AI is good at catching drug theft at hospitals, but only when humans do their part

STAT News – Wednesday, August 25, 2026
By Alexandra Byrne

In late September 2024, patients noticed a nurse at Adventist Health in Bakersfield, Calif., was acting strangely.

The nurse was walking barefoot through the intensive care unit, talking to herself and acting abrasively toward others. Family members of patients noticed the nurse sloppily taking out IV needles but were afraid to confront her. “I know she was on something because you can tell,” one family member told federal investigators after the incident.

Another patient in the post-anesthesiology recovery unit treated by the same nurse was in excruciating pain under her care. “I was white knuckling it,” he said. “You’re not dying,” the nurse replied when the patient expressed his pain. He thought he was being given fentanyl and morphine, but the IV drip didn’t seem to help.

That’s because the nurse, hired through a travel nursing agency several weeks before the incident, was taking the medications from a secured cabinet in the wing and using them herself, while documenting that they had been administered to patients, according to investigators from the Centers for Medicare and Medicaid Services tasked with investigating a complaint about the incident in November 2024.

The incident didn’t happen out of the blue. Hospital managers ignored alerts from a machine learning software that tracks when staff might be stealing drugs, auditors found.

Hospitals are major repositories of addictive substances that can also be crucial medical treatments. Occasionally, employees steal those drugs and are under the influence while caring for patients. While new technologies are helping to identify theft and connect individuals with substance abuse resources, incidents like the one at Adventist show artificial intelligence isn’t foolproof.

Adventist Bakersfield, a 253-bed acute care hospital in Southern California, uses a system called ControlCheck that captures discrepancies in what is taken and returned to locked cabinets, missing parts of a shipment, or unused drugs that aren’t disposed of under supervision. ControlCheck uses those data points to find patterns in staff behavior and assign a risk score to individual staffers.

The AI-based system, developed by Virginia-based company Bluesight, is used in over 2,000 hospitals nationwide. It’s one of several AI systems that hospitals have widely adopted over the past few years to track the use of federally-regulated drugs and flag staff members exhibiting risky behaviors. In many cases, these systems can catch patterns and prompt interventions. But if supervising staff aren’t trained on ControlCheck, the health systems’ drug-diversion programs can be rendered ineffective, and expose patients to risk.

When federal investigators met with senior pharmacists, risk managers, and unit directors at Adventist Bakersfield, the hospital’s senior pharmacist for controlled-substance diversion admitted he did not know how to read the risk scores ControlCheck generated. Nurse supervisors “had received minimal training on Bluesight at best,” they found.

Drug diversion happens frequently at hospitals. “There is diversion going on everywhere,” said Chris Fortier, who previously served as the chief pharmacy officer at Massachusetts General Hospital in Boston. “Whether it’s large, whether it’s small, it’s happening. And it’s very, very difficult to identify.”

In 2021, Fortier conducted what is believed to be the first and only nationwide survey on drug diversion programs. The survey, done on behalf of the American Society of Health-System Pharmacists, found that about two-thirds of the nearly 200 respondents had a formalized drug diversion committee, about half had implemented surveillance software, and a quarter had more than 10 formal drug diversion investigations each year.

Bluesight says its system catches two confirmed cases of drug diversion every day, on average. Over 45,000 cases caught by the software have been investigated since it launched in 2017.

Nurses, pharmacy technicians, and anesthesiologists may dispense thousands of medications each day at a midsize hospital to treat patients. The Drug Enforcement Administration requires hospitals to document these transactions and have policies to prevent and respond to drug diversion.

If a staff member is reported by colleagues or is suspected of diversion, the hospital is supposed to review electronic records, camera footage, and access logs, then conduct interviews. If the individual is found to have diverted, they are usually referred to rehab or treatment.

That process used to be mostly manual, said Michael Fitzsimons, a cardiac anesthesiologist at Mass General who was tapped nearly 25 years ago to start a drug testing program at the institution. Studies have shown anesthesiologists are prone to relapse and death from substance use disorders, and Fitzsimons wanted to help his colleagues get the help they needed.

“It wasn’t just drug testing,” Fitzsimons said. “It was also enhancing education and evolving our systems to include automated medication dispensing systems and medical record surveillance.” The goal was to look for patterns of suspicious or off-protocol obtaining of controlled substances, he explained.

Machine learning and AI can capture diversion at much higher rates than traditional segmented paper methods, Fitzsimons said — and up to 160 days faster, according to a study by the firm Invistics, which developed its own system, now called Sentri7 Drug Diversion and owned by Dutch multinational Wolters Kluwer known for popular clinical-decision support tool UpToDate.

