Clinical decision support was supposed to make us better doctors. For most of its history, it made us better at clicking "override." As an anesthesiologist and pain medicine physician, I deal with these pop-ups every day. AI systems are starting to integrate patient history, real-time vitals, and the entire medical literature, although I can’t say with confidence that the technology has caught up to its promise. There are some very promising products sure, but we are still a ways off. Here’s what needs to catch up: the frameworks for transparency, accountability, and conflicts of interest that determine whether clinicians can actually trust what the algorithm is telling them. Today, we’ll explore some tools that are trying to achieve that. What would you improve?

In today's newsletter:

The original promise of clinical decision support was simple: put the right information in front of the right clinician at the right moment, and fewer mistakes will happen. What was delivered, for most of the EHR era, was something rather different: a constant stream of low-specificity alerts that interrupted workflow, applied to patients who clearly did not need them, and were overridden so reliably that the concept of "alert fatigue" entered the clinical lexicon as a patient safety concern in its own right. The problem was not the information. It was the architecture. Rules-based systems that fire whenever a condition is met, regardless of clinical context, trained clinicians to dismiss alerts the way most of us dismiss terms-of-service agreements.

What is different in 2026 is context. The newer generation of AI-powered CDS tools do not fire based on a single condition meeting a threshold. They synthesize the full patient record — vitals trend, lab trajectory, medication history, comorbidities, recent notes — and surface recommendations that are calibrated to this specific patient at this specific moment. The Sepsis ImmunoScore, the first FDA-authorized AI diagnostic for sepsis, is a useful example: it integrates multiple immune biomarkers with clinical data to stratify patients by risk and predict not just sepsis diagnosis but ICU utilization, mechanical ventilation, vasopressor need, and in-hospital mortality within 24 hours. That is a different kind of tool than a creatinine alert. It requires a different kind of clinical judgment in response.

The accountability gap that opens alongside this improvement is real. When a rules-based alert fires and you override it, the responsibility is clearly yours. When a contextual AI recommendation influences your clinical decision, who is responsible if that recommendation is wrong? The FDA's January 2026 guidance largely exempts CDS tools that present information for clinicians to independently review, which means the tools reshaping clinical judgment the most are among the least regulated. And when the most widely used CDS tool in the country — one that 65% of US physicians are consulting for real-time clinical queries — is funded primarily by pharmaceutical advertising, the question of whose intelligence the algorithm is expressing becomes more than academic.

Mount Sinai Health System announced an enterprise-wide deployment of OpenEvidence directly integrated into its Epic EHR in April 2026, making it the first large-scale rollout of the platform across a full clinical workforce including registered nurses and pharmacists, not just physicians. OpenEvidence draws on real-time synthesis of NEJM, JAMA, ACC, NCCN, and ACEP guidelines and is now embedded at the point of care for over 20 million clinical consultations per month across its user base. The Mount Sinai integration represents a model for how AI-powered clinical intelligence may become standard infrastructure rather than a supplemental tool.

What it means: When a decision support tool reaches nurses and pharmacists at scale inside the EHR, it is no longer a physician productivity tool. It is part of the clinical care infrastructure. The governance and accountability standards need to reflect that.

Epic has unveiled AI agents and its CoMET foundational models — pretrained on its Cosmos real-world patient data network — as the core of its 2026 AI roadmap, with over 150 AI features in development for deployment directly inside clinical workflows. CoMET represents one of the largest scaling-law studies ever conducted on real-world patient journeys, and the resulting models are designed to surface contextual clinical recommendations, predict deterioration, and support documentation across Epic's installations at over 1,000 U.S. health systems. The AI is no longer a layer on top of the EHR. For Epic customers, it is being woven into the EHR itself.

What it means: If your health system runs Epic, AI-powered CDS is coming to your workflow whether your institution has a formal AI governance strategy or not. Now is the time to ask what validation data backs each model and what the opt-out process looks like for specific clinical contexts.

