The Bandwidth: Decoding Health + AI

🚨 Spoiler alert: if you're watching or planning to watch The Pitt, skip the next paragraph.

This spring I binged The Pitt, the HBO medical drama built around a single shift in a Pittsburgh emergency department. In one of the season's tensest sequences, the hospital's electronic medical records systems go down. What follows is instructive: the staff scrambles back to paper, analog processes, and human-to-human coordination. The unit doesn't collapse, but it strains badly, in ways that reveal how much institutional muscle memory has quietly atrophied. The manual fallback exists in theory. In practice, executing it is harder than anyone expected.

The Outage Nobody Planned For

I thought about that scene this past week when Claude, the AI platform I use daily, experienced a near-complete outage. Elevated errors hit claude.ai, the API, Claude Code, and related tools for roughly 85 minutes before the issue was resolved. A minor incident by most measures. But one detail stood out: Claude for Government, which runs on isolated, FedRAMP-authorized infrastructure, stayed online the entire time. The commercial tier went down. The federal tier didn't.

Most health systems are on the commercial side of that line.

That's worth pausing on. The government tier held not because it got lucky, but because it was designed differently — isolated from the commercial pool that strained under load. Dedicated infrastructure, tighter controls, higher uptime commitments. The gap between the two tiers over the past 90 days is nearly 19 hours of downtime versus 90 minutes.

But the more important question isn't about uptime statistics. It's about what happens to an organization when the AI goes down, especially one that has been redesigning its workflows around AI's presence.

The Vanishing Fallback

Abbey Hernandez, Business Resiliency Director at JPMorgan Chase, wrote about this problem in a piece published earlier this month. She calls it the “vanishing fallback”. It doesn't happen all at once. AI is deployed with human oversight intact. Confidence grows. Human involvement lightens. The people who knew how to run the process manually leave, or their roles are restructured, or their replacements are hired to oversee the AI rather than to understand what it replaced. Training programs stop teaching skills that the system has taken over.

The institutional knowledge quietly expires. Then the AI fails.

The recovery procedure still exists, Hernandez notes. The human infrastructure it depended on may not. It's a workforce planning failure that technology made invisible until it mattered.

A Healthcare Problem in Particular

In healthcare, that argument lands with particular weight. When an EHR goes down, hospitals have downtime procedures. Staff remembers paper. There's a generation of clinicians who ran these processes before the systems arrived. But perhaps not for much longer.

AI is moving into clinical workflows differently — faster, and in some cases before a new generation of staff ever saw the manual version. Cadence is automating medication titration for more than 100,000 chronic disease patients (as noted below in the Field Signals). Ambient scribes are recording patient visits. Clinical decision support is nudging diagnoses across specialties.

What is the manual fallback when those systems go down? The question isn't hypothetical; it's operational. And most organizations haven't asked it in a systematic way.

The Harder Layer

This week, the Five Eyes intelligence alliance issued a rare joint warning that frontier AI capable of crippling governments and businesses is close — "the timeline is not years, it is months," the bulletin read.

Healthcare is critical infrastructure. It is not currently being planned for as such in AI deployment decisions. The Five Eyes bulletin put it plainly: boards and executives should ensure cyber resilience is in place and works under pressure. Having controls is not enough. Leaders must be confident those controls will perform during a real incident.

That's a standard most health organizations may not yet meet for their AI-enabled workflows, because most haven't tested it.

The Question That Comes Next

The Pitt gave us a dramatic version of what happens when the system goes down and the manual fallback is harder to execute than expected. The writers put it on screen because it's believable.

Health system leaders are making decisions right now about how deeply to embed AI into clinical and administrative workflows. Those decisions deserve a parallel question: what does this organization look like on the day this system isn't available — whether because of an outage, an adversarial attack, or a government directive that comes without warning — as when export controls grounded Anthropic's most capable models earlier this month?

That's not an argument against adoption. It's an argument for building the resilience that makes adoption sustainable.

