I just finished reading Ben Michaels' Q2 Update on AI, which caused me to go back to an earlier blog I wrote (Who will bell the cat?) and some conflicting reactions to the opinions of CEOs of some of the larger AI firms about slowing the pace of AI model development (lest we harm ourselves) and tried to pile that onto my recent experiences using AI. I came up with the following:
1) Just because you could do something with AI doesn't mean it makes sense to. You have undoubtedly heard about the the panic around the notion that AI will wipe out a number of jobs, including those of healthcare workers. Undoubtedly it will, but the question really appears to revolve around where the value is.
What has been recently nicknamed the "SaaSpocalypse" is an interesting example. The angst arose from the notion that someone could simply ask a general-purpose API system to build a customer-relations system (CRM) or a human-resources system (HR) exactly the way they want it and get rid of the seat-based subscription they currently have. Stocks fell... for a few days.. but then people realized something that the SaaSpocalypse doom-sayers didn't consider. The value in these systems is not just in their software; it also exists in the configurations and rules that have been built into each customer's system and the mass of historic data that has accumulated within those systems.
Sure, someone could work with an LLM system to develop their own software (if they could actually remember or had documented all the customizations they had put into the current system) but then they would have to move all their data into the new system, and verify that the new system operated with that data the way it was supposed to. It wouldn't happen overnight, and the customer would have to continue operating while all that was going on. The sheer inertia of that effort defeats the notion.
That being said, the best of these systems are building AI functionality into their SaaS systems to make their users more efficient, and are likely changing their business models away from the notion of seat-based subscriptions, but many of those supposedly doomed systems are alive and doing well.
What does that mean for pharmacy practice? It means that we need to build value that will be hard for AI to replace. Frankly, a professional whose only talent is to quote the most recent studies will likely get replaced. A professional who can use AI to acquire and use that most current data to derive a coherent and successful therapy plan, especially when there is no clear right answer, will still be irreplaceable. Are we teaching new pharmacists those skills?
2) Current AI systems can only know what is already known. It cannot invent something entirely new. It can perform comparisons based on what is already known. So, when you look at the AI systems being developed for pharmaceutical discovery, what that AI will be used for is to find candidate molecules based on the chemistry of the interaction between a candidate drug product and a human cell.
This is no small feat, and it will likely shear years (if not decades) off the process to generate new drug moieties, but those moieties will still have to go through (possibly accelerated) animal studies, and phase I, II, III and IV human studies to be released.
Note that this is not LLM-based AI. This "agentic AI" aimed very carefully at a specific problem. It is still using a wide range of data, but it won't create an image of a beautiful sunset.
What does this mean for pharmacy practice? Creativity is still a uniquely human property. Using AI can digest a lot of information for us, but it isn't going to get a flash of intuition. The knowledge and, yes, the intuition that comes from years of caring for patients is still ours to provide. Who models these behaviors to pharmacy students and new graduates?
3) The answer you get from AI depends very much how you ask the question. I have been using one of the LLM-based AI tools to build out a drug-related project. I had to invent the premise. I had to excruciatingly review the output data (and get review from colleagues).
There can be no doubt that it has made me more productive. It could review and summarize data I already had in a minute that would have taken me a week to work my way through. It could find resources I wasn't aware of.
But none of my project components came out completely right because I didn't phrase the question entirely correctly. In some cases, it involved my discovering additional questions to ask or requirements to specify based on the information returned from the AI. The fact that data isn't exactly pristine doesn't help.
What this means for our practice is that we need to acquire a new set of skills to properly use AI systems, and we are still going to have a lot of work to do when the AI returns results. Some of this can be learned by just trying. But, for the rest, who will teach us how to use these tools?
As with "Who will bell the cat?", I know the questions to ask. I am not certain I know the answers.
As with all my writings, these thoughts are my own, not AI generated, and not necessarily those of ASHP.
Dennis A. Tribble, PharmD, FASHP
Retired
Ormond Beach, FL
tribbledennis@gmail.com