The Bandwidth: Decoding Health + AI

Someone recently put the 1973 book, Small is Beautiful, on my radar. Not having read it yet (though it’s now on the top of my stack), my understanding of the central argument is this: modern industrialism has developed an almost religious faith in scale. Bigger plants, bigger companies, bigger outputs. Growth as an end in itself. The author, E.F. Schumacher, pushed back. Smaller-scale institutions, he argued, were more adaptable, more humane, and ultimately more durable than the giants that dominated the economic landscape.

For fifty years, the companies that came to define the modern economy were the ones that grew largest, moved fastest, and built the deepest moats. Size was the strategy.

AI may be changing that calculus.

The New Economics of Small

There's a growing case that AI functions as a force multiplier for small teams in ways that fundamentally disrupt the scale advantage. Capabilities that once required armies of analysts, developers, and administrators can now be handled by a lean team with the right tools and the willingness to deploy them. Doing more with less isn't just possible. It's becoming the new economics of how competitive work gets done.

Bo Burlingham made a version of this argument before AI entered the picture. His book Small Giants (one I have read) profiled companies that chose depth over scale — excellence over expansion, community over conquest. They didn't win by being biggest. They won by being best at what they did, with fewer layers between intention and execution. AI may help them gain a competitive advantage moving forward.

A forthcoming piece in Harvard Business Review puts numbers to the shift: AI-native startups are reaching early funding milestones with 80% less capital and 20-40% less time than their predecessors. The conventional product team of six to eight people has compressed to two: a domain expert and an AI engineer. The AI era is giving the small-is-beautiful model real structural backing.

The Bureaucracy Problem

Part of what makes smaller organizations better positioned for this moment is less about technology and more about decision velocity. Large enterprise companies face a compounding challenge. The moat they once had — scale, resources, the ability to do things smaller competitors simply couldn't — is rapidly narrowing. And they're trying to close that gap while navigating the organizational infrastructure that made them large in the first place.

Changing an inventory system at a large organization can mean five levels of approval, a cross-disciplinary task force, weeks of meetings, and careful consideration of how the change interacts. For healthcare organizations, it gets even more complicated given the stakes and regulatory complexity.

At a small company, it’s a different calculus. Decisions can happen in minutes.

That said, mentality matters more than manifest destiny here. Andrew McAfee's The Geek Way — a book I just finished — makes a case that the companies winning in the technology era aren't necessarily the smallest — they're the ones that preserve the operating philosophy of a small team regardless of their size. Rapid iteration, scrutiny, honest feedback loops, willingness to abandon what isn't working. Scale doesn't have to mean slow.

The Model Architecture Argument

There's a third dimension to this that doesn't get enough attention outside technical circles: the AI models themselves may be following the same logic.

Jeffrey Li, CEO of Liquid AI — an MIT spinout — makes a provocative case on the AI Curious podcast. The dominant assumption in AI has been that bigger models are better models. More parameters, more compute, more capability. That assumption has driven the race to build the largest possible systems running in the largest possible data centers.

Li's thesis runs in the opposite direction. Ninety-nine percent of the processors on earth live outside data centers — in phones, laptops, even AirPods. And most tasks that those devices need to handle don't require a model that can solve cancer. A law firm reviewing contracts needs a model that's excellent at tort law. Tailored, in Li's words, like a suit.

The framework he reaches for maps onto something familiar: Daniel Kahneman's Type 1 and Type 2 thinking. Type 2 is slow, deliberate, energy-intensive — the kind of thinking you need for deep research and complex reasoning. Type 1 is fast, pattern-based, efficient. Humans are good at both, and we're good at knowing which one the moment calls for. Li's argument is that mature AI systems will work the same way. And that as AI adoption scales, the share of workloads handled by fast, specialized, edge-deployed models will actually increase.

The Through-Line

Schumacher wasn't arguing that small was always better. Or so I'm told. He was arguing that scale had become an assumption rather than a strategy, that the default toward bigness was crowding out something important.

That's the argument worth revisiting now. In an AI era, what is the right size for the thing you're trying to do? The answer may be smaller than you think.

A New Era for AI Search — At Google I/O, Google announced a sweeping overhaul of Search: AI Mode has surpassed one billion monthly users just one year after launch, and the company is now introducing autonomous Search agents that operate in the background 24/7, a redesigned AI-powered search box described as the biggest upgrade in 25 years, and agentic coding capabilities that build custom dashboards and mini apps on the fly. Personal Intelligence — which connects Search to Gmail, Google Photos, and soon Calendar — is expanding to nearly 200 countries across 98 languages, free of charge.

AI Layoffs Aren't Delivering the Stock Boost CEOs Expected — CNBC tracked 23 S&P 500 companies that announced AI-linked layoffs and found 56% saw stock declines averaging 25%. A separate Gartner survey of 350 executives found no return advantage for companies that cut headcount for AI reasons — while the highest performers used AI to amplify workers, not eliminate them.

