AI could utterly transform healthcare. But reduce its cost? Perhaps not so much.

As Bob Kocher, M.D., sees it, artificial intelligence (AI) will almost certainly have a transformative effect on U.S. healthcare.

It can help ensure clinicians make evidence-based decisions, facilitate better and easier communication between doctors and patients and lead to groundbreaking insights into the genetics of cancer, says Kocher, a former healthcare and economics adviser in the Obama administration who, as a partner in the venture capital firm Venrock, is used to making bets on where healthcare is headed.

But there’s one thing AI won’t do, he says: “AI is not going to make healthcare cheaper, in my opinion.”

As AI has taken center stage in the global economy, proponents of the technology generally cite two value propositions. First, it can perform certain tasks better and faster than any human. Second, those efficiencies can lead to lower prices, in no small part due to reduced labor costs.

When it comes to U.S. healthcare, though, only one of those propositions seems to be holding. AI technologies have demonstrated the ability to streamline and optimize a variety of healthcare-related applications. Yet AI’s ability to lower costs remains murky at best. Experts such as Kocher say that’s because the structural incentives built into our healthcare system make it difficult to translate AI efficiencies into durable savings.

Baumol effect

Perhaps the most straightforward way AI could cut costs is by reducing administrative expenses. Kocher’s research suggests the administrative costs associated with getting a simple medical bill paid range from about $7 to $8, including both payers and providers. More complex claims cost about five times that amount. A separate analysis suggested primary care bills cost approximately $20 per bill.If billing — but also matters such as scheduling and routine paperwork — could be streamlined with AI, health systems could theoretically do more with fewer administrative staff.

Yet, Kocher says most of the existing tools he has seen are better described as “widgets” than as “workers.”

“They can take part of what we do and then make that part better, but they can’t replace you today,” he says.

Nikhil Sahni, MBA, MPA/ID, who leads McKinsey & Company’s Center for U.S. Healthcare Improvement, says healthcare is a prime example of a phenomenon known as Baumol’s cost disease, named for the American economist William Baumol, who formulated the theory. Baumol’s theory explains why labor costs in industries such as healthcare and education tend to rise even without productivity gains. By way of example, Sahni notes that the cost of hiring a string quartet to play a musical composition in the year 2026 is way more than what the cost of hiring a string quartet was in the year 1600.

“The wages are going to go up with inflation,” he says, “but you don’t get more labor productivity.”

Healthcare is naturally a labor-intensive industry. Simply equipping staff to perform the same inefficient processes using AI tools won’t necessarily yield significant savings, he says. Instead, he says healthcare executives should use the advent of AI to develop new, more streamlined processes.

“That’s where I think the labor productivity gains are going to come from, ultimately — process redesign, not necessarily AI taking away tasks from people,” he says.

Collaboration required

Sahni says many of the AI-fueled advances in healthcare administration to date have been in areas where organizations can act alone, such as using AI for nonclinical patient contact or streamlining payment-related processes such as billing and prior authorization. Deeper, system-level efficiencies will be more difficult, he says.

“I think the next wave is going to require opening up new data sets that probably require some kind of industry collaboration,” he says. “And then I think eventually the clinical stuff is further down the road.”

He notes that the health insurance trade group AHIP has already pledged to work with its members to streamline and standardize prior authorization. However, industry-wide collaboration could be more complicated on the provider side, since it would require thousands of the country’s health systems and medical practices to agree on standardized protocols.

Cascades of care

Although many healthcare organizations have been slower to introduce AI for clinical care applications, there is no shortage of ways AI could improve care, according to Kev Coleman, who directs the Health Care AI Initiative at the Paragon Health Institute, a public policy think tank.

“One of the things that’s exciting about artificial intelligence, particularly around clinical applications and medical devices, is that it takes our traditional detect-and-treat paradigm and complements it with a secondary paradigm, predict-and-prevent,” Coleman says.

He notes that as of this summer, the FDA has authorized the marketing of more than 1,500 AI-enabled medical devices. Many of those devices are focused on early detection of cancer and other diseases. But there’s a problem. “The big issue,” says Coleman, “is reimbursement.”

A fundamental issue with any intervention whose primary benefit is preventing disease is that it can be difficult to calculate its impact, both in terms of health outcomes and cost savings. Absent clarity on those factors, it is hard to know how much — if anything — such AI tools are worth.

Coleman says there are a limited number of AI tools that currently receive stand-alone reimbursement. He says AI tools are typically billed using temporary billing codes that do not include a standard or guaranteed payment. That’s not good enough, he says.

“We really need to have a more strategic focus on our use of AI,” he says.

Another problem is that early detection does not necessarily lead to lower costs on a systemwide level. Yes, one patient might avoid hundreds of thousands of dollars in costly care because an AI algorithm identified an anomaly in a radiological image. But such anomalies could lead other patients to seek unnecessary low-value care.

