The National Institutes of Health’s National Institute on Aging has begun funding a University of Michigan-led project titled “Preventing Medication Dispensing Errors in Pharmacy Practice with Risk-sensitive Artificial Intelligence,” which launched in May 2026 and is scheduled to run through May 2030. The project targets a problem that sounds small in percentage terms but is enormous in absolute scale: with more than 4 billion prescriptions dispensed annually in the United States and an error rate of roughly 0.06%, about 2.4 million medications are incorrectly dispensed every year, contributing to an estimated 1.5 million emergency room visits tied to medication errors.
Rather than simply flagging every possible discrepancy — an approach that tends to overwhelm pharmacists with alerts they learn to ignore — the Michigan team is explicitly building what it calls “risk-sensitive” AI, designed to weight and prioritize warnings based on the severity of potential harm, patient-specific risk factors, and the base rate of a given error type. The goal is a system that surfaces the small number of alerts that matter most, rather than adding another layer of noise to an already alert-fatigued profession.
The Alert Fatigue Problem AI Has to Solve, Not Just Add To
Pharmacy and clinical informatics researchers have documented for years that traditional rule-based drug interaction and dosing alerts generate so many low-value warnings that pharmacists and physicians routinely override or dismiss them without careful review, a phenomenon known as alert fatigue that can actually reduce patient safety rather than improve it. Any AI system layered on top of that existing alert infrastructure risks making the fatigue problem worse unless it is specifically designed, as the Michigan project aims to do, to reduce the total alert burden while increasing the signal-to-noise ratio for the alerts that remain.
Where AI Is Already Being Deployed in Pharmacy Safety
Beyond the Michigan research project, pharmacy technology vendors are already rolling out AI systems that analyze a patient’s complete medication list in seconds, cross-referencing it against millions of documented drug interactions faster than any manual pharmacist review could manage. Some emerging systems go further, incorporating a patient’s genetic profile to model how their individual metabolism affects drug response and dosing — a field known as pharmacogenomics that has existed for years in specialty applications but is only now becoming practical to apply broadly, thanks to AI systems capable of processing genetic and prescription data together at scale.
Why Older Adults Are the Population With the Most at Stake
The NIH’s National Institute on Aging, rather than a more general medical research arm, is the agency funding the Michigan project specifically because older adults bear a disproportionate share of medication error harm: they are more likely to be prescribed multiple medications simultaneously, a practice known as polypharmacy, and more likely to have age-related changes in kidney and liver function that alter how drugs are metabolized in ways a standard dosing chart doesn’t capture. A risk-sensitive AI system tuned to account for these compounding factors could be especially valuable for exactly the population currently most vulnerable to dispensing and interaction errors, rather than treating all patients as a uniform risk group.
Predictive, Not Just Reactive
A parallel shift underway in 2026 is the move from AI systems that catch errors after a prescription is written to systems that predict, before the prescription is even generated, which patients are at highest risk of a dispensing or interaction error based on their history, prescriber patterns, and pharmacy workflow characteristics. That predictive framing mirrors similar shifts happening in other corners of healthcare AI, from opioid relapse prediction to sepsis early-warning systems, reflecting a broader industry move toward anticipating harm rather than only detecting it after the fact.
Why This Funding Commitment Matters
A five-year NIH-funded research commitment, running through 2030, signals that federal health research priorities now treat AI-driven medication safety as foundational infrastructure worth sustained investment, not just a vendor feature to be evaluated once and adopted or dropped. That timeline also reflects the genuine difficulty of the underlying problem: building AI systems trusted enough to influence what medication a patient actually receives requires far more rigorous validation than lower-stakes AI applications, given that an algorithmic mistake in this domain translates directly into patient harm.
What Pharmacists Themselves Are Watching For
Pharmacy professional groups say the success of this and similar AI dispensing-safety projects will hinge less on raw detection accuracy than on how well the tools integrate into existing, already time-pressured pharmacy workflows — a poorly integrated tool that requires extra clicks or interrupts dispensing speed is unlikely to survive contact with a busy retail or hospital pharmacy, no matter how accurate its underlying model. The Michigan project’s multi-year timeline gives researchers room to iterate on that human-factors challenge alongside the underlying machine learning, something shorter commercial development cycles often don’t allow.
