AI continues expanding. OpenEvidence launched four clinical AI models, with one already being held back from general release due to dual-use safety concerns. A new Nature Medicine benchmark found that general-purpose AI outperforms specialized clinical tools, which raises important questions about where health systems should be placing their bets. And on the prior authorization front, evidence continues to mount that AI-assisted appeals work, while policy pressure from CMS is pushing payers toward faster, more transparent decisions.

In today's newsletter:

📰 Latest News

OpenEvidence launched four clinical AI models, one already locked behind access controls: Released September 3, the new model family includes Osler (fastest, about 5 seconds per answer), Sackett (deeper evidence review, about 30 seconds), and Snow (a full literature investigation, about 5 minutes). A fourth model, Darwin, was withheld from general access due to dual-use risks in virology and genetics research. All three production models are free to verified clinicians through the OpenEvidence app.

Nurses are being called to lead the governance of AI in their own workflows: A September 9 analysis from Healthcare IT News makes the case that health systems must give frontline nurses authority over AI tools that affect nursing documentation and patient monitoring, from vendor selection through ongoing oversight. Without clinical ownership of these governance decisions, accountability gaps are likely to widen as ambient AI expands into nursing workflows.

Teladoc One is going live this month, and the company is putting 100% of its fees at risk: Teladoc Health launched Teladoc One with select clients in September, a personalized AI-driven virtual care model powered by its Pulse engine that aggregates clinical history, claims, pharmacy records, and connected device data to route patients to the right care. The outcome-based payment model marks a notable bet that coordinated AI-assisted care can demonstrate measurable cost reductions at the population level.

🚀 Technical Advances

General-purpose AI outperformed every specialized clinical tool in a June Nature Medicine benchmark: A study pitting GPT-5.2, Gemini 3.1 Pro, and Claude Opus 4.6 against purpose-built tools including OpenEvidence and UpToDate Expert AI found that frontier general-purpose LLMs outperformed the clinical AI products across all three evaluation stages, including a real-world clinical query benchmark. The researchers were careful to note that regulatory compliance, EHR integration, and liability frameworks do not appear in a benchmark score, which is the more important practical caveat for health systems making purchasing decisions.

The FDA has authorized over 1,500 AI medical devices. The evidence for most of them has not caught up: As of mid-2026, radiology represents the largest share of cleared AI devices, the majority having cleared through the 510(k) pathway rather than full premarket approval. A new Clinical Trial Vanguard analysis argues that clearance via substantial equivalence is a lower bar than clinical performance requires, and that the current authorization pace has outrun the evidence base available to support actual clinical adoption decisions.

CMS is requiring FHIR-based prior authorization APIs by January 2027, now extending to drugs: A 2026 CMS proposed rule extends interoperability requirements to cover drug prior authorizations, building on the existing requirement for FHIR-based APIs across Medicare Advantage, Medicaid, and qualified health plans. Beginning this year, impacted payers must also deliver decisions within 72 hours for expedited requests and 7 days for standard requests. For clinicians, these infrastructure changes are the policy layer that will eventually make real-time prior authorization decisions possible at scale.

🔬 Research Spotlight

AI agents reduced prior auth appeal processing from 45 minutes to 5 minutes with better outcomes: At Hospital for Special Surgery, AI-assisted prior authorization appeals cut processing time from 45 minutes to 5 minutes per case while raising the appeal success rate from 65% to 100%. The result illustrates what this week's skill demonstrates in practice: the bottleneck in the appeals process is not the merits of the case, it is the time required to assemble and write the argument.

80% of prior auth denials that are appealed get overturned. Only 11.5% are ever appealed: The gap between these two numbers defines the problem this week's skill is designed to close. Denials are not being upheld because they are clinically sound; they are being upheld because the administrative burden of filing an appeal exceeds what most practices can absorb on behalf of each individual patient. AI reduces that burden to minutes.

Only 26% of hospitals feel ready for the EU AI Act obligations that took effect in August: A 2026 Frontiers in Digital Health analysis found that most hospital administrators report being unprepared for the conformity assessments, technical documentation requirements, and human oversight mandates that apply to high-risk AI systems under the Act, which became fully enforceable for high-risk systems on August 1. With approximately 75% of deployed clinical AI tools classified as high-risk, the compliance gap is not theoretical.

⚖ Ethics and Policy

The EU AI Act's high-risk enforcement is live, and most clinical AI is in scope: As of August 1, 2026, the EU AI Act's full obligations for high-risk AI systems are enforceable, covering conformity assessments, technical documentation, data quality requirements, and mandatory human oversight mechanisms. The majority of commercially deployed clinical AI tools fall into the high-risk category under the Act, including most radiology AI and clinical decision support tools. For U.S. health systems with European operations or European vendor relationships, the Act is now a live compliance requirement, not a future planning consideration. Fines for non-compliance can reach 35 million euros or 7% of global annual turnover.

🧠 AI Skill of the Week

The Skill: Use AI to draft a complete, evidence-based prior authorization denial appeal in minutes, using the denial letter language and four pieces of clinical information you already have in the chart.

Why It Matters: This week's research makes the case plainly: 80% of appealed denials get overturned, but only 11.5% are ever filed. The barrier is time, not merit. AI eliminates the time barrier.

Try It Now:

Go Deeper: Using AI to Write a Prior Auth Denial Appeal

💭 Final Thoughts

The OpenEvidence model family is worth watching closely, and not just because the models are free to clinicians. The naming is deliberate: Osler for speed, Sackett for evidence, Snow for depth. These are the figures who built the intellectual scaffolding of modern medicine. A company building clinical AI tools named after them is making a statement about what clinical AI should aspire to be: rigorous, evidence-grounded, and oriented toward the patient in front of you, not the efficiency metric above you.

The Nature Medicine benchmark finding is more complicated than the headline suggests. Yes, general-purpose frontier models outperformed specialized clinical AI on standardized tasks. But performance on MedQA and HealthBench items is not the same as clinical trustworthiness, regulatory clearance, or EHR integration. The more useful takeaway for clinicians is this: the AI tools already available to you, the ones you may already be using for other tasks, are likely more capable than you think for clinical reasoning questions. The skill this week requires nothing more than a standard AI assistant and four pieces of information from the chart.

On prior authorization: the policy and the technology are converging at the same moment. CMS is requiring faster payer decisions and FHIR-based transparency by 2027. AI is reducing the appeal burden to minutes. The 80% overturn rate suggests the merit was always there. What was missing was the capacity to argue it.

Let me know what you’d like to learn and I’ll include it in a future newsletter.

email me @ [email protected]