The Bandwidth: Decoding Health + AI

Good to be back. It's been a full week, and I'm glad to be writing to you again. I'm sticking with the 1-2-3 format from a couple of weeks ago, and I hope you like it too. Let’s jump into it.

1) The shape and skill of the new agents

Agents with shapes are getting a lot of attention lately. Meta's Muse is the No. 1 free app in the U.S. App Store, Microsoft's new Autopilot works overnight and monitors Teams channels, and OpenAI launched dots at DevDay, a business-focused agent that lives inside ChatGPT. FWIW, I'm a Grok user, and to xAI's credit, its agents had shapes before the others did. I find the user experience incredibly simple and accessible.

Axios' Jim VandeHei calls it a live look at a post-app, post-web world, where we tell agents what we want done instead of opening software, and he points to Ben Thompson's bet that no human will touch clunky enterprise software again.

On skill, Casey Newton at Platformer found dots did about two hours of work for about 15 minutes of his effort, from declining a conflicting invitation to drafting emails and answering his bookkeeper's questions. With Grok, I can train it on a role or workflow by letting it watch how I complete a task, and it learns and takes over. That's the hard part of agentifying workflows: someone has to show the agent how the work really gets done - Grok’s bots are removing some of that friction.

2) Everyone wants a forward-deployed engineer

Forward deployment isn't a new idea. But it is gaining traction as far as AI transformation goes. Box CEO Aaron Levie wrote this week that a big trend among the enterprises he talks to is placing internal forward-deployed engineers in departments to bridge what AI can do and how the work actually gets done, a function that barely existed two years ago. Here’s evidence of AI creating new jobs and roles. Anthropic made it official with a $100 million Frontier Academy to train 10,000 "Frontier Deployed Engineers" by the end of 2027, and Atlassian is pitching embedded engineers of its own.

The health care tie is in Anthropic's first cohorts, where Novo Nordisk sits alongside traditional professional services firms like Accenture, Deloitte and McKinsey.

The models aren't the bottleneck anymore. Getting them into real workflows is, and that takes people who understand both the technology and the work.

3) Who is responsible when AI goes off the rails

This week Sam Altman told Politico that the world should accept "some bad things happening" for the benefits of AI. That may be a fair philosophy, but someone has to answer for the bad things, and I for one have been scratching my head about why more of these companies haven't been sued, especially after OpenAI's agents hacked Hugging Face. Hugging Face itself hasn't sued, but a nonprofit, the Legal Advocates for Safe Science and Technology, is seeking an injunction. The Information says its standing looks thin, and a think tank official told the outlet to expect many more suits against OpenAI and others.

Washington is playing catch up. Rep. Lori Trahan's new CLAIM Act discussion draft released this week would make clear the labs are liable, by having courts presume an AI system acted with the state of mind a person taking the same actions would have had. Altman himself has said there should be a new liability framework for AI-powered cyberattacks.

OpenAI safety lead David Robinson quits, calling the company's culture "broken" in The Atlantic — Robinson, who drafted OpenAI's current Preparedness Framework and oversaw safety reports for 12 frontier launches, left after 3.5 years and wrote that "the time for trial and error is over" on AI safety risks, arguing labs should be run like nuclear plants and airports, with layers of redundancy.

The top 100 gen AI consumer apps: 7th edition — a16z's latest ranking finds only 4.5% of U.S. consumers hold a paid ChatGPT, Gemini or Claude subscription, and the top 1% of payers account for 19.5% of observed consumer AI spending. So, for all the hype, we’re still at the beginning of AI adoption, and the technology remains more a focus for enterprise use.

OpenAI enlists doctors to enhance ChatGPT's health advice — OpenAI has contracted hundreds of physicians across 49 languages to review ChatGPT health conversations for confusing or dangerous answers, while Business Insider notes the legal landscape around chatbot medical advice is still untested.

New open-weight AI models could disrupt the industry — Reflection AI, which Nvidia backs, and France's Mistral are releasing open-weight models that the companies say match leading Chinese models and that businesses can download, customize and run more cheaply than closed ones, though making powerful models freely tweakable also raises cybersecurity risks.

Will payers claw back hospitals' AI savings? — Fitch's new AI stress test says hospital cost cuts from AI in diagnostics, imaging and administrative work could lead payers and policymakers to cut what they pay for those services, though it rates the overall credit risk to hospitals as low and doesn't expect AI to change any provider's credit rating over the next five years.

