest increase since 2003, according to the latest forecasts from benefits consulting firm Marsh.
The American Hospital Association rejected BCBSA’s analysis. “Today’s patients are older and more clinically complex,” and AI tools are helping health care providers “better understand patient conditions to support care planning. BCBSA’s analysis lacks the context needed to meaningfully assess how these tools impact quality of care, patient access, and spending,” an AHA spokesperson said in a statement to CNBC.
The AHA spokesperson added, “It is particularly concerning that insurers are raising concerns about provider coding while continuing to rely on automated downcoding and denial practices that can prevent coverage of medically necessary care, increase burden on employees, and increase costs from administrative waste.”
Chaulker said Blue Cross Blue Shield companies also use AI for claims review, but “clinical denials are always reviewed by qualified human clinicians.”
AI’s “Administrative Arms Race” in Healthcare
Increasingly sophisticated AI tools being used in all aspects of the health care system could create what Whaley calls an “administrative arms race.”
“Both of these tools and technologies, whether it’s on the hospital side or the insurance company side, are very expensive and have nothing to do with providing proper care to the patient,” he said.
Whaley said those costs ultimately flow through the health care system to consumers through higher insurance premiums and taxes.
Marisa Greenwald, a partner in Oliver Wyman’s health and life sciences practice who advises hospitals and insurance companies on strategy, operations and AI implementation in areas such as revenue cycle management, said AI is already benefiting health systems and physicians.
By reducing administrative tasks, doctors can spend more time seeing patients and less time completing documentation during off-hours. Greenwald said health systems are using AI not only to improve efficiency but also to “bring back some semblance of work-life balance” to doctors.
There are also financial benefits. Greenwald noted that more time with patients allows physicians to increase their patient volume, while more accurate coding allows them to “capture additional acuity and diagnostic factors,” resulting in higher reimbursement amounts.
But she said it remains difficult to know how much of this increase truly reflects improved documentation compared to other factors. “It’s hard to work out how much accuracy will be improved. There will always be misuse, user error, and overcoding.”
Greenwald added that if providers leverage AI to improve coding and insurers respond with their own AI tools, leading to further denials, “the arms race will get even worse. The hope is that stakeholders on both sides of the equation realize that all we are doing is adding cost and burden to a system that is already challenging and labor intensive.”
Otherwise, she added, healthcare could become “robots talking to robots and just fighting each other.”
Tensions between hospitals and insurance companies are not new.
Vanessa Moldovan, author of The Healthcare Revenue Cycle AI Playbook and head of RCM strategy at Magic, which develops AI-powered automation software for healthcare management, said many of the tensions surrounding medical coding existed long before AI.
“AI is new, but everything else is not new,” she says. “You can line up 10 programmers and have them all look at the same chart and code it differently,” she added.
At the same time, providers operate under detailed insurance company requirements that govern which diagnoses and services will be reimbursed. AI allows organizations to review far more records and documents faster. This allows health care providers to be compensated for the care they actually provide, Moldova said.
But she drew a clear line when the diagnosis was not supported by medical records. “If the patient didn’t have those symptoms, then they didn’t have those symptoms,” she says. This is one of the reasons why Moldova opposes fully autonomous coding. “Humans should always be involved.”
He said AI-generated coding should be audited in the same way healthcare organizations have traditionally audited the work of human programmers.
Chalker agreed, saying AI should “support decision-making, not replace human judgment.”
The risk goes beyond whether hospitals receive large payments. Diagnoses could become part of a patient’s medical record, and Moldova said he was concerned that people would place too much trust in information generated by AI.
“I think we’re in danger of trusting AI too much, because we think, ‘Wow, this is AI, it should be smarter,'” she says.
But Moldova also sees AI as a tool to help insurers respond to denials. Revenue cycle teams often lack the resources to respond to document requests and appeals, she said, explaining that AI can help them respond at scale.
“We can file more appeals. We can file more appeals,” she said.
This allows AI to be leveraged at multiple stages of the same claim dispute. This means helping providers manage documentation and codes, helping insurance companies review claims, and helping providers appeal denials.
But Browns-Whaley said the ultimate test should not just be whether AI reduces health care costs.
He said it could still be worth it if the technology increases spending while simultaneously improving access and quality of care.
“If an AI tool is just going to be used to shuffle the cards a little bit more and make sure you have the upper hand without doing anything to the patient, then it’s probably not worth the investment of resources,” Whaley said.