A sweeping new AI to detect heart conditions is coming to OpenEvidence — Pathway Labs received FDA clearance for EchoNext, an AI that screens electrocardiograms for six forms of structural heart disease and will be available to hundreds of thousands of clinicians through the OpenEvidence platform. Independent cardiologists welcome the ambition but note the technology has yet to demonstrate in a randomized trial that it improves hard patient outcomes — and that roughly one in four people without structural heart disease currently receive a false alert.

Cadence raises $100 million to automate chronic disease care with regulated AI — Cadence, which remotely monitors over 100,000 patients with hypertension, diabetes, and heart failure, raised $100 million at a $1.23 billion valuation and is now building AI agents to automate portions of its clinical work — including an FDA-regulated product that would let an algorithm adjust blood pressure medications within parameters a physician pre-authorizes. Health system partners and independent experts are watching closely, with one Brown researcher describing the company as caught between a high-touch clinical model its partners trust and a push toward automation that payers and CMS are actively encouraging.

Anthropic says Claude may want to see your ID — Anthropic updated its privacy policy to allow identity verification — including government-issued ID uploads and biometric data collection — for a subset of users whose accounts are flagged rather than outright banned, with the change taking effect July 8. The move comes as the company remains in an unresolved standoff with the Trump administration, which designated Anthropic a supply chain risk after disputes over the government's access to its AI tools.

OpenAI fires up “Jalapeño,” its first homegrown AI chip — OpenAI announced it has begun testing its first custom-designed chip, built with Broadcom for inference workloads, with commercial deployment at Microsoft and other partners expected by year-end. The move puts OpenAI alongside Google, Amazon, and Microsoft in reducing its dependence on Nvidia — a shift with implications for the cost and availability of compute across the entire AI industry.

1 in 3 Americans use chatbots for health advice. These 6 patients explain why. — Nearly 1 in 3 Americans now turn to AI chatbots for health information, driven by cost and access gaps, according to a recent KFF survey. The reporting pairs that demand with documented risks: one independent evaluation found a chatbot failed to advise an emergency-room visit in more than half of simulated cases of impending respiratory failure or serious diabetic complications, and researchers warn the tools' confident tone can make wrong answers more persuasive than correct human advice.

Health systems must shift AI use from automation to transformation — At the HIMSS AI in Healthcare Forum in Boston, CIOs from Stanford Healthcare and UMass Memorial Health argued that the field has moved past the question of whether to use AI and into harder questions about where, how, and at what cost — including whether human-in-the-loop review is realistic when clinicians stop checking outputs after the first few times they look good. Both flagged a growing digital divide: resource-rich systems can build, experiment, and monitor AI, while safety-net hospitals and community providers increasingly cannot.

Patients prefer healthcare providers' AI agents to public chatbots, with human oversight non-negotiable, survey finds — A Salesforce survey of more than 3,200 global patients found that patients are three times more likely to trust an AI agent embedded in a clinical system than a public chatbot, and that 89-90% say a clear “escalate to human” option is essential for trusting AI administrative or medical support.

FDA gives generative AI in radiology two breakthrough designation nods — The FDA granted breakthrough device designation to two generative AI tools — from Cognita and Aidoc — that can process chest X-rays and draft full radiology reports for radiologist review, a capability that goes well beyond existing AI tools that flag individual findings. A Stanford radiologist who has evaluated more than ten such models finds them weak at measuring and locating findings precisely, prone to hallucination, and raising a harder question: whether accepting AI-generated reports risks diluting radiologist expertise over time.

States are embracing AI to help manage safety-net programs — States are increasingly using AI to handle Medicaid, SNAP, and unemployment caseloads — work that's about to grow under new federal work requirements and more frequent eligibility recertifications — but advocates warn there's little evidence of safeguards. A successful lawsuit over Arkansas's Medicaid algorithm showed how a poorly understood system cut benefit hours for thousands, with officials unable to explain how it worked for more than a year.

J. D. Vance's AI Doctrine — The Atlantic profiles the Vice President's emerging AI framework, which blends Silicon Valley anti-regulation instincts with MAGA worker-protection populism: skeptical of federal rules, wary of letting a few large firms dominate the field, and resistant to ceding the lead to Europe or California. On national security he draws a firmer line, arguing that “decisions over life and death must be made by humans and not machines.”