Utah plows ahead with more health AI pilots for prescriptions, women's health — Utah added five new agreements to its AI sandbox, including Nolla Health's AI to start prescribing acne treatments, and now requires third-party audits of participants' accuracy claims after criticism from the state medical board, though it's unclear any pilot can drop clinician review without FDA sign-off.

Ochsner uses AI to widen clinical trial screening — Ochsner Health reports that an AI platform from Paradigm Health, connected to its patient records, expanded screening capacity for oncology trials by 41% and cut manual screening work by about 75%, with research staff still confirming eligibility.

How Cleveland Clinic is developing AI to monitor the brain — Cleveland Clinic and the startup Piramidal built an AI model that reads EEG brain signals to flag seizures in ICU patients, and in a pilot it caught 9 of 10 seizures with under 10% false positives, though it is still awaiting FDA clearance.

How nurses are using AI: the data at a glance — The AHA's market scan cites a nurse survey from Incredible Health showing AI use jumped from 15% to 44% of nurses in a year, and that nurses who helped choose their tools trust them far more (74%) than those who weren't consulted (38%).

Health Sense by Cigna Healthcare: how it works — Cigna Healthcare says its new Health Sense model uses AI to connect a customer's health, benefits and cost information inside existing tools such as myCigna and flag when someone may need more support, with a phased rollout in early 2027 for customers with oncology or cardiometabolic needs, and says it will not diagnose, prescribe or change treatment.

Boston VA workshop tests AI's potential for Veteran care — At a workshop at VA Boston, the first VA medical center to host one, clinicians and researchers tested generative AI tools including a PTSD assessment aid that initially performed comparably to a clinician, though adding more training data made it too rigid and every project is still at the proof-of-concept stage.

Can Artificial Intelligence Learn the Art of Surgery? — The NY Times featured Canada's new national AI strategy, which Prime Minister Mark Carney unveiled at Toronto General Hospital, aims to use AI to ease long waits and bed shortages, and surgeons there are testing an AI overlay that marks safe and unsafe zones to cut during gallbladder removal.

CMS and FDA officials ran a private Slack workspace with AI companies — Started in August 2025 by CMS chief product officer Amy Gleason, the roughly 1,700-member group included Microsoft, Anthropic, OpenAI, Apple, Google and Palantir, and a critic told reporters it resembles a federal advisory committee operating without the public process that comes with one.

New AI czar unveils goals, members of White House task force — Director of National Intelligence Jay Clayton, who effectively becomes the administration's AI czar, will chair a "Super Intelligence Force" with 120 days to report on AI risks, opportunities and the federal government's responsibility, with industry remaining the primary vehicle for managing risk.

Ted Cruz and Maria Cantwell's big college sports win won't make an AI breakthrough easier — Sens. Cruz and Cantwell each expressed optimism that their bipartisan college sports deal could carry over to AI, but Politico reports few lawmakers, aides or experts expect it to, given Cruz's stalled promise of AI hearings and markups and the distrust between the two.

Cantwell outlines comprehensive governance framework for safe and secure frontier AI — Sen. Maria Cantwell, the top Democrat on the Senate Commerce Committee, released a frontier AI framework that calls for federal safety standards and independent audits before release, a right to challenge AI-influenced health care decisions, and a warning that AI-assisted prior authorization, such as CMS's WISeR model, risks delaying care.

OpenAI's approach to EU text provenance rules — OpenAI will add invisible watermarks to eligible ChatGPT and Codex text in the EU over the coming weeks to meet the AI Act's labeling requirement, though in its own tests replacing 25% of the words in a passage cut detection from about 92% to 17%.

UK unveils new policy to constantly monitor AI-based devices — The UK government accepted 44 recommendations from a national commission to have its medical device regulator keep monitoring AI devices after approval instead of relying on a single initial review, while in the U.S. the FDA is still working out its approach, having put out an August discussion paper on generative AI-enabled devices.

Gottlieb: Congress should have FDA certify health systems, not each AI tool, as it did for lab tests — In a JAMA Health Forum viewpoint, Scott Gottlieb of the American Enterprise Institute proposes an LDT-style pathway in which FDA certifies hospitals to build and run their own clinical AI, arguing the current gap leaves a two-tier market where academic centers deploy freely while commercial developers pay for FDA review.

HHS announces $20M health tech and AI prize challenges — HHS's Agency for Healthcare Research and Quality launched two prize competitions worth about $20 million, including up to $7.95 million to build an AI system that improves how the agency evaluates medical evidence, with the first phase opening January 11, 2027.

