Medicare is getting a hard lesson in the cost of administrative obstacles to care, courtesy of Artificial Intelligence (AI).

More Perfect Union reports that traditional Medicare is introducing AI-driven prior authorization requirements, similar to those already utilized in privately run Medicare Advantage plans. In several cases, AI was shown to lead to high rates of treatment denial, leaving patients without care unless a clinician overruled the algorithm.

Worse, the program’s payment structure rewards the AI contractors with percentages of the savings from denied claims, inarguably incentivizing refusal of treatment.

Since clinicians are burdened with reviewing AI-generated treatment denials, correcting what appear to be significant numbers of incorrect denials can create more work for doctors, not less. One provider reportedly complained that “the additional administrative burden resulting from these process errors was negatively impacting their staff.”

AI cannot think, critically or otherwise. It can make algorithmic predictions based on training data, but it cannot act rationally, let alone compassionately, making it profoundly unqualified to serve as a physician. It cannot abide by the Hippocratic oath, because not only is it incapable of understanding what it means to “do no harm,” it is incapable of understanding, period.

The expansion of prior authorization into Medicare is predicated on the notion that it will increase efficiency. As we’ve seen repeatedly, placing obstacles between patients and the care their physicians recommend is just as likely to have the opposite effect.

Add deeply fallible AI programs to the prior authorization equation, and the results can be regrettable.

In healthcare, as in many industries, AI can create more problems than it eliminates by making errors humans wouldn’t make and creating messes humans must clean up. However, the private equity interests behind AI authorization technology benefit regardless of how those technologies affect people’s health.

With only vague accountability mechanisms and no direct financial penalties for incorrect treatment denials by AI contractors, even at scale, the oversight of this program so far is inadequate to say the least. As a result, doctors and Medicare patients must be prepared to fight the machine.

Medicare and the Plague of WISeR

Historically, Medicare treatment is subject to prior authorization only in narrow circumstances. But this year, Medicare launched the Wasteful and Inappropriate Service Reduction (WISeR) program, establishing prior authorization in traditional Medicare, similarly to Medicare Advantage plans (which, though carrying the “Medicare” moniker, are administered by private companies).

According to KFF, a preeminent non-profit health policy research firm, insurers administering Medicare Advantage plans have faced lawsuits and congressional investigations over AI tools that “substantially increased denial rates” and “often operated without human oversight.”

Traditional Medicare is now infected with this model. According to HEALTH CARE un-covered, the results have been unwarranted denials, treatment delays, and delays of reimbursement to providers, as evidenced by documents that a Freedom of Information Act (FOIA) request forced CMS to release (emphases ours):

“...although the documents only cover the first three months of the pilot, they paint a damning picture: thousands of inappropriate denials and extended wait times for patients in two of the six states, thousands of providers waiting on payments that were weeks or months late, and complaints by individual patients that went all the way to the state Attorney General, and eventually, Congress before being resolved in the patients’ favor.”

Those documents reveal a disturbing number of AI-driven “Non Affirm” decisions (i.e., decisions not to approve the requested care) in response to prior authorization requests, which human clinicians must then evaluate. As shown below, one of the documents pried from CMS under FOIA shows “Non Affirm” rates for two WISeR Model Participants (WMPs):

• Virtix Health in Washington State refused to “Affirm” 53% of requests
• Zyter in Arizona refused to “Affirm” 19% of requests

Moreover, WMPs often took far longer than the CMS-mandated 72 hours to respond to treatment requests. Though the CMS documents obscure the real timelines by making “1-5” days a single category, thousands of decisions took between 6 and 10 days, with hundreds taking more than 10 days, for just two WMPs in two states:

The Profit in Delayed Care

According to the CMS’s WISeR Participant Guide, the AI vendors participating in the program receive up to 25% for “each payable non-affirmation,” i.e., a cut of the value of each treatment their models deny.

As a memorandum from the CMS Office of the Actuary bluntly states (emphases ours):

“The participants are paid based on the claims denied. The ICIP mentions a quality adjustment and while the determination of this adjustment has not yet been fully specified, it appears as though it would not significantly change the portion of savings the participants will receive. As a result, model participants will have an incentive to deny as many claims as possible.”

Medicare is pioneering a new frontier in profitable friction, predicated on a fiction of efficiency. As we’ll explore in a future article, California workers’ comp offers a cautionary tale about exactly this kind of approach.

As daisyNews explored recently, prior authorization in all its forms is fraught enough without the specter of powerful, essentially unregulated and error-prone technology. AI’s ability to convincingly pantomime decision-making may be an impressive illusion, but the consequences for people’s health are all too real.

daisyAuth submits compliant Requests for Authorization for California workers’ comp treatment in seconds. Click to learn more.

LEARN MORE: DAISYAUTH