We're back ! | The 'Robot Referee': How AI Could Approve Your Treatment in Seconds Before a Human Ever Reads Your Chart

Picture this. You're 68, your knee has been failing for a year, and your surgeon finally recommends a procedure. Their office submits the request. Somewhere, a piece of software scans the paperwork, compares it against a rulebook, and makes a call. It might approve you in seconds. It might also leave you waiting for weeks, or say no.

Meanwhile, back at your appointment, a different AI was listening. It turned your conversation into clinical notes and quietly suggested a few extra diagnoses for the billing record. Diagnoses you may never hear about.

Two algorithms, one on each side of your medical bill, are negotiating over your care. Neither has ever met you.

This isn't a thought experiment. It's this week's healthcare AI news.

Three Headlines, One Story

Three developments that look separate are really one story:

• The machine that says "no." Roughly 1,000 pages of records, pried loose by the Electronic Frontier Foundation through a Freedom of Information Act lawsuit, show how Medicare's new AI-assisted prior authorization experiment has actually performed since January: high denial rates, missed deadlines, and technical failures.
• The machine that says "more." An analysis from the Blue Cross Blue Shield Association (BCBSA), whose independent member companies cover more than 100 million Americans, found that AI scribes and automated coding tools helped hospitals bill nearly $1 billion more over two years, while treatment patterns barely changed.
• The humans who said "enough." About 150 physicians at two Minnesota hospitals held a four-day strike, reportedly the first by inpatient doctors at a private U.S. hospital. One of their central demands was a real voice in how AI is used in diagnosis and billing. On September 24 they reached a tentative deal.

Put them side by side and a pattern emerges. AI is no longer only a clinical tool that reads scans or predicts sepsis. It has become a financial weapon, and both sides of the healthcare economy are arming up.

Round One: The Machine That Says "No"

Prior authorization is the process where a doctor must get an insurer's permission before providing certain treatments. It exists to curb unnecessary or harmful procedures, but it has become one of the most hated parts of modern medicine.

In January 2026, Medicare launched a six-year pilot called WISeR (Wasteful and Inappropriate Service Reduction) in six states. For the first time, traditional Medicare began using private technology vendors, armed with AI, to review prior authorization requests for a list of services regulators consider prone to overuse.

The records obtained this month paint a troubling picture:

• Thousands of denials, fast. Two vendors alone denied 5,944 requests in the program's first three months.
• More "no" than "yes." One contractor made more than 6,000 decisions through the end of March and denied 53% of them. It was placed on a corrective action plan, in part for missing the program's 72-hour turnaround target.
• Waiting in limbo. One request went unanswered for 83 days. Providers described patients calling in tears, in pain, with surgeries postponed.
• Built-in incentives. Vendors are paid in part based on requests they deny (denials reversed on appeal don't count), while penalties for poor timeliness or accuracy reduce payments by only about 5–10%.
• Expansion ambitions. Planning documents discussed adding MRI scans, cancer treatment, air ambulance transport and certain medications to the program.

Why does this matter so much? Because most people never fight back. In 2024, Medicare Advantage insurers made nearly 53 million prior authorization decisions and fully or partially denied about 4.1 million of them. Only 11.5% of those denials were appealed. Yet when patients did appeal, 80.7% of the denials were overturned.

When someone challenges a denial, it is reversed four times out of five. An automated system that denies quickly and at scale doesn't need to be right to save money. It only needs most people to give up.

Round Two: The Machine That Says "More"

Now flip to the other side of the table.

Hospitals and clinics have embraced "ambient" AI scribes: tools that listen to a visit and draft the clinical note, so doctors can look at patients instead of keyboards. It's one of medicine's most popular AI deployments, because documentation is a leading driver of physician burnout.

But these tools do more than transcribe. They also surface every condition that can be documented and billed. And in healthcare billing, detail equals dollars.

