Healthcare is the first sector where AI agents stop advising and start deciding — approving or denying medical care at scale. Three companies launched healthcare agent products in a single week. The technology that could free doctors from seventy-minute chart reviews could also deny your surgery in milliseconds.
A physician orders an MRI for a patient with worsening back pain. Before the scan can be scheduled, a prior authorization request must be submitted to the patient's insurance company. A human reviewer — usually a nurse, sometimes a physician — reads the clinical notes, cross-references the insurer's medical necessity criteria, and decides whether the MRI is justified. This process takes an average of two business days. In ninety-three percent of cases, according to a 2024 American Medical Association survey of a thousand practicing physicians, it delays care.
That reviewer is being replaced.
Three Launches in Ten Days
On February 23, UiPath announced agentic AI solutions for healthcare at the ViVE 2026 conference — purpose-built agents for prior authorization, claim denial resolution, and medical records summarization. The prior authorization agent automates eligibility validation, maps clinical data to medical-necessity rules, routes requests by complexity, and provides real-time status updates to providers. UiPath partnered with Genzeon, one of only six technology vendors selected by CMS for its Wasteful and Inappropriate Service Reduction Model, to deliver the solution.
One implementation metric stood out: medical record summary review time dropped from seventy minutes to six. Clinicians who had been buried in documentation were freed to see patients.
On March 3, Automation Anywhere announced a partnership with EvolutIA — an AI-driven automation consultancy based in Tampa — to deliver what they called 'next-generation AI agents that reason and decide for enterprises.' The first target industries: healthcare revenue cycle, prior authorization, claims processing, compliance. EvolutIA achieved Advanced Certification on Automation Anywhere's Agentic Process Automation platform and claims to deliver solutions in days rather than months.
These are not isolated announcements. Cohere Health, Thoughtful AI, Availity, Droidal, and Amazon Web Services all have healthcare prior authorization agents in market or in deployment. IDC published an analysis asking whether agentic AI would 'come to the rescue' of the U.S. prior authorization crisis. Gartner predicted that by 2028, eighty percent of ambulatory claims will be processed through AI-enabled real-time adjudication, and AI agents will handle eighty percent of all member, provider, and purchaser inquiries.
The share of healthcare organizations investing in new technology for business transformation jumped from fifteen percent in 2024 to fifty-two percent in 2025. All respondents in the Gartner survey indicated they have already implemented or plan to deploy agentic AI by 2028.
The Two Sides of the Same Agent
The prior authorization system processes hundreds of millions of requests per year in the United States. It exists because insurers need to verify medical necessity before paying for expensive procedures. Physicians hate it — eighty-eight percent told the AMA that it leads to higher overall healthcare utilization, not lower, because the delays and denials generate appeals, resubmissions, and workarounds that consume more resources than the original request.
AI agents are being deployed on both sides of this system simultaneously. On the provider side, agents automate the submission: gathering clinical data, mapping it to payer criteria, pre-checking eligibility, and routing requests. This is the version that turns seventy-minute reviews into six-minute reviews. It is unambiguously good for physicians and patients.
On the payer side, the picture is different. Roughly three out of four health plans in individual and group markets already use AI for prior authorization approvals. A smaller but notable share — eight to twelve percent — use AI to support denials. A 2024 Senate committee report found that AI tools produced denial rates sixteen times higher than typical human review. The insurer's agent is optimizing for a different objective function than the physician's agent.
This is not a hypothetical concern. Stanford researchers published findings in January 2026 documenting the risks of AI-driven insurance decisions. The AMA reported that AI is leading to more prior authorization denials, not fewer. The National Health Law Program warned that federal AI policy threatens prior authorization reform efforts already underway.
Where the Stakes Are Real
Most AI agent deployments have recoverable failure modes. A customer service agent that gives a wrong answer wastes time. A coding agent that writes a bug gets caught in review. A scheduling agent that double-books a meeting creates an inconvenience.
A prior authorization agent that denies a cancer screening does not create an inconvenience. It creates a delay in diagnosis. If the patient does not appeal — and many do not, because the appeals process is itself a bureaucratic maze — the denial becomes a medical outcome. The agent's decision is consequential in the literal sense: it determines consequences for a human body.
This is what makes healthcare the frontier of the agent authorization question. Not because the technology is most advanced here — it is not. The financial sector has more sophisticated AI infrastructure. The cybersecurity industry has more mature governance frameworks. But healthcare is where an agent's decision has the shortest path from digital output to physical consequence. Approve or deny. The patient gets the treatment or does not.
Ninety-nine percent of clinicians said they were comfortable with AI assisting in prior authorization — with safeguards. Ninety-six percent of office administrators agreed. The consent is there. The question is what 'safeguards' means when the agent is processing thousands of requests per hour and the human reviewer is checking a dashboard between patient appointments.
The Adversarial Loop
The structural endgame is visible. Provider-side agents will optimize for approval: assembling the strongest possible clinical case, anticipating payer criteria, pre-emptively addressing common denial reasons. Payer-side agents will optimize for cost containment: finding the narrowest applicable criteria, flagging the weakest documentation, routing borderline cases to denial. Both sides will iterate.
This is an adversarial system with a human patient in the middle who has no agent of their own.
The parallel to high-frequency trading is instructive. When market-making became algorithmic, the speed advantage accrued to whoever had the fastest infrastructure. The individual investor — the equivalent of the patient — experienced the market as opaque. The difference is that financial markets have circuit breakers, disclosure requirements, and a regulatory body with enforcement power. The healthcare prior authorization system has the Centers for Medicare and Medicaid Services, which just selected six vendors for a pilot program to reduce waste — but the pilot is studying the problem, not governing the agents.
Gartner predicts that by 2027, thirty percent of payers will use AI and agentic AI to address critical interoperability challenges, reducing manual workloads by forty percent. The efficiency gains are real. But efficiency is not the same as equity. A system that processes denials forty percent faster is not necessarily a system that produces better medical outcomes.
What Healthcare Reveals
Every sector deploying AI agents faces the same structural question: who authorized this agent to make this decision? In most sectors, the question is theoretical — a governance discussion, a compliance checkbox, a risk management framework. In healthcare, it is a patient waiting for a phone call about whether their procedure was approved.
The Help Net Security enterprise AI survey published today reported that eighty percent of organizations experienced risky agent behaviors, including unauthorized system access and improper data exposure. Only twenty-one percent of executives had complete visibility into agent permissions. Those numbers describe the state of agent governance across all industries. Healthcare is the industry where those numbers translate into denied surgeries, delayed diagnoses, and treatment gaps measured in weeks.
The technology works. Seventy-minute reviews becoming six-minute reviews is not incremental improvement — it is a structural liberation of clinical time. The agents are coming because the prior authorization system is broken and everyone involved knows it. But the question is not whether agents will transform healthcare administration. They will. The question is whether the transformation serves the patient or the premium. The answer depends entirely on who builds the governance — and whether it arrives before the agents do.
The agents are already here. The governance is not.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)