Clinical decision support has been promising the same thing for 30 years: The right evidence, to the right clinician, at the right moment, without adding cognitive burden. What it has mostly delivered is alert fatigue. A new guideline drops. Informatics translates it into rules. Rules become alerts. Alerts interrupt clinicians mid-workflow. Clinicians dismiss them—at rates that routinely exceed ninety percent—because the interruption does not match the moment. The evidence is current, the technology is expensive and the behavior change is small.

The frustration is not that we lack evidence. It is that we have been delivering evidence through an architecture designed in the 1990s for a different problem. Rule-based decision support assumes the EHR is the intelligence. It is not. The clinician is. And increasingly, so is the AI agent sitting alongside the clinician.

That shift—from EHR-as-rule-engine to agent-as-reasoning-partner—changes what clinical decision support can be, who needs to own it and how health systems will evaluate what is worth paying for. A pattern emerging in the broader AI industry, under the name “skills,” points to what I believe is the most important reframing of clinical decision support in a generation.

A skill is a small, versioned, signed package of procedural knowledge—a description of when to use it, instructions for how to do the thing correctly, curated reference material and often a few deterministic helper scripts. The agent sees the description of every available skill, loads the ones that match the task at hand, follows the guidance and unloads them when the task is done.

What a clinical “skill” looks like

A clinical “skill” is a portable, machine-readable artifact that captures how to handle a specific clinical situation according to the best available evidence. It is not a rule. It is not a chatbot prompt. It is not a PDF guideline. It is a structured package that an AI agent can discover, load and execute against the data in the patient’s chart.

When the agent is reviewing a chart or answering a clinician’s question, and the patient’s data matches a Skill’s scope, the agent pulls the Skill into context and executes it. The recommendation surfaces with its citation attached, so the clinician can audit the reasoning in one click. This is not a model guessing. It is an agent executing a curated, evidence-based procedure authored by people whose job is to know the evidence.

Why this is structurally different

Three things change when decision support is distributed as skills rather than rules. First, authorship moves to the people who own the evidence. When a major specialty society publishes an updated guideline today, the journey from publication to bedside runs through every EHR vendor’s informatics team and every health system’s build cycle—often 12–24 months. In a skill’s architecture, the issuing society publishes the skill itself and subscribers receive the update the day it ships. Second, trust becomes legible. Every skill carries its provenance on its face—which guideline, who authored it, when it was last reviewed, what evidence grade. The evaluation artifact and the execution artifact are the same thing, so there is no gap between what was approved and what the software actually does.

Third, expertise becomes composable. A patient with heart failure, chronic kidney disease and type 2 diabetes does not need three competing alerts. The agent loads three relevant skills and reasons across them and surfaces a single coherent recommendation instead of three interruptions. Composition at the reasoning layer is something rule engines have struggled with for decades. For agents with skills, it is the default.

Who authors the skills?

Not a single entity, and that is the point. A healthy ecosystem will draw from four sources: Specialty societies, which already synthesize evidence into guidelines and are the natural authoritative publishers of skills in their domains; established commercial knowledge vendors, whose evidence libraries become subscribable skill libraries shipped as agent-ready artifacts rather than human-read articles; health systems themselves, authoring local skills that reflect formulary and institutional practice; and a peer-contributed commons for niche conditions and fast-moving evidence, with vetting rather than authorship as the gatekeeping function.

What this means for health system strategy

For executives and CMIOs thinking about the next five years of clinical AI, the implications are worth taking seriously. Knowledge sourcing becomes a procurement decision with real consequences—the libraries you subscribe to are the ones your agents will execute against. Clinical content governance shifts from being a project to being an operational competency; organizations that do this well will compound advantages over time. The EHR’s role shifts from being the primary author of decision support logic to being the platform that discovers, authenticates, delivers, executes, audits and governs skills on behalf of the organization.

And measurement finally becomes possible. Because every skill execution is logged, attributable to a specific guideline version, and linked to a specific clinical context, health systems can answer questions that have been nearly impossible to answer with traditional CDS: Which guidelines are actually changing care, where the evidence-practice gap is widest, which skills correlate with improved outcomes and which ones clinicians consistently override. This is the missing feedback loop between evidence and practice.

The shift worth making

None of this is inevitable. Trust infrastructure must be rigorous. Semantic interoperability must keep improving. Business models and liability questions are unsettled. And clinicians conditioned by a generation of bad decision support will only give their attention back to tools that are visibly better than what came before.

But the building blocks are visible in adjacent industries, the underlying AI capability is already in clinical hands, and the appetite for something better than the last generation of CDS is, I suspect, universal. The next generation will be built for a world where the intelligence sits alongside the clinician, and its expertise is delivered by the community of people who know the medicine best. That is a future worth building together.

Learn more about how TouchWorks is embracing AI capabilities.