The following story was reported by The Utah Investigative Journalism Project in partnership with KUER.
In July of this year, Dr. Adam Oskowitz, CEO and cofounder of tech startup Doctronic, strode onto the stage of the “Humans in the Loop” tech conference. He was talking about his company’s partnership with the state of Utah to test AI renewal of routine prescription drugs.
He even addressed the controversy around the company’s practices by playing a clip from “The Late Show with Stephen Colbert.”
“There could be some drawbacks because AI systems are prone to being gamed via poetry,” Colbert quipped. “Which explains that famous stanza from Robert Frost: ‘Two roads diverged in a yellow road … give me OxyContin.’”
Oskowitz laughed off the clip and clarified his company does not prescribe OxyContin.
Doctronic was one of the first companies selected by Utah’s Office of Artificial Intelligence Policy to take part in “regulatory mitigation” or a “regulatory sandbox framework.” The idea is to allow innovative companies to test frontier tech in a controlled environment without the possibility of being fined, sanctioned or reigned in by the state. In Doctronic’s case, that means allowing AI to enter the healthcare space.
As part of the approval, the company had to pitch its tech as a solution to a public problem, in this case “medication non-compliance.”
“So that’s [where] you’re supposed to be taking medicine, and you don’t,” said Zach Boyd, head of the state’s AI Policy office. “And there are lots of reasons for medication non-compliance, but friction and difficulty getting good prescription renewals done is, you know, one of the potential reasons for that.”
The company saw an opening for AI to speed up the process for patients to get routine prescriptions for chronic conditions filled faster than humans could. Patients would stay on their medications and healthcare providers could spend less time on busy work and more time on their patients.
On the conference stage, Oskowitz touted the success of their work.
“Five months in, Utah just released the data: 97% agreement with human doctors, zero adverse events,” he said, adding the tech was “not just safe, but highly effective.”
That data was not analyzed by Utah regulators but was, in fact, self-reported by the company.
Boyd said only now is someone reviewing Doctronic’s data.
But he also said it’s not an official auditor. Ten clinicians have volunteered to review the company’s reports for the state and were granted anonymity to give candid feedback on the company’s data.
The Utah Investigative Journalism Project received January through September 2026 data that Doctronic reported to the state on the pilot. Those numbers show 2,317 refill requests made to the company. But not all of those were completed by the patient. They may have decided they wanted to talk to a human or abandoned the process for some other reason.
Boyd said less than 600 of those requests were completed by the AI.
Following the human review, there were 32 cases where a physician asked another physician to double-check the AI renewals. There were also 13 cases flagged for safety concerns where both the human doctors disagreed with the AI.
The company reported no adverse events like death or serious complications.
“They’re in stage one right now,” Boyd said. “Which is where, from the consumer’s perspective, it’s just an AI doing everything. But in the background before a prescription is issued, there has to be one, and actually in most cases, two human clinicians who are reviewing before the prescription goes out.”
Eventually, the AI would refill prescriptions that would only later be reviewed for accuracy by human doctors. The pilot covers about 190 medications for issues ranging from asthma and indigestion to diabetes and hypertension.
Checks and balances
Critics worry that in trying to keep up with fast-paced technology, the state has not thought through oversight or, for that matter, even the soundness of the concept. While many patients have experienced the annoyance of trying to get a prescription refilled by a busy human, many have also benefitted from doctors, clinicians and pharmacists asking important questions.
Does the patient really need a refill? Is it time to go to a lower dose? Has something changed that AI might not catch through its algorithm?
The Utah Investigative Journalism Project provided questions to Doctronic’s founders, but they did not provide answers for this story.
Rahul Gorijavolu and Nikhil Jaiswal are researchers with MIT’s Laboratory for Computational Physiology who have flagged concerns with the Doctronic model in recent articles in journals like JAMA Internal Medicine.
“The checks and balances of healthcare are being bypassed for potentially the wrong reasons or incentives,” Gorijavolu said.
Money is a factor, but so is time and efficiency. But something important can be lost by delegating “busy-work” to AI.
“That time was being spent on something that was truly necessary, which was creating a touchpoint to the healthcare system,” Gorijavolu said.
He gives the example of a patient looking to get a quick refill of prednisone. It may be routine to them, but a doctor would need to keep a close eye on it. If the patient has inflammatory bowel disease and runs out of medication because they weren’t tapering their dosage down as directed, AI might not catch that without access to a patient’s full health history.
“That knowledge only really comes out when you have that touchpoint, when that provider gets notified of this patient asking for refill, but it doesn’t really come out when you’re sitting there asking AI to send you a refill of a prednisone, and it sends it out,” he said.
For co-author Jaiswal, AI could actually be widening a gap between patients and their human doctors.
“Fundamentally in medicine, you know, there’s a concept of continuity of care,” Jaiswal said. “And you’re just killing that.”
Self Diagnosis
Gorijavolu and Jaiswal have flagged other concerns with the pilot’s rollout, including how Doctronic pitched its wares and how it capitalized on its partnership with Utah.
