The AI Clinic · Issue 003Agentic AI
AI is learning to run the visit without you.
Autonomous AI takes on clinical tasks
The AI world continues to operate, grow, and change at warp speed. Patients are using AI more and more, digital interfaces that let them message their practitioners have us buried in more documentation, and AI is beginning to show signs of triaging and treating in specific medical pathways.
The more I read, the more I recognize that if we want to stay relevant in healthcare, we need to know how to use these tools.
- Two new Nature papers introduce autonomous AI agents that coordinate multi-step clinical encounters without a human directing each stage.
- A JAMA study found patient portal messages to clinicians jumped 153 percent since 2020, and office visits rose in the same stretch.
- Two live security incidents this month show AI tools wired into your inbox or admin systems are a real practice-data risk.
The Big Story
Autonomous AI takes on clinical tasks
Eric Topol reviewed two new Nature papers this week: MIRA, from a German research group, and AIME, from Google. Both introduce agentic AI, systems that can coordinate multiple steps of a clinical encounter without a human directing each stage. That is a shift from AI as a single-task tool toward AI as an active participant in clinical reasoning.
Mark's read: It's getting scary how good AI is getting. One thing this article speaks to is how AI models are remarkably precise with respect to sticking to guidelines, which of course is good, but what happens to the patient who doesn't fit the guideline? We know this occurs regularly in clinical practice, so will AI become a new source of invalidation for people who don't fit into population statistics?
From Modern Pain Care
Want the whole toolkit, not just the newsletter?
This newsletter is the free layer. Modern Pain Pro is the community and training ground where these AI skills, and the clinical reasoning behind them, actually get built, monthly drops, live coaching, and clinicians doing the work alongside you.
In the Clinic
AI post-op answers: reliable or not?
Researchers scored several LLM chatbots on 27 common patient questions after total hip arthroplasty, rating reliability, accuracy, and readability across post-op domains using blinded expert review. Some domains came back consistently accurate. Others showed meaningful errors.
In the clinic: Before directing orthopedic or rehab patients to use AI for post-operative questions, identify which topics AI handles accurately and which require direct guidance. One caveat: the models tested were from May 2025, and AI has moved fast since then. Would it do better now?
The Business of Care
Patient messages to clinicians are up 153 percent since 2020
A JAMA study found patient-written portal messages to clinicians rose 153 percent between 2020 and 2025. Office visits climbed in the same stretch, so messaging is adding to the workload rather than replacing in-person care.
For owners and staff: This volume increase is a documented driver of clinician burnout and uncompensated time. AI-assisted inbox triage and draft-reply tools are a practical response to a real, measured problem, not a nice-to-have.
Security Watch
Copilot searched your mailbox. LiteLLM handed out admin keys.
Two enterprise AI security incidents landed in the same two weeks. Microsoft 365 Copilot was exploited to search and exfiltrate mailbox content through a crafted URL, and a trust-boundary flaw in LiteLLM exposed admin credentials. Both share the same design weakness: AI systems that accept external input without adequate trust controls.
For owners: Practices using AI tools connected to email, scheduling, billing, or EHR data should treat those integrations as potential HIPAA exposure points. You cannot be too diligent here, the five-check security audit in the article is a practical place to start.
Leadership & ROI
CEO-led AI gets three times the ROI, KPMG finds
A KPMG quarterly survey found organizations where the CEO personally owns the AI strategy report three times the ROI of organizations that delegate AI entirely to IT or operations. The framing shifts from technology adoption to leadership accountability.
For clinic owners: AI investment decisions made at arm's length, delegated to a vendor or a staff member without owner-level engagement, are the ones most likely to underperform or stall. Leadership needs to be in the loop.
Research Watch
AI is straining peer review norms
As AI tools accelerate manuscript writing, peer review logistics, and global collaboration, they are also creating new pressure on editorial integrity. The piece looks at how AI-generated or AI-assisted papers challenge authorship norms, reproducibility standards, and reviewer capacity.
Worth reading: A good look at the real tension in scientific publishing. Alexander Muacevic captures the pull between AI adoption and preserving the integrity of the scientific process. Journals need to adapt without losing what makes research trustworthy.
AI for Researchers & Faculty
RAG-grounded LLMs show promise for med students
This study evaluated source-based LLMs, tools grounded in student-uploaded material via retrieval-augmented generation, for preclinical dermatology education. Tools grounded in a student's own course materials may outperform off-the-shelf ChatGPT for learning, framed through cognitive load theory.
For educators: Early evidence that LLMs grounded in your own course materials may beat generic ChatGPT for student learning, worth watching if you are considering AI tutoring or study-support tools in a training program. More reason to start learning how to train agents to become course tutors.
Money Matters
AI moves from pilot to financial impact
This sponsored piece from a healthcare technology vendor looks at how organizations are moving from AI experimentation to measurable revenue-cycle returns: prior authorization, coding accuracy, denial management, and workflow efficiency.
Read with a skeptical eye: The revenue-cycle applications here, coding accuracy, denial reduction, auth workflows, are areas where smaller practices most commonly leave money on the table. Treat the specific ROI claims with skepticism and ask vendors for comparable data from practices your size and specialty.
The last word
Agentic AI is showing up in Nature journals, patient message volume just spiked, and two real security incidents landed in the same two weeks. Different stories, same thread: AI is moving from a tool you pick up to a system running quietly in the background of your practice. Know where the human still signs off.
If you have thoughts on the newsletter, good or bad or something missing, reply to any issue and let me know. I read every one.
Get the next issue.
One email a week. The AI that matters for clinical work, in plain language.