How AI Is Turning UnitedHealth, CVS, and Elevance Into Software Companies — The country's largest payers are commercializing their internal AI investments: Optum is directing one-third of UnitedHealth's $1.5 billion AI spend toward becoming an "AI-first software and services firm," while CVS and Elevance are rolling out their own external-facing platforms for clinical and administrative workflows. The defining tension is whether these products serve providers and members — or simply add another revenue layer to an already complicated system, particularly as AI-driven prior authorization faces intensifying state legislative scrutiny.

UnitedHealth Tracks Workers' AI Use in Push to Transform Company — UnitedHealth is monitoring whether some Optum employees perform at least one AI query per day, using an internal dashboard that tracks usage, training progress, and "adoption gaps," as it deploys $1.5 billion in AI investment toward what executives call a fundamental operational transformation. The company reports its AI tools have already avoided more than 15 million calls and adjudicated hundreds of millions of claims — while its latest annual report simultaneously flags that growing AI use "presents legal, regulatory and business risks."

AI-Driven Change and the Next Big Thing — A Strategy&/PwC survey of 50 German CIOs finds Europe falling further behind China and the US on AI-driven revenue gains and cost reduction, with only about one-third of organizations having structurally embedded AI into core workflows despite broad executive expectation that AI will dominate strategic decision-making within five years. The identified bottleneck isn't model capability — it's data readiness, with just 22% of organizations having the foundation required to scale advanced AI use cases.

Andrej Karpathy Joins Anthropic's Pretraining Team — Andrej Karpathy — an OpenAI founding member and former head of AI at Tesla — has joined Anthropic's pretraining team, the division responsible for the large-scale training runs that underpin Claude's core capabilities. Karpathy is the second high-profile OpenAI co-founder to join Anthropic, following John Schulman's move in 2024, and his hire signals continued talent consolidation at the frontier model level.

Doximity Inks Partnerships with Aledade, Photon as It Ramps Up AI Spending in 2026 — Doximity CEO Jeff Tangney called 2026 the company's "AI investment year," announcing new partnerships with Aledade and Photon Health to embed its ambient notetaking and clinical AI tools into primary care workflows, while reporting that nearly 800,000 quarterly active prescribers now use its platform — up 30% year over year. Wall Street was skeptical: shares dropped 24% after earnings as guidance came in well below analyst expectations, with management flagging pharma marketing headwinds and near-term margin pressure from increased AI spend.

Jury Rejects Musk's Claims Against OpenAI — A jury unanimously sided with OpenAI and Sam Altman in Elon Musk's lawsuit, rejecting his claims of unjust enrichment and breach of charitable trust on the grounds that the statute of limitations had expired — clearing the way for an OpenAI IPO. Musk has pledged to appeal, but legal experts noted that overturning a statute-of-limitations finding is an unusually difficult bar.

BMS Taps Anthropic's Claude for Enterprise-Wide AI Adoption to Speed R&D, Global Workflows — Bristol Myers Squibb has announced an enterprise-wide deal to deploy Anthropic's Claude as its "shared intelligence platform" across 30,000+ employees, targeting agentic AI integration across drug R&D, manufacturing, and commercial operations — with a stated goal of halving the time from target selection to lead molecule identification. The deal is the latest in a rapidly intensifying pharma AI arms race: Merck went with Google/Gemini in a potential $1 billion deal, Novo Nordisk picked OpenAI, and Eli Lilly and Roche have aligned with Nvidia.

New AI Models Raise Alarms Among Health Care Leaders — Advanced AI models capable of identifying and exploiting software vulnerabilities at unprecedented speed — including Anthropic's Mythos, deemed too dangerous for public release — have put healthcare cybersecurity leaders on high alert, with the AHA's deputy cybersecurity adviser warning that these capabilities are "accelerating the attack cycle time frame at an alarming rate." Healthcare organizations are calling for stronger federal coordination at a moment when CISA has been weakened by staffing and budget cuts, while rural and under-resourced hospitals — already operating below what experts call the "cybersecurity poverty line" — face the sharpest exposure.

Two Hours That Changed AI — In a compressed two-hour window Wednesday, the AI industry produced a defining news cycle: Anthropic reported it's on track for its first profitable quarter with $10.9 billion in Q2 revenue — two years ahead of internal projections — while Nvidia posted $81.6 billion in quarterly revenue and Anthropic signed a $1.25 billion per month compute deal with SpaceX through 2029. The backdrop: Meta laid off 8,000 workers the same day, and new polling shows 70% of Americans fear AI is advancing too quickly.

