Advertisement

Blog Viewer

AI Pharmacist verification is a data problem…..not an AI problem.

By Ben Michaels posted 12-03-2025 09:47

  

Last year, at the 2024 ASHP Conference for Pharmacy Leaders in Chicago, I recall receiving a particular question following our second presentation. During the Q&A session, an audience member asked, "When will there be an AI pharmacist verification product?"

What was particularly interesting about this question is that the question was not “if”, but “when” will this product exist.  Until that point, I had only been asked if it was possible.  At the time, I responded within the next 5 years and with recent developments, the possibility of AI verification built into our EMRs seems more realistic now than it was a year ago.

Before we dive into the specifics around what a solution would look like and the barriers to why we aren’t using this currently, I want to outline a parallel track where I have not only been an observer but also an active user.  Coding assistants have and continue to be a large focus for large foundation-model companies.  This should come as no surprise since most of the individuals involved in the building and configuration of these models have or previously held a position where writing code was a part of their role.  That subject matter expertise of what an individual needs in a coding assistant is easier for them to translate into products than something that they don’t have experience in such as healthcare. 

The evolution of coding assistants through the various model versions and AI related infrastructure improvements has seen drastic improvement over the past 5 years.  In my experience it has gone from basic suggestions and corrections to the ability to produce a working prototype or template within a short amount of time.  I rarely go to a documentation website anymore and instead ask the assistant how to fix or modify code.  When it can’t produce the right answer, I supply the website URL with the question and let it figure out the answer.  Within the past year, this has improved to a point where I will let AI agents work asynchronously on bug fixes or scaffolding a new feature.

I bring this up because it has changed the way that I work along with the time and effort it takes me to accomplish tasks.  Pharmacy verification will go through a similar process of reference, assistance, and then automation.  The technical barrier to why this is not already integrated into the process is the data.

Current AI models excel in answering questions and retrieving information.  The breakdown in healthcare is that the model does not have access to the specific patient information that it needs to answer the question.  In our case, think of what you evaluate when verifying a medication.  To make sure it hits all the “Rights” for a patient, you need allergy information, past medical history, other current medications, labs, cultures, vitals, etc.  before determining if the medication is appropriate for the patient and this is where models don’t have the data needed to accurately determine this. 

It gets even more complicated.  Think about verification of orders for investigational medications that rely on protocols or individual policies for medications that are specific to your institution.  Also, guidelines on proper use of medication are constantly being updated which requires continual continuing education.  A data pipeline with access to real-time patient specific data elements along with a continually updating reference store is what will be required.

Why am I bringing this up?  Is the profession of pharmacy obsolete once the data flow issues are resolved?  Not if we stay relevant.  What I mean by this is that outside the current legal and regulatory restrictions requiring human verification of medications, we need to expand the services and demonstrate the value that pharmacists provide.  Additionally, we need to be involved with context and data engineering that has become a focus for AI along with becoming experts on the integration of this technology into our workflow.  Last, our leaders need to be thinking now about what redeployment of their pharmacy workforce would look like.  If we wait until AI starts verifying orders, it will be too late. 

 The next 10 years will be filled with change for pharmacy.  Declining pharmacy school admissions, lower rates of licensure exam completion, and retirements from the workforce will increase the demand on the workforce.  As a profession, I encourage us to continue to be medication experts and expand this knowledge into the application and development of AI systems that help provide more effective patient care while supporting the demands of the workforce. 

2 comments
56 views

Permalink

Comments

01-02-2026 14:36

Thanks for this post, Ben. I also think we need our accrediting organizations, CMS, and state laws to contemporize their stance on pharmacist review of medication orders so we can feel less constrained when it comes to creating these efficiencies and focusing on activities that require our full expertise.

12-10-2025 13:15

Ben,

I think that the interesting question is “verifying what?”.

If we are talking about verifying products for dispensing, we should be willing to admit that all that really entails is verifying that two labels describe the same thing and than none of them appear to be u usable. It’s not like pharmacists have eyeball that are calibrated for this purpose.

if we are talking about inspecting CSPs, most of which involve injection into another, I would be surprised that anyone, even a pharmacist could look at an IV bag and tell that anything, much less the right amount of the right thing had been injected into it. That leaves inspection for signs of degradation or particulates all of which could be automated. Particulates are particularly worthy of discussion since human visual acuity fails at particulate sizes smaller than 50 microns while potentially damaging particles go down to about 5 microns.

So, the only thing that leaves is verification of orders for which we already have some auto verification processes. 

in my not too humble opinion, the AI opportunity may not be too hard for order verification.