The Future of Pharmacy
Better outcomes require better architecture.
- Home
- Solutions for Providers
- Future of Pharmacy
Clinical AI is not a future question,
it's a present one.
Clinical pharmacy has fundamentally expanded in scope. Complex decisions involve synthesizing patient data, institutional protocols, pharmacokinetics, and evolving guidelines, and happen dozens of times a day across entire health systems.
The infrastructure to support that level of reasoning now exists. But not all of it is built to the same clinical standard, and the differences are consequential. What follows is a framework for knowing what to demand.
From computational models to agentic clinical AI.
Fifty-five years of computation in service to clinical judgment have lead us to this point.
Computer-aided dosing
Sheiner at UCSF demonstrated that a mathematical model combined with patient-specific data could guide dosing as well as expert clinicians. The first proof that computation could individualize therapy.
Population PK and Bayesian methods
Sheiner, Beal, and colleagues established population pharmacokinetic modeling as a scientific discipline. NONMEM gave clinicians a principled method for individualizing doses from sparse, real-world patient data.
Early bedside dosing tools
The science found its first clinical implementations in standalone dosing software for narrow therapeutic index drugs. The clinical evidence was compelling; the infrastructure was not: manual data entry, no EHR connection, no feedback to the chart.
Clinical data infrastructure
The HITECH Act drove US hospital EHR adoption from single digits to near-universal, with 96% of acute care hospitals possessing certified EHR technology by 2015. FHIR interoperability standards, cloud infrastructure, and bidirectional APIs then removed every structural barrier that had kept MIPD out of routine clinical practice.
MIPD in production
EHR-integrated precision dosing platforms launched at scale, bringing Bayesian individualization directly into clinical workflows. Patient data flowed in automatically, and dosing guidance flowed back into the chart without manual entry.
Clinical proof and continuous learning
A decade of published clinical evidence validated MIPD's patient impact. Models deployed at scale began retraining on real-world institutional data, improving accuracy over time as continuous learning entered clinical practice.
The rise of LLMs
General-purpose large language models entered healthcare and were rapidly adopted for informal clinical queries. Fast and fluent, they were not designed for clinical use, and the gap between consumer AI and clinical-grade AI became visible.
Agentic AI
Specialized agents for clinical retrieval, reasoning, and safety, orchestrated to deliver transparent, defensible guidance at the point of care. Glass box architecture: the LLM communicates the output; it does not generate the clinical logic.
The Problem
Most AI tools are not built for clinical use.
Large language models (LLMs) inherently contain structural features. However, these are not simply limitations that will be fixed in a new release to be suitable for clinical use.
Hallucination
Confidently generates plausible-sounding but factually wrong information. Confidence is not accuracy.
No context interrogation
Fills gaps with assumptions rather than asking clarifying questions. A system that doesn't know what it doesn't know is a liability.
No native constraints
Treats dose limits, allergies, and contraindications as soft suggestions rather than enforced boundaries.
Determinism gap
Two identical prompts can produce meaningfully different clinical guidance. Probabilistic reasoning, clinical consequences.
Knowledge frozen in time
Trained on stale data, unaware of updated guidelines, new evidence, or outcomes specific to your institution.
of clinicians report using an AI tool for clinical work.
Yet only 1 in 4 have received training from their employer to evaluate what they are using.
Elsevier, Clinician of the Future 2025; Medscape / HIMSS 2024
The gap between a plausible answer and a precise one is where patient safety lives. That gap is exactly where off-the-shelf LLMs fall short, and why the future of clinical decision support belongs to specialized agentic architectures.
Sirj Goswami
Co-founder & CEO, InsightRX
The Solution
Six things to demand from an AI-powered clinical decision support platform.
These are architectural requirements, not features. A platform that meets all six is built for clinical use; one that meets fewer has been adapted for it.
01
Glass box transparency
Every decision traces to its source: a national guideline, a local formulary entry, a patient data point. That trace must be explicit and accessible. Black box outputs are not defensible at the bedside.
02
Agentic reasoning, not generative text
The language model acts as an orchestrator, not an oracle. Clinical queries decompose into sub-tasks resolved by deterministic tools: validated PK models, curated knowledge bases, hard-coded safety agents. The LLM communicates the output. It does not generate the clinical logic.
03
Continuous learning from outcomes
Models retrain on real outcomes from actual populations, not static training data. The slow loop, retraining across institutional data over months and years, separates a static tool from one that improves with your practice.
Learn about Continuous Learning
04
Embedded precision dosing
Bayesian MIPD is embedded natively, not bolted on. Dosing adapts in real time to each patient's pharmacokinetic profile, using population models validated against your patients and updated as evidence evolves.
05
Bidirectional EHR integration
Patient data flows automatically from the EHR into the platform, and guidance flows back into the chart. A system requiring manual data entry — or living outside the clinical record — will not be used consistently enough to change outcomes.
06
Enterprise-scale standardization
The platform translates specialist expertise into consistent workflows across departments, facilities, and skill levels. A tool that works in one unit is not a platform.
Read on
Go deeper on the architecture behind clinical AI.
Learn more about implementing proper systems, and about our broader vision for the future of AI in healthcare.
Maximize therapeutic success with precision medicine.
