Pharmacy General Intelligence: Q3 2025 Update
Welcome to the third in a series of quarterly updates tracking the evolution of AI through the lens of Pharmacy General Intelligence (PGI). As a reminder, PGI focuses specifically on AI's potential to perform at or beyond the level of a pharmacist, which could lead to AI agents integrating into pharmacy workflows for tasks like medication verification, dose adjustments, and patient counseling notes.
My goal is to provide a clear overview of the AI landscape, highlighting both the advancements propelling us towards PGI and the remaining hurdles. Expect insights into policy changes, industry trends, and technological breakthroughs. Your feedback is invaluable, so please share your thoughts if this is (or is not) helpful!
GPT-5 Release
Before we get into the standard policy, industry, and technology trends, I wanted a special section on the most anticipated model release…..GPT-5. The initial OpenAI GPT-4 was made publicly available on Nov 30th, 2022 and many cite this as the date that kicked off the AI conversations that exist today. The combination of performance, availability, and ease of use was the first time that a transformer-based model was readily available and accessible to the public. This availability and resulting familiarity led to the multiple use cases that we now see with AI. Since that initial release, the flagship OpenAI models have all been based on the GPT-4 architecture which led to intense anticipation of GPT-5.
To summarize the August launch, it did not meet the expectations that users held. One of the key differences with the model, is that is uses intelligent selection to essentially “route” the inputted context to the most appropriate type of model to “answer” the question.¹¹ This ended up being one of the most controversial parts of the model as power users wanted to be able to control which model they felt was most appropriate for the context being submitted. At launch, GPT-5 left many users discouraged and failed to meet expectations of many power users.
Since launch though, there has been a reversal of opinion. Power users have realized that there are capabilities that previous models couldn’t perform due to an improved “meta understanding” of the task at hand. Casual users appreciated the automatic model routing which resulted in better answers appropriate to the question being asked. In a way this is a micro trend within the seasonal trend that will be discussed below where there is a boom bust cycle of AI expectations and reality.
Q3 2025
The third quarter of 2025 has been incredibly busy. We've witnessed the release of next-generation foundation models with improved reasoning capabilities, a continued federal push to dismantle regulatory and data-sharing barriers, and an acceleration of AI integration directly into the core EHR platforms that health system users interact with daily. This momentum has also been met with foundational research questioning whether today's AI can truly develop the deep "world models" necessary for complex clinical reasoning, alongside growing legal and governance challenges that accompany rapid adoption.
Policy Shifts:
The regulatory environment this quarter had a new push from the Trump Administration to accelerate AI innovation and data interoperability in healthcare. The administration unveiled "America's AI Action Plan," which is a strategy aimed at achieving "unquestioned and unchallenged global technological dominance" in AI.¹ A core part of this plan is to "Remove Red Tape and Onerous Regulation," reflecting a clear policy to allow private-sector innovation to expand.¹ For healthcare, this includes establishing regulatory sandboxes where organizations can rapidly test AI tools.¹ This deregulatory stance could significantly speed up the development and testing of PGI agents.
This policy was quickly followed with the "Make Health Tech Great Again" initiative in late July. This effort secured voluntary pledges from over 60 major healthcare and technology companies—including Amazon, Google, and OpenAI—to a CMS Interoperability Framework.²,³ The stated goal is to finally "kill the clipboard" by enabling seamless digital data exchange between patients, providers, and applications.²,⁴ As mentioned in my previous post, the existing data silos are one of the largest barriers to successful integration of AI into healthcare software.
Many health system CIOs remain skeptical, pointing to the failure of past interoperability efforts and the initiative's reliance on voluntary participation.⁴ Key challenges persist, including:
- Patient Privacy: Once patient data is shared with third-party apps, it may lose HIPAA protection, creating significant new privacy risks that lack governance under a comprehensive federal privacy law.⁴
- Behavioral Barriers: Many experts argue that interoperability is not a tech problem but a business one, with information blocking by vendors and providers remaining the biggest impediment to progress.⁴,³
- Data Usability: Even if data becomes interoperable, its heterogeneity and poor quality mean that it may not be usable for complex tasks without significant cleaning and transformation, a crucial step for training reliable PGI models.⁴,⁷ I can provide a personal example of this where I consulted with a start up to ingest data from state public health databases. We wanted to concentrate first on just returning patient HbA1c values. In pulling the data, we discovered there were almost 100 variations of HbA1c lab type labels, with no standard key to be able to classify the lab.