Beyond the Invistics study, few academics have researched the advantages and disadvantages of these systems, but Fitzsimons said he’s “optimistic that AI is going to expand our capability to surveil medical transactions at a much higher level than we do now.”

Even as AI-based systems hold promise, experts cautioned that the technology is only as effective as the humans behind drug diversion programs are attentive. If nurse managers and pharmacists don’t correctly document transactions, or if they ignore alerts about risky behavior, drug diversion might go unnoticed. And unless staff get continuous education, “you’re probably not going to have as much of an impact as you would really like to have,” Fitzsimons said.

At a different Adventist hospital in Glendale, Calif., a registered nurse was found unresponsive in the ICU locker room on July 26, 2025, with narcotics and a syringe. The nurse, who was responsible for two patients’ care, tested positive for fentanyl and opioids, federal investigators reported the next month.

The nurse had dispensed a 100 mL IV bag of fentanyl, but only 75 mL was administered to a patient. The rest was unaccounted for.

A review of the nurse’s ControlCheck history revealed a “highly unusual number of variances” and a history of long stretches of time between taking controlled drugs out of storage and administering it to patients or throwing it away with witnesses. Despite the surveillance system flagging the nurse’s behavior, the hospital’s staff did nothing to act upon the report.

The software had been in use at the facility since 2018, but “the hospital did not have a policy and procedure regarding the process of using Bluesight or ControlCheck,” federal investigators found. Staff did not manually review the risk scores the system produced, and would generally disregard alerts related to IV drips because they felt the system couldn’t accurately monitor them.

“ControlCheck is highly accurate at surfacing early indicators of potential diversion, but it can’t act on them for the hospital,” said Kevin MacDonald, the co-founder and CEO of Bluesight. “That takes people — whether centralized or decentralized — reviewing and following through on the highest risk signals.”

Jessica Hodges Schaef, who leads implementation at Bluesight, explained that the company conducts live training with new clients and makes available a suite of training videos that can be referenced at any time. If hospitals have a question about the technology, they always have a point person at Bluesight, Schaef said.

Bluesight says it is working to expand ControlCheck’s ability to monitor infusions, which is an add-on feature that hospitals must opt-in to. The company has recently started offering a generative AI component that allows users to query the system with plain-language questions to make it easier to use.

Adventist declined an interview and did not address a detailed list of questions about failures identified by federal investigators. Adventist also declined to share its plan of correction, a document that typically details how a hospital plans to respond to federal investigators’ concerns.

“Adventist Health maintains a comprehensive program to prevent, detect, and respond to drug diversion,” spokesperson Japhet De Oliveira said in a statement. “This includes established policies and a combination of system-level and individual-level controls.” De Oliveira said Bluesight “is one of the multiple tools and processes used to support monitoring and compliance” and “represents one component within this broader oversight framework.”

The problem extends beyond Adventist. At Virginia Mason Medical Center in Seattle, a nurse took 62 vials of fentanyl and 21 cartridges of hydromorphone, an opioid, between November 2023 and July 2024, according to a federal inspection report. The hospital eventually revoked the nurse’s access to the drug dispensing cabinet, but unit managers ignored verbal concerns from other staff members, discrepancies in drug stock, and alerts from ControlCheck.

Staff at Virginia Mason “did not recall receiving any recent training about diversion surveillance or how to run surveillance reports using ControlCheck,” investigators noted. Virginia Mason declined to comment.

While ControlCheck is widely adopted among health systems, it’s not the only surveillance tool experiencing implementation challenges.

Like ControlCheck, Sentri7 Drug Diversion also relies on hospitals training staff and developing policies for how to use it effectively. Federal investigators found that Penn Medicine Princeton Medical Center failed to develop those policies for Sentri7 after 150 syringes of ketamine went missing in June 2024. (The hospital did not respond to a request for comment.) In 2025, the Sentri7 system failed to flag missing drugs at Erlanger Baroness, a hospital in Tennessee, according to a consent order released by the state’s Board of Nursing, first reported by KFF Health News.

Kristy Drollinger, who formerly implemented Sentri7 in hospitals for Wolters Kluwer until leaving the company recently, said it’s difficult to separate signals from the noise without AI, but the system does not work without people reviewing the results.

“It’s not magic,” she said. “It’s kind of magic, but it’s not all-the-way magic.”

The ultimate goal of drug diversion programs should be to identify and help staff members who may be struggling with addiction, said Mass General’s Fortier. It’s a monumental task, he said, especially in hospitals where resources are often stretched thin.

Even though they’re not perfect, machine learning systems can carry some of that weight, Fortier said. “You cannot do this manually,” he said. “You’re literally looking for a needle in a haystack every single day.”

Access this article at its original source.