The BCBSA analysis, released September 24, found:

• AI-assisted documentation and coding added $942 million in inpatient costs across 2024 and 2025 compared with 2023 levels.
• About $653 million came from "secondary conditions" such as anemia or low sodium being added to records, raising payments by nearly $12,000 per affected case.
• Among bowel surgery patients, recorded secondary intestinal blockages rose 55% and acid reflux diagnoses rose 33%.
• Yet the treatment you'd expect for those conditions, such as blood transfusions for anemia, stayed flat.
• Outpatient care added at least another $1.67 billion, bringing the estimated total to roughly $2.3 billion.

BCBSA's verdict was blunt: AI "is identifying more billable conditions, not sicker patients."

Hospitals see it differently. For decades, they argue, rushed clinicians under-documented how complex their patients really were, and insurers under-paid as a result. AI simply captures the full picture. There's truth on both sides. Even industry observers acknowledge that insurers and providers privately agree AI scribes are increasing coding intensity; the fight is over whether that is accuracy or inflation. One Michigan health system has openly said its AI documentation tools add about $1 million in revenue every month.

For the everyday person, this isn't abstract. Higher billing flows into higher premiums, deductibles and employer health costs. And a diagnosis added to your record for billing purposes doesn't disappear. It becomes part of your medical history, and can surface years later in life or disability insurance applications, where medical records are still fair game in many places.

The Arms Race Nobody Voted For

Here is the uncomfortable dynamic. Providers deploy AI to document more, so insurers deploy AI to scrutinize and deny more, so providers deploy AI to write more persuasive appeals, and so on. Each side points to the other's algorithm to justify its own.

The costs of this escalation land on two groups who never agreed to it: clinicians and patients.

The American Medical Association's latest prior authorization survey, released in May 2026, found that physicians spend an average of 13 hours per week on prior authorization paperwork. 93% say it delays patient care, 94% say it harms clinical outcomes, and 82% say it can lead patients to abandon treatment altogether. Only a third believe the industry's latest reform pledge will make a meaningful difference.

When Doctors Walk Out Over Algorithms

That's the backdrop for what happened in Coon Rapids and Fridley, Minnesota. Hospitalists, OB-GYNs, psychiatrists and critical care doctors at Mercy Hospital and its Unity campus, after three years of trying to secure a first union contract, went on strike for four days.

Pay and sick leave were on the table. But the doctors also raised something new: AI was already suggesting billing codes and diagnoses in their electronic records, and they feared it would soon start nudging treatment decisions, pressuring them to follow the machine rather than their own judgment.

The tentative three-year agreement reached on September 24 doesn't ban AI. Instead, it protects physicians' rights to advocate for their patients and for patient safety amid powerful economic forces, explicitly including the emergence of AI. The deal still needs ratification, but the precedent is already set.

AI governance used to live in ethics committees and vendor contracts. Now it is being negotiated at the bargaining table, much as Hollywood writers fought for AI protections in 2023. Expect other clinician groups to follow.

The Rulebook Is Being Written in Real Time

Regulators worldwide are racing to catch up, and the patchwork is uneven.

United States. A federal rule taking effect in 2026 requires Medicare Advantage and Medicaid plans to answer urgent prior authorization requests within 72 hours and standard ones within seven days, to give a specific reason for every denial, and to publicly report their approval and denial rates. By 2027, they must also support standardized electronic prior authorization. California already requires that a licensed clinician, not an algorithm, make the final call on medical-necessity denials, and other states are following.

Privacy law: HIPAA and PIPEDA. HIPAA governs how your health information is protected when it's shared with the vendors running these systems. But it says nothing about whether an algorithm's decision is fair or explainable. Canada has a similar gap. PIPEDA, the federal private-sector privacy law, requires meaningful consent and gives you the right to access your personal information, but it was written long before algorithmic decision-making and does not spell out a specific right to contest an automated decision. Quebec's Law 25 goes further: organizations must tell you when a decision about you was made exclusively by automated processing, explain the main factors on request, and let you submit your observations to a person who can review the decision. And while Canada's public system doesn't use U.S.-style prior authorization for most services, private drug plans and provincial drug programs rely heavily on "special authorization" forms, exactly the kind of workflow that is ripe for automation.