Medicare's AI Push Snarls Patients and Doctors in Errors and Delays — KFF Health News reports that WISeR, the Trump administration's AI-powered Medicare prior authorization pilot now running in six states, has produced a cascade of errors, payment backlogs, and delays — including denials that doctors attribute to AI hallucinations that garbled or invented clinical information. Providers in the pilot states describe the rollout as “horrendous,” with some patients rerouted to more expensive care while waiting weeks for approvals that Medicare promised within 72 hours.

States surge ahead on healthcare AI regulation — With Congress largely inactive on AI and the Trump administration more focused on deploying the technology than regulating it, states have moved to fill the void — all but a handful now have some law on the books governing AI in healthcare, most targeting how insurers use it in prior authorization decisions. The resulting patchwork has insurers calling for a federal standard that, by most accounts, isn't coming anytime soon.

Bipartisan Senate duo to unveil bill protecting kids from AI chatbots — Senators John Curtis (R-Utah) and Adam Schiff (D-Calif.) introduced the SAFE KIDS Act, which would require rigorous risk assessments before chatbots go public, mandate independent child safety audits, ban targeted advertising to minors, and compel platforms to surface crisis resources when a child may be at risk. The bill is angling for consideration at an upcoming Senate Commerce Committee markup alongside Cruz's existing CHATBOT Act, signaling rare bipartisan momentum on AI and child safety.

Health Accreditors Advance AI Programs With House Appropriators' Support — URAC and the Joint Commission are both rolling out voluntary AI accreditation programs for healthcare organizations, filling a governance gap that federal legislation hasn't closed — and House appropriators signaled support in a recent HHS funding bill report, encouraging HHS to leverage accreditation as a mechanism for responsible AI deployment. The programs are voluntary for now, but given the Joint Commission's role as a major hospital accreditor, what it certifies could eventually shape Medicare conditions of participation.

Trump administration asks OpenAI to limit next model release — The White House asked OpenAI to restrict the rollout of its next model, GPT-5.6, to a small set of government-approved partners before any wider release — marking the first time the U.S. government has preemptively asked an American AI company to limit a model launch on security grounds. The move follows the government's earlier export controls on Anthropic's frontier models and reflects a broader administration posture that models of sufficient capability require government testing and approval before public release.

The Last Thing US Healthcare Needs Is an AI Takeover — Two physicians and Medicaid advocates writing in The Nation argue that AI adoption in healthcare is being driven by market pressure and investor enthusiasm rather than evidence of patient benefit, and that the real risk is further financialization of a system already failing its most vulnerable.

Anthropic, OpenAI join $500 million AI jobs push — Former Commerce Secretary Gina Raimondo and former Indiana Gov. Eric Holcomb are launching Raise Us, a $500 million workforce initiative co-funded by Anthropic, OpenAI's Foundation, Amazon, Microsoft, and others, aimed at helping states and employers prepare workers for AI-driven labor market disruption.

Employees Aren't Questioning AI Advice Enough — A Harvard Business School study put 2,512 participants in the role of loan officers reviewing real $10,000 loan applications, where an AI flagged each borrower as high or low default risk — and found that while 80% wanted the AI's risk assessment, fewer than half chose to view the explanation for how it was reached. Participants were even less likely to look at those explanations when their bonuses depended on loan repayment or when they were told the explanation might reveal that race or gender had influenced the AI's decision, leading the researcher to conclude that transparency alone isn't enough: organizations need incentives and oversight that require people to actually engage with AI reasoning, not just have access to it.

Human-AI Collaboration in Healthcare: A Scoping Review — Researchers reviewed 140 studies on what happens when clinicians and AI tools work a task together, rather than measuring the AI on its own. Teams generally did better, but almost all the strong evidence comes from one narrow area — reading scans and images — and most studies measured quick wins like speed and accuracy rather than whether patients actually fared better. The authors' bottom line: keeping a human “in the loop” isn't enough on its own, because the AI quietly shapes what clinicians notice and trust, leaving the doctor accountable for a decision the system increasingly steers.

