Artificial intelligence tools have been developed to predict adverse drug events, assist clinical decision support systems with medication-related decisions, optimize medication dosages, detect drug–drug interactions, and identify and prevent medication errors and drug diversion.
But as with any other novel technology, it can be hard to separate facts from hype, and discern what is actually beneficial from what’s just the latest shiny new toy. To determine what AI technology you need right now and choose the vendor that best meets your needs, it’s important to set up a structured evaluation process, said Benjamin Michaels, PharmD, MIDS, a data science pharmacist at UCSF (University of California, San Francisco) Health.
“First, you need to take the time to really look at your current state and ensure that a clearly defined need exists,” he said.
Questions to ask as part of this process include:
What problem will AI solve? For example, if you’re using AI to streamline your prior authorization (PA) process, how long does it take your staff to do PAs currently? “You need to define the baseline and agree on success metrics,” Dr. Michaels said. “It’s a lot easier to get agreement on these metrics at the beginning of the process than at the end.”
What is the use case? “Try and put yourself in the shoes of the end user,” he said. “How will you get assessment from them to know if it is helpful to achieve what they are trying to get it to do?”
How will you assess readiness for implementation? What data are required? “You have to think about [whether] the data you need even exist, and what the quality of the data is that will be going into the AI product,” Dr. Michaels said.
“You don’t want to fix something that isn’t broken,” said Andrea Sikora, PharmD, an associate professor of biomedical informatics at the University of Colorado School of Medicine Anschutz Medical Campus, in Aurora.
Dr. Sikora is working on developing a clinician-friendly health IT tool that will visualize AI-informed predictions from the perspective of the critical care pharmacist with the goal of preventing adverse drug events. “Tools like the sepsis and cardiac collapse alerts have shown that they save lives,” she said. “That’s what I’d be working toward: a tool that can improve outcomes and ideally save lives, and that has gone through implementation studies to show that people will actually use it.”
Evaluating Potential Vendors
Once you’ve decided what problem you want an AI tool to solve, the next step is evaluating the capabilities of potential AI vendors.
First, determine what technologies the vendor uses, Dr. Michaels advised. “It’s helpful to ask about specific software and hardware. If they start throwing out names of programs you’re not familiar with, you can look it up later.” Then ask what other vendors use and what separates them from those other vendors. “Also, how is this vendor technology custom, or are they just using the basic model and throwing your data in? What does it do that you can’t do just by copying and pasting a prompt into a publicly available system?”
AI vendors who want to work with you in the pharmacy space should also understand the subject matter. “Ask questions about the processes their product interacts with,” Dr. Michaels recommended.
He gave these examples:
- “Explain how a prescription goes from being signed to a 340B accumulation at a contract pharmacy.”
- “Why would a patient go to the PACU [post-anesthesia care unit] and what happens there?”
- “Explain the required steps in the preparation of a sterile compound.”
“These questions can really set vendors apart,” he said. “I’ve had vendors give me answers ranging from very advanced, extremely technical definitions and the ways they were working on them, to ‘I don’t know.’”
Sara Trovinger, PharmD, an associate professor and the director of clinical education for pharmacy at Manchester University College of Health Professions, Nursing and Pharmacy, in North Manchester, Ind., suggested that any AI tool used for pharmacy applications should have had at least one pharmacist involved in its development. “It’s cheaper to hire someone else to train these models, but those other people are not trained the way pharmacists are,” she told Pharmacy Technology Report. “If I’m going to use something in my pharmacy, I’d want to have at least one and preferably more pharmacists involved [in its development].”
Evaluating Performance
“One of the things you’ll find is that AI vendors will say it improved a certain outcome, but it’s not necessarily a patient-centered outcome like mortality or length of stay,” Dr. Sikora noted. “Just because you have a fun, fancy way of doing things, does it actually improve the things you care about? It’s kind of like buying a specialized grilled cheese maker. I’m not saying that’s not a cool contraption, but grilled cheese on a pan is still pretty darn good.”
Once you have a short list of vendors who might meet your identified needs, ask for demonstrations and proof-of-concept projects, as well as references from existing customers, advised Kendall Gross, PharmD, the pharmacy informatics manager at UCSF Health. “These should demonstrate their current functionality and intended use cases, expected outcomes, and return on investment, with time lines.”
Factors to consider are:
- the tool’s usability by your clinicians or other end users;
- integration into your existing workflows; and
- whether the tool integrates with your current systems or is a stand-alone product.
References should be able to tell you about the AI’s real-world benefits, as well as their measurement strategy, actual implementation time lines, and how much education and/or change management was needed to adopt it. “Were there any FTE [full-time equivalent] needs or changes? How much vendor support did they receive? What safeguards were there against possible errors?” Dr. Gross said.
Dr. Trovinger advised tempering expectations about how much AI can transform your hospital pharmacy practices, at least for now. “It’s very unlikely that [the technology is] going to be able to do something that you’re not already able to do,” she said. “But is it allowing you to automate processes that would otherwise take you a long time? That’s where the space really is at present. [AI is] going to continue to do better things, but probably not as fast as people think.”
The sources reported no relevant financial disclosures. Portions of this article are based on a session at the ASHP Midyear 2024 Clinical Meeting & Exhibition, in New Orleans.
This article is from the June 2025 print issue.