Europe. The GDPR already gives people a right not to be subject to decisions based solely on automated processing when those decisions have significant effects. The EU AI Act goes further, classifying AI used to determine eligibility for essential public services, including healthcare, and AI used to price or assess risk in health and life insurance, as "high-risk." However, under the Digital Omnibus package that entered into force in July, the obligations for these systems have been pushed back to December 2027.

Middle East and Asia. In markets with mandatory health insurance, such as Dubai and Abu Dhabi, fast-growing claims volumes make automated triage attractive, and across Asia, insurers are using machine learning to flag suspicious claims. The same question will arrive everywhere: when the software says no, who is accountable?

The Case for the Robot Referee

It would be easy to frame all of this as "AI bad." That would be a mistake.

The old, human-only system was already failing: fax machines, phone trees and weeks of waiting. Used well, AI could make prior authorization almost invisible. A routine, clearly appropriate request could be approved in seconds, while human reviewers focus their time on genuinely complex cases. In mid-2025, dozens of major U.S. insurers pledged to answer 80% of electronic prior authorization requests in real time by 2027, and to ensure that a qualified clinician reviews any request that isn't approved. That future is only achievable with automation.

The same is true on the documentation side. AI scribes give doctors back hours every week, and more complete records can catch real problems. An anemia flagged by AI that is genuinely present deserves follow-up, not suspicion.

The problem isn't the technology. It's the incentives it's plugged into. An AI paid to deny will deny. An AI tuned to maximize revenue will find revenue. The most important design decision isn't the model; it's what the model is rewarded for.

What This Means for You: A Patient's Playbook

You can't opt out of the arms race, but you don't have to be passive:

1. Always ask for the specific reason. If a treatment is denied, request the exact reason in writing, and ask whether an automated tool was involved.
2. Appeal. Seriously. With roughly 8 in 10 appealed Medicare Advantage denials overturned, an appeal is often the single most powerful thing you can do. Ask your doctor's office to help.
3. Read your visit notes. Most patient portals let you see your clinical notes. If you see a diagnosis you don't recognize, ask about it. You have the right to request corrections.
4. Ask about AI scribes. It's reasonable to ask whether your visit is being recorded or transcribed by AI, and how that data is stored and used.
5. Know your regional rights. In Quebec, Law 25 lets you ask how an automated decision was made. In Europe, GDPR gives you a right to human review of significant automated decisions. In the U.S., check whether your state requires clinician review of denials.

What Professionals Should Watch

For health leaders, clinicians and technologists, this week offers a governance checklist:

• Make human review real, not ceremonial. A clinician who approves 200 AI recommendations an hour isn't reviewing; they're rubber-stamping.
• Audit outcomes, not just accuracy. Track denial rates, appeal overturn rates and turnaround times by vendor and by service line.
• Measure the "diagnosis-treatment gap." If coded conditions rise but treatments don't, that's a signal worth investigating, from either side of the table.
• Fix the incentives. Contracts that pay vendors per denial, or tools marketed purely on revenue uplift, will eventually become front-page news.
• Bring clinicians in early. Minnesota just showed what happens when frontline staff feel AI is being done to them rather than with them.

The Road Ahead

The next phase is already visible: AI-to-AI negotiation. As standardized electronic prior authorization becomes mandatory, a provider's system will assemble the clinical evidence, a payer's system will evaluate it, and many decisions will happen machine-to-machine in seconds.

That could be wonderful or terrible, depending on three things: transparency (can both sides, and the patient, see why a decision was made?), accountability (is a named human responsible for every denial?), and alignment (are the algorithms rewarded for good care, or for money?).

Wrapping Up

This week, we learned that the most powerful AI in healthcare may not be the one reading your MRI. It may be the one deciding whether you get the MRI at all, and the one deciding how much it costs.

The Robot Referee is already on the field. It can make healthcare faster, fairer and less exhausting for everyone, or it can turn medicine into an automated tug-of-war with patients as the rope. Which future we get depends on whether we demand that both algorithms show their work.

The next time you get a denial letter, remember: there's a good chance a machine wrote the first draft. And if Medicare Advantage is any guide, there's a good chance it won't survive an appeal.