AI wades into a vexing medical mystery: What causes sudden cardiac death? — A study published in Nature used AI to analyze EKG tracings and found that cardiac fibrosis — scar tissue scattered through the heart — is highly prevalent among patients at the greatest risk of sudden cardiac death, a condition that kills upward of 350,000 Americans a year and is preventable with a defibrillator if you can identify the right patients. The model needs further study before clinical use, and an independent electrophysiologist notes it is a “long road to even getting close to the patient,” but the finding opens a new line of inquiry into why current guidelines misidentify so many high-risk cases.

How GPT-5 helped immunologist Derya Unutmaz solve a 3-year-old mystery — OpenAI profiles Jackson Laboratory immunologist Derya Unutmaz, who used GPT-5 Pro to revisit a shelved 2022 experiment on how glucose affects T cell development — the model suggested a mechanistic explanation involving a protein called IL-2 that Unutmaz and his team had missed, and correctly predicted the outcome of a separate unpublished experiment.

Researchers Examine Safety Risks in AI-Generated Translation of Emergency Department Discharge Instructions — University of Colorado Anschutz researchers presented findings at the AMIA Amplify Conference showing that automated metrics used to evaluate AI-translated Spanish discharge instructions frequently missed clinically dangerous errors — including a translation of “take with food” rendered as “do not take with food” — that human clinician review caught. The study argues health systems need a layered evaluation approach rather than relying on automated scores alone, particularly for high-risk content where a mistranslation could alter patient behavior or compromise safety.

EY: we found your biggest AI blind spot. It's called the 'tempo gap' — Two EY consultants argue the real risk in AI adoption isn't slow uptake but “tempo” — systems now interpret intent and push interactions forward faster than people can process them, including patients whose sensitive data auto-populates before they grasp how it'll be used. Their prescription is “intentional friction”: deliberate pauses built into high-stakes workflows so human judgment and trust can catch up.

Why Healthcare AI Governance Breaks Down After Deployment — Two health IT consultants make the case that most AI governance is front-loaded — rigorous at go-live, then quietly outdated as models drift, vendors push updates, and clinicians adapt workflows in ways no policy anticipated. Their practical prescription: a named owner for every tool, a monitoring baseline reviewed on a standing schedule, and escalation paths that clinicians can actually find and use.

Introducing Claude Tag — Anthropic launched Claude Tag, a Slack-native feature that lets teams tag @Claude as a shared team member in channels — it builds context over time, works asynchronously on delegated tasks, and can take initiative by surfacing relevant updates without being asked. Available today in beta for Enterprise and Team customers, the announcement notes that 65% of Anthropic's own product team's code is now generated through its internal version of the tool.

The Human Impact of AI: Why Redesigning How We Work Matters as Much as the Technology Itself — IBM's Chief Talent Officer writes in the company's newsroom about a new IBM Institute for Business Value study finding that executives report AI-driven role change at twice the rate employees experience it in their own jobs — and that 43% of employees say their employer provides no AI training, even as 81% of executives believe workers are being rewarded for building AI skills. The practical prescription: treat major AI workflow changes as people adoption challenges first, update performance systems to reward judgment and responsible challenge, and build in structured opportunities for employees to practice deciding when to trust AI and when to push back.

Copilot in Excel: Built for the era of Frontier Finance — Microsoft announced new Copilot in Excel capabilities aimed at finance professionals, including custom "skills" that guide the AI through repeatable workflows like variance analysis, forecast updates, and board reporting models using your organization's own process definitions. The update also adds financial data connectors from providers including FactSet, Morningstar, PitchBook, and S&P Global, and a "Plan with Copilot" feature that shows what it intends to change before acting — addressing the traceability demands that financial work requires.

👉 July 5, 2026 — HHS comment deadline: RFI on addiction and mental health

👉 Jul. 7–10 — AI for Good Global Summit 2026, Geneva, Switzerland

👉 Aug. 4–6 — Ai4, Las Vegas

👉 Oct. 13–16, 2026 — AIxPH 2026: 1st Annual Conference on Artificial Intelligence and Public Health, Baltimore, MD

👉 Oct. 22–23, 2026 — HIMSS AI in Healthcare Forum San Diego + AI Leadership Summit — San Diego, CA