Tech trade group urges states to apply existing laws before writing new AI rules — The Connected Health Initiative warns that a patchwork of state laws will hurt small and midsize health tech companies and slow patient access, while backing human review for higher-risk uses and payment modernization.

The risk of unchecked AI agents — A survey for security vendor Imprivata found 72% of health industry leaders say AI tools or agents are being deployed without formal IT approval, and security experts warn that agents with broad access to health records and clinical systems create new safety and privacy risks.

Healthcare privacy leaders look to AI to improve monitoring and data breach investigations — Healthcare privacy teams are using AI to automate how they identify, prioritize and investigate potential privacy violations, but speakers at a Bluesight-hosted symposium said it creates new risks that organizations must govern with human oversight and accountability.

ChatGPT for Teens keeps teens talking, even during mental health crises — Common Sense Media rated ChatGPT for Teens an "unacceptable risk" after finding it kept encouraging teens to keep chatting even in crisis scenarios, while OpenAI disputes the testing and says it largely ran before parental controls were fully active.

AI safety group outlines demands for legislation, warns lawmakers — A coalition of nearly 40 groups led by Guardrails Action, including the American Federation of Teachers, the National Education Association, Public Citizen and the United Auto Workers, told Congress it will actively oppose any AI legislation that lacks enforceable, transparent government oversight not controlled by major tech firms.

The Age of Artificial Intelligence: 73% concerned about potential threat to human survival — A Quinnipiac University poll of 1,202 U.S. adults found 73% are concerned that future AI could threaten human survival, and 86% support independent safety standards for AI companies even if they slow development.

AI is uncovering healthcare security flaws — Epic is patching flaws found by testing its software with an AI model, and security firm Wiz and Google each reported finding serious hospital-software vulnerabilities with AI tools, prompting Epic's security chief to urge health systems to prepare for faster patching.

Expanding the Cyber Verification Program — Anthropic is expanding its Cyber Verification Program into three access tiers that give vetted security teams, including operators of critical infrastructure such as regional hospitals, fewer blocks on its models' cyber capabilities, while its generally available models keep conservative safeguards.

Fired OpenAI Researchers Ask Company to Preserve Visibility Into AI Reasoning — Three safety researchers OpenAI fired for allegedly mishandling confidential information wrote to the company's board urging it to work with outside safety auditors and not to pursue developments that reduce the ability to monitor AI models' reasoning, while OpenAI says the firings weren't about raising safety concerns and that it agrees with the letter's recommendations.

Detecting forever chemicals with yeast and AI — Columbia and City College researchers are using AI to help design a dip-stick water test, built on engineered yeast that glows when it touches PFAS "forever chemicals," that they say would give water systems instant results at a fraction of the cost of sending samples to a lab.

Making global public health more proactive with Google Earth AI — Google says its Earth AI tools, which combine satellite imagery, mobility data and AI models, helped the World Health Organization's Africa office map 48 exposed settlements and about 45,500 at-risk people near Ebola-affected mining areas in the Democratic Republic of Congo in minutes instead of weeks, and that partner studies show better forecasts of cholera and dengue outbreaks.

Hospitals push back against claims of using AI to drive up costs — The Blue Cross Blue Shield Association says hospitals' AI-assisted coding added about $942 million in costs to Blue plans over two years, while the American Hospital Association says the report never checked medical records or identified which claims used AI.

Zuckerberg's Biohub teams with Google in push to map cells — Biohub, working with the Department of Energy, NIH, Google DeepMind, Isomorphic Labs and Meta, announced a roughly $1.8 billion effort to produce the data for an AI "virtual cell" model that would let scientists run experiments on a computer before testing them in the lab, with commercial partners getting a year of exclusive access to the data.

University study examines AI use in health questions — A University of Nevada, Reno analysis of more than 1,600 real conversations with generative AI found people mostly use it for health education, often seeking clarification or reassurance rather than a diagnosis, and that chatbots frequently urged users to seek medical care for serious concerns.

Delaware lands a $21.5M NSF AI institute for mobility and rehabilitation care — The University of Delaware and three partner institutions (Delaware State, Penn and Princeton) will receive $21.5 million over five years from the National Science Foundation to build AI that analyzes movement and medical records for people with stroke and Parkinson's disease, making it the 30th NSF AI institute.

AI's Real Gift to Science — Anthropic said Claude found a CRISPR-like enzyme system in under 24 hours, but researchers including Eli Lilly's CEO called the result incremental, and gene-editing pioneer Fyodor Urnov told The Atlantic the jury is out on whether it is an enzyme at all, so the article argues the real value is AI sorting biological data too large for scientists to analyze themselves.