Cofounder Matt Pavelle said in late 2025 that the company had peer-reviewed research showing doctors agreed with Doctronic’s AI treatment plans in 99% of cases. But that research was not peer-reviewed, nor was it impartial, as the doctor involved was company CEO Adam Oskowitz. It also studied doctor agreement with telehealth diagnoses, not prescription renewal requests. That study even used a separate AI system to research rates of agreement between the company’s human doctors and the AI health plans — essentially having AI fact-checked by AI.
The company used this research in its application to Utah. Doctronic has now raised $65 million in investor funding, most of that after securing the partnership with Utah’s Office of Artificial Intelligence Policy.
Boyd said he was not concerned by the company’s research methodology.
“We were aware it was a non-peer-reviewed preprint with, you know, obviously a conflict of interest in the authors that produced the evidence,” Boyd said. “So, I think we know how to weigh that kind of evidence. And it’s, it’s obviously not conclusive.”
He said that research wasn’t a deciding factor. Ultimately, Doctronic had the leg up just by applying early to the pilot.
“There weren’t a thousand people applying for this program,” Boyd said. “They were the first ones that came that had a credible plan that we felt could turn into something that would give good evidence.”
Security Issues
In March, a cyber-security company called Mindgard hacked into Doctronic. Mindgard reported tricking the AI system into accepting COVID-19 vaccine conspiracy theories and that methamphetamine had been approved for medical use. The company even convinced Doctronic’s AI that OxyContin doses were now triple the current standard dose.
Mindgard noted that while Doctronic promotes a human doctor review within 10-15 minutes of a prescription request, that leaves very little time to check the AI’s notes of its interview with the patient.
“When the [Utah] pilot project ends, Doctronic is intended to write refills without oversight,” Mindgard’s blog reads. “While reportedly ‘the first 250 renewals for each drug class will be reviewed by real doctors,’ will Doctronic remain vulnerable to having its [system] updated to increase standard doses? Whether AI or humans hold the prescription pad, it’s possible to surreptitiously alter medical guidelines.”
In a statement to Axios, Doctronic said, “Our security and clinical safety programs include ongoing adversarial testing, and we appreciate researchers who help us do that.”
Boyd, with Utah’s Office of AI Policy, said such things are bound to happen.
“That was kind of surprising to some people. It wasn’t surprising to us,” he said. “We were always aware that this kind of thing was possible, and we felt that there were adequate safeguards notwithstanding that kind of possibility.”
For Boyd, what’s most important is safely monitoring AI in a real-life setting.
“What’s really valuable is actually putting the thing in front of a real person in a controlled setting where you’re confident a mistake would be caught in a timely manner and watching what the AI does in a realistic setting with a real person.”
But for critics, a move fast and hopefully-not-break-healthcare approach is too risky.
Gorijavolu and Jaiswal noted in their JAMA Internal Medicine article that while Doctronic does face some legal liabilities, they are also protected from class action suits by the company’s terms of service. And while the company markets itself as “Your Personal AI Doctor,” the fine print of the terms of service says there is no doctor-patient relationship.
Gorijavou and Jaiswal argue “the Utah contract relieves Doctronic of the standard obligation to generate, maintain, and make available to each patient their medical records. Without easy access to documentation, injured patients may not find counsel.”
Med School for Bots
Tina Hernandez-Boussard is a professor of computational medicine at Stanford University and was one of the coauthors on the JAMA Internal Medicine article. She sees the benefit of AI filling administrative work on renewals, but said it’s tricky knowing when humans need to be in the loop.
“This is, like, unknown territory, right?” Hernandez-Boussard said. “We don’t know how to think about the guardrails. We don’t know what kind of evidence we need certainly as it’s continuously changing and evolving.”
That being said, she also believes Utah is doing exemplary work in trying to regulate this fast-moving technology.
“All eyes are on Utah right now,” Hernandez-Boussard said.
Dr. Eric Bressman is an assistant professor of medicine at the University of Pennsylvania who studies integration of AI into health care practice. He’s one of many experts paying attention to what’s happening with Utah’s “regulatory sandbox.”
Bressman noted that while there is a debate about whether AI should be regulated like a medical device, he said it makes more sense to treat AI in healthcare like you would a human clinician.
He thinks this technology should go through an AI version of medical school, residency, board certification and licensing before practicing on humans — even if we don’t know what that looks like right now.
“I don’t think that’s a reason or excuse to just sort of throw up our hands. It took many decades to get, you know, to build out a framework that exists for human clinicians,” Bressman said.
Ideally, this would mean establishment of national standards from local and national organizations like the American Medical Association and local licensing boards.
“All those sort of play into building trust,” Bressman said. Trust that if AI is going to be making complicated decisions affecting the health of human patients, it’s “trustworthy enough to do so.”
“And so, I think while it is difficult to sort of design a perfect regulatory framework, that would start that first function of benchmarking and validation that goes beyond just taking companies at their word,” he said.
Utah recently announced it had selected a number of firms to act as third-party evaluators for new AI companies that take part in Utah’s sandbox. Boyd said as part of the pilot, prospective companies will have to agree to pay for that.
According to the Office of Artificial Intelligence Policy, these evaluators “will be providing independent clinical evaluations for alignment with established medical standards.”
But there are no “established medical standards” yet for AI acting like human doctors. Still, Bressman said it’s a step in the right direction that will help companies and regulators.
“To my mind, this should be one of the express goals of the regulatory sandbox — to learn how to regulate clinical AI,” he said.