I recently gave a talk at the 2026 Presbyterian College School of Pharmacy Preceptor Workshop on the history of AI in precision dosing and where the field is headed next. One point I wanted to make clear from the start is that AI in healthcare did not begin with ChatGPT.
Over the last few years, the conversation around AI has exploded. Every company has an AI strategy. Every product has an “AI-powered” label attached to it. It can feel like AI in healthcare suddenly appeared in 2022. But in precision dosing, AI has been evolving for more than five decades.
The journey starts in 1969 with Dr. Lewis Sheiner, one of the founders of modern pharmacometrics. Long before anyone used the term “artificial intelligence” in medicine, Sheiner published work exploring whether computers could help individualize warfarin dosing using patient-specific data and mathematical models. The core idea was remarkably similar to what we still pursue today: use computation to improve clinical decision-making. That work eventually led to population pharmacokinetics, Bayesian dosing, NONMEM, and the broader field of model-informed precision dosing (MIPD).
By the 1980s and 1990s, much of the underlying science existed, and early clinical dosing tools had started to emerge, but adoption remained limited. What didn’t exist was the infrastructure to scale it. Systems lacked EHR integration, workflows were manual, and the tools remained largely confined to academic centers.
The 2010s changed that. EHR adoption accelerated, cloud computing matured, and interoperability standards like FHIR made scalable clinical platforms possible. Precision dosing systems could finally operate at the bedside across large health systems. And something important happened as these systems scaled: the models themselves improved from real-world clinical data. Population PK models became more accurate across special populations like pediatrics, obesity, and renal impairment as more data flowed through the platforms. In many ways, this was already machine learning. The systems were continuously improving predictions based on accumulated data. It just didn’t resemble the modern consumer AI wave people associate with ChatGPT.
What changed in the 2020s was not the arrival of AI itself, but the emergence of a far more unified and capable generation of AI systems. Large language models (LLMs) unified and dramatically advanced capabilities that previously existed across fragmented NLP techniques: understanding unstructured text, synthesizing broad information, and communicating naturally with humans. Rather than relying on brittle multi-model pipelines that often felt clunky and unnatural, LLMs enabled AI systems to perform these tasks in a far more adaptable, conversational, and intuitive way for clinicians and patients.
That shift matters, but it also created confusion; general-purpose chatbots are often being mistaken for clinical systems, and those are not the same thing. Clinical decision-making, especially in precision dosing, requires deterministic calculations, hard safety constraints, validated workflows, and reproducible outputs. General-purpose LLMs were not designed for that. They can hallucinate, fail to ask clarifying questions, ignore institutional constraints, and generate different answers to the same problem. That does not mean LLMs are useless in healthcare. Quite the opposite: I believe they will fundamentally reshape clinical workflows. But the architecture matters enormously.
The future is not a smarter chatbot making autonomous clinical decisions. The future is agentic systems built around transparency, specialized reasoning engines, and enforceable clinical constraints. A general-purpose chatbot is fundamentally a black box. You ask a question, get an answer, and have limited visibility into how the system arrived there. Clinical care cannot operate that way.
Agentic systems flip that model. The LLM acts as an orchestrator, not an oracle. A clinical question is broken into smaller components, each handled by specialized agents or validated tools designed for a specific task.
Some agents encode knowledge such as formularies, protocols, antibiograms, and evidence-based guidelines. Others perform PK/PD simulations, Bayesian dose optimization, and broader clinical calculations through validated pharmacometric and rules-based engines. Safety constraints, dose ceilings, contraindications, formulary restrictions, and institutional policies are enforced architecturally rather than left to probabilistic interpretation by the LLM. This is where precision dosing becomes especially important in the broader AI conversation. The PK/PD layer is not replaced by the LLM; it becomes one of the foundational reasoning engines within the agentic architecture itself.
In other words, precision dosing evolves from a standalone dosing workflow into a broader clinical decision support capability. The same validated PK/PD engines that optimize dosing can now operate alongside agents synthesizing microbiology, laboratory data, institutional guidance, patient-specific risk factors, and longitudinal clinical context simultaneously.
In the talk, I walked through what this could look like within an agentic clinical workflow. Different agents operate across specialized domains, retrieving and synthesizing patient context, evaluating evidence and institutional knowledge, performing pharmacometric calculations, validating safety constraints, and coordinating broader therapy optimization tasks. Importantly, these agents do not operate freely or independently. They function within a safety harness built around deterministic tool calls and enforceable clinical rules. The agent itself is not inventing clinical logic or “guessing” recommendations. It is orchestrating computational and clinical decision-making frameworks underneath it.
One of the important tools I highlighted is the clinical knowledge graph. Healthcare decision-making is deeply relational: drugs interact with organisms, organisms connect to resistance patterns, therapies carry toxicities, and patient-specific factors influence every layer of the decision process. Knowledge graphs allow agents to reason across these relationships in a structured, traceable, and clinically grounded manner rather than relying solely on unstructured language prediction.
The clinician ultimately receives a synthesized recommendation with transparent calculations, supporting rationale, citations, and institutional constraints already incorporated into the workflow. Importantly, this also expands what precision dosing platforms can become. Historically, many PK/PD systems were confined to isolated dosing workflows. Agentic architectures allow pharmacometric engines to operate as components within broader clinical decision-support systems spanning therapy optimization, medication safety, and longitudinal medication management.
Through all of it, the clinician remains at the center, which might be the most important through-line across the last 57 years of this field. From Sheiner’s early work in 1969 to modern agentic AI systems, the goal has remained remarkably consistent, which is to use computation to augment clinical judgment, not replace it.