Finally, there has been an industry first that has been predicted for some time. In a landmark case, OpenAI is facing its first wrongful death lawsuit, alleging its chatbot contributed to a teenager's suicide.⁵ This case underscores the responsibility and risk associated with deploying AI in sensitive domains like healthcare.⁵
Industry Trends:
The third quarter saw major EHR vendors integrating AI directly into their platforms, shifting AI from a niche, bolt-on external tool to a core component of healthcare intelligence. Oracle Health launched a new "AI-first" EHR (current only for outpatient use), built from the ground up as a cloud-native, voice-first solution with agentic AI embedded in its workflows.⁶ This represents a fundamental redesign intended to move beyond the legacy Cerner infrastructure and compete directly with Epic on AI capabilities.⁶
In response, Epic unveiled its concept of "healthcare intelligence," weaving AI into all aspects of its EHR.⁷ A key announcement was its development of a "large medical model" trained on 16 billion patient encounters from its Cosmos research database, which CIOs believe will offer unparalleled accuracy because it is trained on real clinical data, not common crawl internet data.⁷ Epic's planned entry into the AI-powered ambient scribe market is being hailed as a "watershed moment" that will accelerate the technology's adoption from pilot programs to the standard of care.⁸ The creation of these new tools by EMR vendors has had a drastic impact on healthcare AI startups as their products suddenly need to compete not only with other startups, but also the EMRs that they integrate into.⁸
2025 has become the "year of AI agents in healthcare," with health systems moving beyond pilots to create strategic roadmaps for deploying agentic AI in both clinical and operational settings.⁷,⁹ In many instances, the healthcare systems do not yet have a formal oversite process. A recent HFMA survey found that while 88% of health systems are using AI, only 18% have a mature governance structure and strategy in place.¹⁰
Technological Advancements:
This quarter was filled with new model releases. As mentioned, OpenAI released GPT-5, which it describes as its smartest model yet and a "unified system" that uses a router to decide when to use a fast model for simple queries and a deeper "thinking" model for complex problems.¹¹ Interestingly, the focus with this release has moved from hallucinations to sycophancy which is described as the tendency of a model to agree with a user’s input.¹¹ This shift is in part due to massive improvements in decreasing the hallucination rate for GPT-5.¹¹ Before GPT-5, xAI launched Grok 4 and Grok 4 Heavy, which it claimed were the most intelligent in the world.¹² These models feature native tool use and real-time search integration, allowing them to autonomously seek out and process information to answer questions—a key function for any PGI agent needing to consult external drug information resources or patient records.¹² The bundling and integration of AI agents into LLMs in a trend which started in Q2 25 and I predict will continue with each new model release.
While the models continue to progress, a paper published this quarter questioned the maximum capabilities of models built on the current infrastructure. The research introduces "inductive bias probes" to test whether a foundation model has learned a true "world model" (e.g., the underlying laws of physics) or is merely using "task-specific heuristics".¹³ The findings from this paper bring into question just how close current models can get to AGI: a transformer trained to predict planetary orbits became exceptionally good at predicting trajectories but completely failed to learn Newtonian mechanics, instead inventing nonsensical "laws" for different situations.¹³ This research highlights the "AI chasm"—the gap between high performance on a prediction task and genuine, generalizable understanding.¹³ For PGI, this is the central challenge: an agent that can predict drug interactions without a true "world model" of pharmacology and physiology could be dangerous as it does not understand the entire picture of the patient’s drug therapy.
Challenges and Opportunities
Q3 2025 has continued to build both the expectations and reality around AI. The mixture of decreased regulation partnered with better models continues to allow advancements for the efficiency and capabilities of what models can do. With EHR giants like Epic and Oracle building AI into their core products, the infrastructure for deploying these agents is moving forward. If you follow the headline AI news, currently there are quite a few articles describing the AI bubble that could potentially exist. Every presentation I give has a slide that looks at the difference between technological expectations and reality. What has been fascinating over the last 4 years is that the boom-bust expectation cycle timeline has shrunk to the point that every summer seems to be filled with skepticism, followed by 3 quarters of expectation inflation.
Looking Ahead
As we move into the final quarter of 2025, pharmacy leaders should be preparing for new AI tools delivered directly through their core EHR systems and the evaluations/discussions around them. The challenge will be to establish robust governance and evaluation frameworks to manage these tools. The question is no longer if AI will be integrated into EMRs, but how can we ensure it is built on a foundation of genuine intelligence that is safe, reliable, and trustworthy? This is a critical opportunity for pharmacy to be involved with the decision making around these new tools and included in the evaluation processes.
References
- The White House. America’s AI Action Plan. Published July 2025.
- Centers for Medicare & Medicaid Services. White House, Tech Leaders Commit to Create Patient-Centric Healthcare Ecosystem. CMS.gov. Published July 30, 2025.
- Beavins E. Trump's HHS just embarked on a thorny journey in health IT, industry experts say. Fierce Healthcare. August 14, 2025.
- Bruce G. Can Trump finally ‘kill the clipboard’ in healthcare? Becker's Hospital Review. August 1, 2025.
- Robins-Early N. AI industry pours millions into politics as lawsuits and feuds mount. The Guardian. September 2, 2025.
- Landi H. Oracle Health debuts AI-powered EHR designed as a 'voice-first' solution embedded with agentic AI. Fierce Healthcare. August 13, 2025.
- Bruce G. ‘This isn’t just hype’: CIOs react to Epic’s latest AI moves. Becker's Hospital Review. August 22, 2025.
- Diaz N. ‘A watershed moment’: CIOs react to Epic’s AI scribe launch. Becker's Hospital Review. August 21, 2025.
- Dyrda L. 2025 is becoming the year of AI agents in healthcare. Becker's Hospital Review. June 27, 2025.
- Healthcare Financial Management Association. Health system adoption of AI outpaces internal governance. Published August 12, 2025.
- OpenAI. Introducing GPT-5. Published August 7, 2025.
- xAI. Grok 4. Published July 9, 2025.
- Vafa K, Chang PG, Rambachan A, Mullainathan S. What Has a Foundation Model Found? Using Inductive Bias to Probe for World Models. Proceedings of the 42nd International Conference on Machine Learning. 2025.