⚠️ When AI Makes the Wrong Call, Who Catches It?

Abridge just signed a $775 million VA contract. Medicare's AI prior auth pilot left patients waiting 83 days for decisions. Those two things happened in the same week. I've been thinking about what separates them, and it comes down to one thing: someone was watching. Abridge spent a year in VA pilots before earning an enterprise contract. WISeR launched in January with AI making coverage decisions and apparently nobody checking whether it was getting them right. The technology isn't the variable. The oversight is.

In today's newsletter:

📰 Latest News

Abridge wins the VA's ambient AI enterprise contract at a $775.7M ceiling: Announced September 22, the five-year multiple-award contract gives the Department of Veterans Affairs ambient clinical AI across more than 75 medical centers spanning primary care, behavioral health, rehabilitation, and medical and surgical specialties. Abridge is currently the only ambient scribe vendor integrated with both the VA's legacy and new EHR systems, which is the reason it survived the enterprise evaluation when others did not.

UNC Health and Duke launch a $4.4M statewide AI network for rural North Carolina hospitals: Funded by the Duke Endowment and announced September 23, the three-year initiative will help rural hospitals and small clinics in North Carolina evaluate and implement AI tools, with a focus on giving smaller institutions the infrastructure to make AI adoption decisions responsibly rather than reactively. The network is one of the first state-level efforts to treat rural AI equity as an implementation challenge rather than purely a funding challenge.

Anthropic and OpenEvidence are giving physicians in 100 low-income countries free access to clinical AI: The partnership, reported September 23, provides physicians in low- and middle-income countries with free access to an AI tool that answers clinical questions using peer-reviewed research and treatment guidelines. For health systems where access to clinical decision support has historically depended on institutional subscriptions that cost more than many hospitals earn in a month, this represents a meaningful shift in who gets to benefit from clinical AI.

AI scribe company Heidi raises $340M at a $900M valuation: Australia-based Heidi secured $100 million in equity and a separate $240 million customer acquisition financing commitment, one of the largest combined healthcare AI capital events in recent months. The distinction between the equity raise and the financing facility matters: Heidi is not just raising growth capital, it is building the financial structure to acquire large health system customers at enterprise scale, which signals confidence that ambient AI is moving from pilot to long-term contract territory.

🚀 Technical Advances

FDA denied a partial 510(k) exemption for AI radiology software on September 17: The decision means radiology AI vendors cannot bypass premarket review based on substantial equivalence alone, and must demonstrate clinical performance directly. Given that radiology accounts for the largest share of AI-cleared devices and most cleared under 510(k), this is a meaningful tightening of the regulatory floor at exactly the moment the market is scaling.

Wolters Kluwer released a structured validation framework for evaluating clinical AI at the point of care: The framework addresses a practical gap: health systems are deploying AI tools faster than their governance structures can evaluate them. Wolters Kluwer's approach gives clinical informatics teams a structured process for assessing accuracy, evidence quality, update frequency, and transparency before a tool reaches a clinician's workflow. It is not the only framework available, but its publication by a major clinical information vendor signals that evaluation infrastructure is becoming a commercial expectation, not just an academic one.

Nature Medicine: physicians plus explainable AI outperformed either alone in predicting lung cancer outcomes: The I3LUNG study, published September 13, analyzed 2,396 patients with advanced non-small cell lung cancer and found that when physicians used an explainable clinical-and-blood-model support tool, disease-control prediction accuracy rose from 57% to 65%. The result is one more piece of evidence that the value of AI in clinical settings is not replacement but augmentation, and that explainability is what makes augmentation possible rather than dangerous.

🔬 Research Spotlight

CDS AI shows strong diagnostic potential but falls short on the trust and transparency clinicians need most: A 2026 systematic review published in The Clinical Teacher found that current AI clinical decision support tools demonstrate real potential to improve diagnostic accuracy and reduce clinician workload, but that adoption remains limited by insufficient transparency, poor explainability, and clinicians' inability to audit what the system actually did. The gap between AI performance on benchmarks and AI adoption in clinical practice is not a technical gap; it is a trust gap.

Systematic review: clinician trust in AI-CDSS depends more on explainability than accuracy: A 2026 PMC systematic review on trust in AI-based clinical decision support among healthcare workers found that trust is not driven primarily by demonstrated accuracy; it is driven by whether the clinician can understand the reasoning, trace the evidence, and detect when the system is wrong. Health systems deploying AI tools that produce recommendations without visible reasoning are solving the wrong problem.

Community health centers are using AI-CDS, but infrastructure barriers are preventing systematic evaluation: A July 2026 Commonwealth Fund case study found that while community health centers are adopting AI clinical decision support tools, most lack the informatics infrastructure to evaluate those tools systematically after deployment. The result is that tools go live, generate recommendations, and continue operating without any structured process for assessing whether the recommendations are accurate, appropriate, or safe for the specific patient populations they serve.

⚖ Ethics and Policy

Medicare's AI prior auth pilot left patients waiting 83 days and denied more than it approved: Documents released in September 2026 by the Electronic Frontier Foundation revealed that the WISeR model, CMS's AI-assisted prior authorization pilot active in six states, produced widespread care delays since its January launch, with one request left unanswered for 83 days. One vendor, Virtix, denied more prior authorization requests than it approved and was placed on a corrective action plan. Roughly 5,944 requests were denied by two vendors in the program's first three months alone. The program represents the federal government's most direct experiment with AI replacing human review in high-stakes coverage decisions, and the early results are a textbook case for why AI that operates without meaningful human oversight in healthcare creates systemic harm rather than efficiency.

🧠 AI Skill of the Week

The Skill: Use AI proactively at the point of care to strengthen your own clinical reasoning. Instead of waiting to receive AI outputs, you query AI actively to generate differentials, cross-check treatment plans, surface current guidelines, and flag what might be missing before you make a decision.

Why It Matters: The I3LUNG study in this week's Research Spotlight found that physicians using explainable AI to support their own reasoning outperformed both AI alone and physicians working without it. That improvement comes from active engagement, not passive receipt. The WISeR prior auth failures illustrate what passive receipt looks like at scale. These five habits are the difference.

Five ways to use AI for your own CDS:

Try It Now:

Go Deeper: Stay tuned for upcoming courses at intelligentmedicine.ai

💭 Final Thoughts

The Abridge VA contract and the WISeR prior auth failure share something important: they are both products of the same moment in healthcare AI, one that moved faster than our governance structures did. Abridge earned its contract by integrating carefully with existing systems and demonstrating performance across more than 75 VA medical centers over a year of pilots. WISeR launched in January and by September had produced 83-day delays, 5,944 denials, and a corrective action plan. The difference is not the technology. It is the oversight.

The research from this week reinforces something that should be obvious but often gets lost in the conversation about AI performance: accuracy is not the same as trustworthiness. The systematic reviews on clinician trust in AI-CDSS are consistent on this point. Clinicians do not trust AI tools because they are accurate; they trust them when they can understand the reasoning, audit the sources, and detect when the system is operating outside its competence. A tool that is 90% accurate but opaque is harder to use safely than a tool that is 80% accurate but transparent. The 10% matters less than your ability to recognize it.

This week's skill exists because that recognition requires a structured habit. Five questions asked consistently before acting on a CDS recommendation do not slow you down; they make you faster in the cases where the AI is right and safer in the cases where it is not. The I3LUNG data shows what happens when you combine physician judgment with AI transparency: accuracy rises from 57% to 65%. That improvement is not the AI doing more. It is the physician doing better because the AI made its reasoning visible.

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

email me @ [email protected]