Two recent developments have sharpened a question health systems will eventually have to answer: How much care can AI deliver on its own?

A Sept. 28 McKinsey & Co. analysis estimated that AI can perform the care represented in 16% to 22% of U.S. outpatient claims, led by low-complexity evaluation and management visits. The firm also said payers stand to gain more than providers as AI becomes patients’ “front door” to care. A day later, HHS Secretary Robert F. Kennedy Jr. said at the MAHA Summit in Washington, D.C., that AI offers “a second opinion that is much better informed than any doctor in the country.”

Becker’s asked CIOs, chief medical information officers and other health system leaders whether McKinsey’s estimate is realistic for their organizations and what would need to be true before they let AI handle more of that care. Four also responded to Mr. Kennedy’s remarks.

The leaders split on the number. TidalHealth’s CIO called it realistic for primary and urgent care, and several others called it plausible in theory. Parkview Health’s CMIO put the real figure closer to 1% to 2%, and Stanford Health Care’s chief data scientist said it doesn’t fit an academic medical center. On what has to come first, though, five of the eight landed in the same place: Someone has to be accountable when AI gets it wrong.

Here is what they told Becker’s in written responses, edited for length:

Nigam Shah, MD, PhD. Chief Data Scientist at Stanford Health Care (Palo Alto, Calif.): It’s probably not accurate for a tertiary care facility like ours. Low-complexity, outpatient care isn’t the biggest volume of care done at an academic medical center.

I’d interpret [Mr. Kennedy’s] comment as “much better than the average doctor,” and the essence of the comment makes sense. Our physicians would gladly engage with such a request. Some of them are already leading the evaluation of the quality of advice offered by a broad range of AI models.

Mark Mabus, MD. Chief Medical Informatics Officer and Senior Vice President of EHR at Parkview Health (Fort Wayne, Ind.): I am very doubtful that we are actually already at 22% of all outpatient claims. My generous estimate is more like 1-2% of primary care/urgent care only. This is one of the flaws of the [McKinsey] analysis: They say that the effect will vary by care setting, but care setting is not the same as specialty. Additionally, the study lumps all specialties of providers together.

The bigger flaw in this study is its reliance on claims data, which cannot capture what actually occurred during a patient encounter. Claims are designed for reimbursement, not to represent the full clinical interaction.

Deepti Pandita, MD. Chief Medical Informatics and AI Officer at UCI Health (Orange, Calif.): My reaction is that the [McKinsey] estimate is directionally plausible, but the reality is far more nuanced than a single percentage suggests. No two health systems are alike, particularly when you consider differences in patient populations, payer mix, state regulatory requirements, care delivery models, and access to digital infrastructure.

The greatest value will come when incentives across payers, providers, and patients are aligned rather than optimized for any one stakeholder.

Ashis Barad, MD. Executive Vice President and Chief Transformation and Digital Information Officer at Hospital for Special Surgery (New York City): AI can recommend care; the harder question is who stands behind that recommendation. I see tremendous opportunity here. But someone has to decide what good care looks like, translate that into instructions AI can reliably follow, and take responsibility when a patient acts on its advice.

Before we talk about what percentage of care AI can deliver, we need clear answers to those questions. Patients need to know who is accountable for their care and how to reach a clinician when they need one.

Mark Weisman, MD. CIO of TidalHealth (Salisbury, Md.): I believe it is realistic for primary care and urgent care. There are components to a visit that AI can do as well as anyone else, like recommend preventative care based on guidelines. Routine maintenance checkups for chronic stable conditions like hypertension, hyperlipidemia, and hypothyroidism can certainly be managed by AI.

The other challenge is liability. Who will be responsible for the outcome if the AI is wrong? The AI vendor is not going to take that liability and there is little incentive for the doctor to take on that risk. There is not enough money in these low level visits to pay for the AI and take the liability risk.

The ability to have the world’s knowledge available at a doctor’s fingertips is very appealing and AI makes that possible. However, AI is a tool and like all tools a doctor may use, it has its place. The models learn from a variety of sources and the ones which learn from Reddit and X will have different results than ones which learn from randomized, blinded clinical trials in peer-reviewed literature. The source of the recommendation from AI matters and must be reviewed. We cannot blindly accept what AI generates today.

Muhammad Siddiqui. CIO of Reid Health (Richmond, Ind.): AI can help us expand access and reduce clinical workload. McKinsey’s estimate is plausible as a measure of potential capability, but we haven’t validated that percentage for Reid. And Secretary Kennedy’s claim goes further than the evidence supports.

The claim that AI can provide a better-informed opinion than any doctor is too broad. Reviewing more information can be useful, but it doesn’t establish that AI understood the record or reached the right conclusion.

Nipa R. Shah, MD. Chair of Community Health and Family Medicine at University of Florida College of Medicine-Jacksonville: McKinsey’s estimate feels realistic for selected low-complexity visits given LLMs’ capacity at this time, and my limited experience utilizing validated AI-based tools. … For our organization, 16-22% is plausible if we define “perform” carefully. AI can replicate many steps in care, but healthcare is not just a series of steps.

Secretary Kennedy’s comments highlight AI’s remarkable ability to process information, but information is not the same as wisdom. In our organization, we are preparing for patients who arrive with AI-generated opinions (Dr. AI) by treating them as informed starting points for discussion, not final answers. … The future may be less about doctors versus AI and more about doctors helping patients navigate an abundance of AI-generated advice.

Darrell Bodnar. CIO of North Country Healthcare (Whitefield, N.H.): I do think the [McKinsey] estimate is plausible, although I would distinguish between what AI may technically be capable of performing and what we should be comfortable allowing it to perform independently.

I also think the financial question is important. If AI becomes the front door for low-acuity care and most of the economic benefit accrues to payers, we risk weakening the delivery systems patients still depend upon when their needs exceed what AI can manage. That is particularly significant in rural healthcare, where maintaining access and clinical capacity is already challenging.

I believe AI will deliver more care directly over time. The measure of success, though, shouldn’t be how much healthcare we can remove humans from. It should be how much safer, more accessible and sustainable healthcare becomes because humans and AI are working together.