Notes from inside health-tech.
Short reads on AI tools, the products actually shipping, and the regulatory notes worth paying attention to. Two posts a week, no filler.
- Regulatory
The FDA Now Lets CDS Software Give One Answer. Here's What That Actually Changes for AI Teams.
FDA's January 2026 revision to its Clinical Decision Support guidance quietly dropped the rule that non-device CDS had to offer multiple options. For teams building clinical AI, the single-recommendation change matters, but the bigger signal is what the guidance still refuses to say about LLMs.
Jameson DainesMay 30, 2026 - Research
When the Scribe Can See: A Vision-Enabled AI Cut Medication-History Omissions From 358 to 10
An npj Digital Medicine study put smart glasses on pharmacists and gave the AI scribe both audio and video. Omission errors in medication histories dropped from 358 to 10, and audio-only accuracy fell to 81%. The result reframes what ambient documentation is actually missing.
Jameson DainesMay 30, 2026 - Health-Tech Strategy
The FDA changed what counts as a breakthrough AI device. Most product teams haven't noticed.
The FDA's breakthrough device criteria have shifted toward multi-problem AI solutions and away from single-indication tools. Here's what that means for product teams deciding where to invest.
Jameson DainesMay 23, 2026 - AI in Clinical Tools
LLMs leak patient information even when you tell them not to
A new Frontiers study tested LLaMA, Mixtral, and Claude generating clinical notes from patient conversations. Every model leaked third-party personal information, and prompt-based privacy instructions didn't fix it.
Jameson DainesMay 23, 2026 - Industry
Ambient AI Scribes Cut Documentation Time by 16 Minutes. Here's Why That Number Is Both Impressive and Insufficient.
A study of 1,800 clinicians across five academic medical centers found ambient AI scribes reduce documentation time by 16 minutes per encounter. Reading that finding carefully tells health-tech product teams something more useful than the headline does.
Jameson DainesMay 16, 2026 - Industry
LLMs Are Strong at Naming Diagnoses. They're Still Weak at the Harder Part.
A Mass General Brigham study evaluated 21 LLMs on iterative diagnostic reasoning rather than final-diagnosis recall. The distinction between those two tasks is the most important thing in the paper for health-tech product teams to understand.
Jameson DainesMay 16, 2026 - Research
An AI Model Out-Diagnosed ER Doctors in a Real Hospital Study. The Fine Print Matters.
A Harvard and Beth Israel study published in Science found OpenAI's o1 model correctly diagnosed ER patients in 67% of triage cases vs. 50-55% for attending physicians, but the study's design and scope raise questions that the headlines largely skipped.
Jameson DainesMay 15, 2026 - Research
One Model, 14 Diagnoses: What Aidoc's FDA Clearance Says About Radiology AI's Next Phase
In January 2026, the FDA cleared Aidoc's CARE foundation model to triage 14 acute conditions from a single abdominal CT scan, marking the first time a multi-indication AI platform built on a single foundation model has cleared regulatory review.
Jameson DainesMay 13, 2026 - CLINICAL AI
Abridge just bought NEJM into the exam room. Here's what that actually changes.
Abridge announced multi-year content deals with NEJM Group and JAMA Network this week. Most coverage framed it as a hallucination story. The actual significance is distribution, and whether clinicians will trust the citations enough to change what they do.
Jameson DainesApril 18, 2026 - CLINICAL AI
The 16-minute paradox: why ambient scribes work even when the time savings don't
The largest real-world study of ambient AI scribes found clinicians saved 16 minutes per 8-hour shift. That's smaller than the pitch promised, and bigger than the number suggests. Here's what teams are measuring wrong.
Jameson DainesApril 18, 2026 - AI IN HEALTH-TECH
The AI health tool review most teams aren't doing
Accuracy benchmarks dominate AI tool evaluations. For patient-facing apps, five other questions matter more. Here's the checklist product teams should actually run.
Jameson DainesApril 15, 2026 - AI IN HEALTH-TECH
The AI health boom is missing something
Microsoft launched Copilot Health. Amazon opened Health AI. Google is integrating CGM data into fitness coaching. All of it misses the same thing.
Jameson DainesApril 14, 2026 - AI IN HEALTH-TECH
Why behavior matters more than features in AI health tools
Most AI health tools are shipping features patients never ask for. The teams that win in the next five years will be the ones who built around behavior first, and let the feature list follow.
Jameson DainesApril 14, 2026