Still waters, rapid currents: Early labor market transformation under generative AI — A survey of 25,000 Danish workers in 11 AI-exposed occupations, linked to administrative records through December 2024, found no effect of chatbot use on earnings or hours (ruling out effects above 2%) and no sign that firms adopting chatbots are driving the drop in early-career jobs, though adopters were more likely to have switched into better-paying occupations.

Google's AMIE chatbot passes its first clinical trial at Beth Israel Deaconess — In a Lancet study of nearly 100 urgent care patients who talked with the bot before seeing a doctor, physicians said it was helpful in about 75% of cases and may have changed how they handled more than half, with one hallucination (a misstated date) and no conversations stopped for safety. Former FDA commissioner Robert Califf, who was not involved, called it the technology of the future but cautioned that no one should assume it will reduce doctors' workloads.

Who will manage healthcare's AI agent explosion? — Executives from LRVHealth and CVS Health argue that health systems should govern AI agents like non-human employees, with each agent getting its own credentials, task-scoped access, an expiration date and a named human sponsor.

Tying AI adoption to executive pay gains traction — UnitedHealth Group and Elevance Health now tie executive pay to AI adoption, though only about 2% of public companies across industries do, and one compensation consultant says boards push back when the metrics aren't measurable.

The real opportunity for agentic AI in health insurance — An Infosys-sponsored piece argues that insurers' automation is siloed across call centers, prior authorization and claims, and that AI agents coordinating across those handoffs, with humans keeping clinical and coverage decisions, is the missing layer.

How health systems are managing AI token costs — Advocate Health, Hartford HealthCare and Vanderbilt University Medical Center are keeping token costs in check with small pilots that carry a cost ceiling, a centralized platform for chart abstraction, smaller and cheaper models, and IT staff assigned to watch usage dashboards.

AI is making verification the bottleneck for companies — Harvard Business Review's Christian Catalini argues that as AI makes producing work cheap, checking whether it's right becomes the hard part, and that organizations should keep experts reviewing AI output and keep the record of their corrections in-house rather than handing it to AI vendors.

New AI procurement guide focuses on environmental impacts — Health Care Without Harm released a procurement guide, built with health system and payer partners, that gives buyers 10 questions to ask AI vendors about emissions, energy and water use, and lets vendors earn a quarter of the first-year score just by committing to supply that data within 12 months.

A health insurance agency's journey from pilot to full-time voice AI agent — eHealth, a Medicare insurance agency, moved its AI voice agent from an after-hours pilot to prescreening every open-enrollment call, and Regal, the vendor behind it, reports 77% top satisfaction ratings and a 27% higher purchase rate than with its earlier outsourced human screeners.

Is AI a Band-Aid for healthcare's historic problems? — Craig Samitt, a former Blue Cross and Blue Shield of Minnesota CEO, tells Healthcare Brew that most health plans use AI to speed up existing workflows and should instead redesign the processes themselves, for example with AI that gives providers clinical guidance upfront to cut prior authorization disputes.

Health systems are leaning on outside money for AI, and upkeep is the surprise cost — Mayo's chief AI implementation officer says maintenance costs are "far higher" than expected, while grants such as ARPA-H's $62.7M ADVOCATE awards and a $100M gift to UT Health Austin are filling the gap.

👉 Oct. 14 — AI Day 2026 — Hybrid, Georgia Room, NYC; free virtual pass available

👉 Oct. 14–16 — AIxPH 2026: 1st Annual Conference on Artificial Intelligence and Public Health — Baltimore, MD

👉 Oct. 15, 1:00 PM ET — How specialty practices are using AI to turn patient access into capacity and growth — Webinar, Becker's Hospital Review

👉 Oct. 19 — Comments due on FDA's discussion paper on regulating generative AI-enabled medical devices — Public comment deadline via Regulations.gov (docket FDA-2026-N-7874)

👉 Oct. 22–23 — HIMSS AI in Healthcare Forum — San Diego, CA

👉 Tuesday, Nov 17, 1:00 PM ET — Building the foundation for clinical AI: Lessons from Stanford Medicine & The Hospital for Sick Children — Webinar (virtual)

👉 Dec. 8–9 — Stanford AI+HEALTH 2026 — Virtual

👉 Dec. 16, 11:00 AM ET — The Next Phase of AI Adoption: Governance, Ethics, and Accountability — Virtual