Clinical AI stops proving it can answer, and starts proving it can operate

Licensed guidance at the point of care

The American College of Physicians and OpenEvidence have entered a collaboration to integrate ACP content into the OpenEvidence platform, with OpenEvidence licensing ACP's evidence-based clinical guidance and content from the Annals of Internal Medicine. The partnership is designed to provide internal medicine specialists with evidence-based clinical recommendations directly at the point of care, and the organizations say it aims to support more informed clinical decision-making and improve efficiency for physicians. Internal medicine is the largest specialty on the OpenEvidence platform, and the ACP has 163,000 members located in more than 172 countries worldwide. m:…0185

OpenEvidence claims to be the most popular AI for doctors in the United States, and says it is free for all verified US clinicians. The licensing model matters more than the marketing claim: rather than training on scraped medical text, the platform is licensing the guidance itself, which gives it a provenance story that matters to clinicians and to regulators. The term evidence-based medicine first appeared in print in an ACP journal in 1991, a detail that underlines how old the standard is that these tools are now being measured against. m:…0185

Agents that show their evidence

SkinMind, a preprint multimodal agentic retrieval framework for dermatology AI, iteratively probes a large knowledge graph using eight hierarchical graph-probing tools to identify relevant entities and inspect node-level information. Its knowledge graph is constructed from a 62M-token corpus of dermatology guidelines and textbooks and 423K images, containing over 1.0M nodes and 2.0M edges. The system constructs query-specific evidence subgraphs, and improves the average score over direct answering by 8.70 points with GPT-5.4-mini and 8.89 points with Gemini-2.5-Flash across six dermatology VQA benchmarks. m:…7631

ClinFusion, a vision-centric multimodal system for medical understanding, unifies 2D and 3D image processing through a cascaded spatial-aware architecture, and outperforms proprietary models GPT-5.2 and Gemini-3-Flash on 13 out of 16 medical benchmarks, the authors say. Its RoI-grounded evaluation metric achieves the strongest correlation with expert radiologist judgment among all automatic evaluation metrics examined, the authors say, which is the more important result. A benchmark win matters less than a metric that agrees with radiologists, because the radiologists are the ones who decide whether the system is safe to use. m:…8858

From spot audits to full populations

Pharmacy leaders at Prisma Health and Baptist Health South Florida are using AI to automate administrative and compliance tasks, and the numbers describe a change in kind rather than degree. Prisma Health previously maintained a manual spot audit process for 340B contract pharmacy relationships that covered only 10% of claims; with Plenful it audits 100% of its pharmacy claims, and audited approximately 80,000 pharmacy claims in the last fiscal year, with 95% requiring no further review. The Plenful platform paid for itself within the first year at Prisma Health, the company says, through expanded audit capacity and identified 340B savings. m:…2989

Baptist Health South Florida reduced infusion patient intake delays from two to three weeks, and pharmacy owns the infusion strategy from authorization to denials, the company says. Madeline Camejo, a pharmacy leader quoted in the report, views AI as a patient experience and workforce strategy rather than a purely technical implementation. That framing is worth taking seriously: the benefit here is not a better model, but a shorter wait for a patient who needs an infusion, and an audit that covers every claim instead of one in ten. m:…2989

Observability is the new requirement

Dynatrace argues that healthcare organizations embedding AI agents into electronic health records, Microsoft 365, Workday and ServiceNow need an Enterprise AI Control Plane, and that observability supplies the evidence and context that governance depends on. AI-era observability requires correlating intent, model, infrastructure and outcome layers in real time, the author says, and traditional observability focused on system health, whereas AI-era observability focuses on decision defensibility. Without a unified observability layer, health systems face risks from model drift, fragmented workflows and regulatory exposure under the False Claims Act and updated HIPAA requirements. m:…8382

The regulatory pressure is specific. Several states have enacted laws requiring human oversight when AI influences healthcare coverage or care decisions, and the Department of Health and Human Services proposed updates to the HIPAA Security Rule in December 2024 to strengthen requirements for technology asset inventories and risk analysis. The Department of Justice prioritizes healthcare fraud enforcement under the False Claims Act. The thread running through all five sources is that clinical AI is now judged the way clinical work is judged: by provenance, by coverage, by whether a decision can be defended after the fact, and by whether a patient waited less. Benchmark wins open the door. Operations decide whether the system stays. m:…8382