The FDA's revised CDS guidance, effective January 2026, clarifies that AI-enabled clinical decision support tools that present information for clinicians to independently review are largely exempt from medical device regulation. The guidance is intended to reduce regulatory friction for innovation, but it creates a significant gap: the CDS tools with the most influence over clinical judgment — contextual AI recommendations, differential diagnosis generators, risk stratification systems — are exactly the category the guidance exempts from oversight. EHR developers and CDS vendors welcomed the guidance; patient safety advocates flagged the absence of post-market surveillance requirements for deployed AI systems.

What it means: The FDA is not the safety net for most AI CDS tools entering clinical practice in 2026. Health systems and clinicians are. Ask vendors specifically: what is your post-deployment monitoring process, and what is the mechanism for reporting CDS-related adverse events?

NEJM AI, 2024-2025

The Sepsis ImmunoScore, the first FDA-authorized AI diagnostic tool for sepsis, was validated across five U.S. institutions in a prospective study of 3,457 adult patients suspected of infection. The tool integrates immune biomarkers with clinical data to produce a risk stratification score that predicts not only sepsis diagnosis (AUC 0.81 in external validation) but also ICU utilization, mechanical ventilation, vasopressor requirement, and in-hospital mortality within 24 hours — simultaneously, from a single assessment. Read the study.

Key Finding: A single FDA-authorized AI tool can stratify sepsis patients by severity and predict multiple downstream outcomes with clinically meaningful accuracy across diverse institutions.

Clinical Implication: For intensivists and emergency physicians, this is the model for what evidence-based AI CDS should look like: prospective validation, external cohort testing, FDA authorization, and outcome prediction beyond the index diagnosis. It is a high bar. Most CDS tools in clinical use today do not meet it.

Journal of Medical Internet Research, 2026

This systematic review synthesized evidence on AI-powered CDS deployment across clinical settings and identified alert fatigue as the primary implementation failure mode: systems that generate high-sensitivity, low-specificity alerts train clinicians to override regardless of content, reducing effectiveness to near zero in high-alert environments. Studies showed that AI-powered CDS with contextual filtering — recommendations calibrated to patient-specific risk rather than population thresholds — reduced alert override rates by up to 54% compared to rules-based systems. Read the review.

Key Finding: Alert fatigue is not a clinician problem. It is a CDS design problem. Systems built around contextual AI outperform rules-based systems on both accuracy and clinical adoption.

Clinical Implication: When evaluating a CDS tool, ask for the alert override rate in comparable clinical settings. A system with an override rate above 80% is not supporting clinical decisions — it is training clinicians to ignore it.

OpenEvidence, now used by an estimated 65% of US physicians for real-time clinical queries, is free to clinicians and funded primarily through pharmaceutical advertising — a business model it has not prominently disclosed to its users despite raising over $735 million at a $12 billion valuation. The platform has content partnerships with NEJM, JAMA, ACC, and NCCN, and its recommendations are drawn from peer-reviewed evidence. Whether pharmaceutical sponsorship influences which evidence is surfaced, how prominently, or in what framing is a question the company has not publicly addressed, and no independent audit of the platform's outputs for sponsorship bias has been published.

Why This Matters: A clinical decision support tool that influences 20 million patient encounters per month and is funded by the companies that sell the treatments it recommends should carry a disclosure. Clinicians deserve to know who is paying for the intelligence they rely on at the point of care.

Clinical decision support has cleared the lowest bar. It is no longer a system of nuisance alerts that clinicians learn to dismiss. The best tools now do what the original promise described: surface the right information, calibrated to this patient, at the moment it is needed. That is real progress, and it is worth acknowledging.

But progress on the technical side tends to outrun progress on the accountability side, and the gaps here are meaningful. The FDA has exempted most CDS from regulatory oversight on the premise that clinicians independently review the output. In practice, when a tool is embedded in the EHR, accessed in under 30 seconds, and used in 20 million clinical encounters a month, the "independent review" assumption is doing a lot of work. The influence of a widely used CDS tool on clinical judgment is real even when — especially when — clinicians do not register it as influence.

The question worth carrying into your next clinical day is not whether to use these tools. Most of you already are. The question is whether you know who built them, how they were validated, who funds them, and what happens when they are wrong. That is not skepticism. That is the standard we apply to every other tool in medicine, and AI is not exempt from it.

Send me an email letting me know what you’d like me to discuss in future issues @ [email protected]