“This is where the doctors are having trouble with [AI],” Sahni says, “because they see those cascades of care and they’re like, this is going to trigger more cascades, not prevent cascades from happening.”

Adds to costs

In theory, AI could help sniff out fraud, waste and abuse. Coleman says payers can leverage AI solutions to identify inconsistencies associated with unnecessary procedures or upcoding. They can then resolve those inconsistencies and save money. Yet, he notes that payers’ vision into the operations of individual healthcare organizations is limited.

“It’s harder to detect things like resources that were never delivered, especially if you don’t have access to hospitals’ procurement systems,” he says.

Moreover, he notes that if a health system realizes significant administrative savings using AI, payers generally have little way of knowing about those savings or benefiting from them.

“Unless you have competitive cost pressures, if you introduce a technology that is going to reduce your overhead, what is your incentive to pass that along to the consumer?” he says.

Kocher’s research suggests that AI is more likely to increase healthcare costs in the short- and medium-term under the current fee-for-service paradigm. He says the fee-for-service system is not designed to reward proactive vigilance.

He uses the example of wearable devices. Devices such as Apple Watches and Oura Rings are commercially popular and can enable remote monitoring and provide clinicians with a wealth of personal health data. Yet, most clinicians do not utilize such data. “It costs them money to look at the data and call you if something is off,” he says.

AI could change that, since it can autonomously monitor and interpret the data and alert patients and clinicians of any anomalies.

Not only that, Coleman says similar tools can be used to help patients substitute telemedicine for in-person visits or be monitored at home with hospital-at-home procedures, rather than occupying a costly hospital bed.

Yet, Kocher says even those efficiencies may not translate to lower healthcare costs.

“If AI makes us all healthier — which would be a wonderful thing — and the demand goes down for hospital beds, the price will go up,” he says, “which is the opposite of what happens in a competitive market.”

After all, hospitals still have to keep their doors open. They still need highly trained medical professionals. They still need high-cost equipment. And they still need to pay costly utility bills.

Kocher compares it to the first decade of the 2000s, when the U.S. saw a wave of specialty hospitals crop up, siphoning off the highest-margin procedures from traditional hospitals.

“They took out a bunch of patients from the regular hospital, and the regular hospital still had to run, and so it raised the prices to offset the lost revenue on every other patient in the hospital, and so in general, healthcare got more expensive,” he explains.

Shifting toward value

Kocher sees hope in value-based care, a model that has been tried in pilot programs and by some insurers. A 2023 report from AHIP found that approximately 4 in 10 insured Americans were enrolled in a plan with an “advanced payment model.” AHIP defines advanced payment models as those that use a value-based care structure or supplement a fee-for-service model with incentives or penalties tied to value-based outcomes.

Kocher says such models could incentivize health systems to use AI tools aggressively to optimize care.

“The continued emphasis toward pushing value-based care for health systems is the most powerful thing they could do to get AI to be adopted in ways that I think they would find beneficial,” he says.

The CMS could also incentivize the use of AI by, for instance, giving clinicians enhanced protection from malpractice claims if they run a diagnosis, treatment plan or course of care through an AI tool before finalizing it, Kocher says.

“[AI] could do a bunch of quality and safety checks that might reduce errors and make care more likely to be effective,” he says.

Still, Kocher says the primary impact of such regulatory changes would be to improve care. Whether and how much money it ultimately saves could be difficult to quantify, since such savings may only manifest as fewer malpractice claims or better long-term outcomes.

AI hospitals

Coleman says regulators should do more to enforce pricing transparency, data accessibility and standardization.

“If you’re not going through and monitoring noncompliance and doing hefty fines for noncompliance, you’re going to get a bad situation,” he says.

However, he says regulators also need to be careful not to misregulate. For instance, although keeping a human in the loop is considered very important in many situations, overly broad rules that require human involvement in easily automated aspects of healthcare might add unnecessary costs as technology improves.

Coleman believes the best way to optimize AI and achieve durable savings is by building a better system that other healthcare organizations will want to emulate. He notes that China is already experimenting with “AI hospitals,” which aim to optimize the healthcare journey through the use of AI, including leveraging telemedicine and AI intake and triage.

“To achieve transformative savings in healthcare for Americans, we need a parallel competing system,” he says. “And it can’t be the same one, slightly improved. It has to look very different.”

Sahni says an often-overlooked component of AI optimization is change management, the art of helping people within an organization adapt to new paradigms. While he believes designing new processes is better than inserting AI into existing, outdated processes, he also argues that if organizations change too much, too fast or with too little buy-in, they are setting themselves up for failure.

Indeed, while considering the overall cost of healthcare is an important macroeconomic topic, Sahni says most healthcare organizations focus on different questions when considering AI implementation.

“It’s not about savings. It’s ‘Is this improving patient experience or clinician experience?’” he says. “That’s the start[ing] premise.”