“The autonomous revenue cycle.” That is the phrase healthcare’s billing vendors are now using out loud, and it should stop you cold. The thesis, spoiled up front: agentic AI in healthcare has moved fastest not into diagnosis, where the public imagined it would, but into the money layer — the revenue cycle that decides who gets paid, whose claim is denied, and whose care is authorized. And the machinery deciding is increasingly acting on its own.

From demo to deployment

According to TechTarget, agentic AI in the revenue cycle has crossed from demonstration into real deployment. Waystar, a major revenue-cycle-management vendor, introduced agentic AI to its platform in an announcement dated January 2026, describing a push toward what it calls the “autonomous revenue cycle.” The company says the technology has helped providers prevent more than $15 billion in denied claims, with clients reporting a 90% reduction in time spent on appeals and documentation.

Those are the vendor’s own figures, and they describe genuine pain being solved. Appeals and documentation are grinding, expensive administrative work. Major revenue-cycle-management vendors now run prior authorizations, denials, appeals and coding autonomously — tasks that until recently required rooms full of trained staff. On efficiency grounds, the case nearly writes itself.

The risk cuts both ways

The trouble is that the same automation that clears backlogs can also deny care at scale, and one court case now hanging over the industry shows why that matters. In Lokken v. UnitedHealth Group, a March 2026 federal court order compelled broad discovery into UnitedHealth’s “nH Predict” algorithm. Plaintiffs allege the algorithm drove post-acute-care denials — and, strikingly, that those denials carried a roughly 90% appeal-reversal rate.

Read that number carefully. If an appeals process overturns roughly nine of every ten denials it reviews, it implies the original denials were, in the vast majority of reviewed cases, wrong. An automated system that produces decisions overturned that often is not merely inefficient; it is a system whose errors fall directly on patients, most of whom never file an appeal at all. The court’s willingness to force discovery into the algorithm itself signals that “the model decided” is not going to work as a defense.

Where the highest risk concentrates

Not all agentic tasks carry equal danger. TechTarget’s reporting flags the highest-risk categories as those that directly affect a patient’s access to care:

• Prior-authorization denials — deciding whether a treatment is approved before it happens.
• Clinical documentation and triage — shaping the record and the order in which patients are seen.
• Discharge planning — determining when and how a patient leaves care.
• Inpatient monitoring — watching patients whose condition can change fast.

What these share is that a wrong call is not a billing inconvenience; it is a gate on treatment. When an agent handles a coding edge case and errs, a claim is reprocessed. When an agent denies a prior authorization and errs, a person may not get care in time. The stakes are asymmetric, and they land on the patient.

Efficiency versus accountability

This is the real fault line, and it is not diagnosis versus treatment. It is efficiency versus accountability. Automation is arriving fastest precisely where the incentives are financial — in billing, denials and appeals — because that is where the money is and where the return on removing human labor is most immediate. But it is also the layer where the patient has the least visibility and the least power, and where an error is hardest to catch before it does harm.

Waystar’s $15 billion figure and the Lokken case are two readings of the same trend. One says autonomy prevents wasteful denials and frees clinicians from paperwork. The other asks what happens when the autonomous system is the thing generating wrongful denials in the first place — and who is answerable when it does. The 90% appeal-reversal allegation and the 90% reduction in appeal time are, unsettlingly, describing two sides of the same automated machine.

The unresolved question is liability. When an agent denies a claim and a patient is harmed, who is responsible — the vendor that built the model, the payer that deployed it, or the clinician who deferred to it? Healthcare has answered that question for human decisions for a century. It has not yet answered it for autonomous ones, and the March 2026 discovery order suggests the courts, not the vendors, may be the ones to force the answer.

Written for Red Robot with AI assistance and human editing. Based on reporting by TechTarget, Waystar and court filings.