AI Sepsis Tools Are Gaining Traction—and Raising Expectations — AI-powered sepsis detection is accumulating a real clinical track record: Duke University Health System's Sepsis Watch has achieved a 27% decrease in sepsis-related deaths since 2018, UC San Diego's Composer showed a 17% mortality reduction in a 2024 study, and Bayesian Health just received FDA 501(k) clearance for its EHR-integrated detection tool. The next challenge clinicians identify is tools that can explain their reasoning and distinguish sepsis from other causes of deterioration — not just fire an alert.

AI Chatbots for Health: How to Use Them Safely and Effectively — The AMA has published a patient-facing guide — including five prompts — for using AI chatbots to get health information safely, while simultaneously calling on federal lawmakers to establish stronger safeguards ensuring these tools complement rather than replace physician guidance. The resource reflects a growing institutional reckoning with patient chatbot use: AI tools are increasingly where people turn first, and the AMA is trying to shape how that happens before regulators do.

Healthcare's Real AI Breakthrough May Be Getting Proven Care to More Patients — Speaking at a Mount Sinai/NYAS conference, former NYC Health Commissioner Dr. Dave Chokshi reframed AI's highest-value opportunity in healthcare: not discovering new treatments, but closing what he calls the "discovery-delivery gap" — the stubborn distance between proven interventions and the patients who never receive them. His examples are concrete: fewer than a third of diagnosed hepatitis C patients receive curative antivirals, half of patients with hypertension aren't controlled, and 100 million Americans lack regular access to a doctor — problems where AI-assisted case finding and care navigation may do more good than the next diagnostic breakthrough.

Expert Perspectives on US Regulatory Approaches to Large Language Models in Healthcare — In an interview published in npj Digital Medicine, FDA AI advisor Dr. Shantanu Nundy articulates a two-dimensional framework for thinking about which LLM-enabled health tools warrant regulatory scrutiny: clinical severity crossed with tool autonomy — how serious the condition, and how much the tool acts versus informs. He also flags an unsettled frontier: when general-purpose AI models routinely provide health guidance at scale, the FDA's function-based approach may need to extend beyond traditional medical device boundaries.

Trump AI Executive Order Seeks Early Government Access to Advanced Models — The White House had been preparing to release an executive order this week establishing a voluntary framework requiring AI labs to share frontier models with the government at least 90 days before public release — a response to alarm over cyber-capable models like Anthropic's Mythos and OpenAI's GPT-5.5-Cyber. The signing, originally expected Thursday, was postponed, per Politico; the reason and a new timeline were not immediately provided.

Congressional Democrats Try to Force a Vote to End Medicare AI Prior Authorization Pilot — Senate and House Democrats invoked the Congressional Review Act to overturn the WISeR pilot — a Medicare experiment using AI to approve or deny care in six states — following a GAO ruling last week that the program should have been submitted to Congress before taking effect in January. Critics have focused on a particularly sharp structural concern: WISeR contractors are paid based on a formula that includes how many procedures they deny, which outside experts say directly incentivizes denial of care.

California Signs First-of-Its-Kind Executive Order to Prepare Workers for AI Disruption — On the same day the Trump administration postponed its AI executive order, Governor Newsom signed a first-in-the-nation order directing California state agencies to study AI's workforce impact and develop policy responses — including potential WARN Act updates, severance standards, worker ownership models, and a 90-day dashboard tracking AI-driven job displacement. The order creates no immediate worker protections; it launches a study and policy development process in a state home to 33 of the world's top 50 AI companies, amid a year that has already seen over 70,000 job losses attributed to AI.

HHS Cracks Down on Years of Unchecked Audit Findings — HHS has launched AERO — the Audit Enforcement and Risk Oversight initiative — deploying AI analytical tools to comb through at least five years of single audit data across all 50 states, identifying states and grantees with persistent unresolved findings, some dating back three to five years. The department has sent formal notices to all 50 governors warning that chronic audit noncompliance will now carry real consequences, including payment withholding, award suspension, and debarment proceedings.

MindBench.ai — A new initiative from the Division of Digital Psychiatry at Beth Israel Deaconess Medical Center and NAMI is building an independent evaluation framework for LLMs used in mental health contexts — an area where adoption has outpaced rigorous assessment. The project, led by Dr. John Torous, centers evaluation on what users in distress actually need, including input from people with lived experience of mental illness, rather than narrower clinical accuracy or safety simulation testing.

How Google's Partnership with Consultants Could Derail Enterprise AI Adoption — Writing in Fast Company, Sam Shar — himself a consultant and therefore not a disinterested observer — argues that Google's $750 million fund tying Accenture, Deloitte, and McKinsey to its AI stack creates a structural conflict: firms incentivized to sell AI fast are less likely to prioritize model flexibility, security, or compliance. The same critique applies to OpenAI's Frontier Alliance; if rushed deployments continue producing poor returns, he argues, it damages both the consultancies and the technology they're selling.

Generalization Uncertainty in AI-Enabled Medical Devices: A Safer Way Forward — A paper from the Paragon Health Institute — a market-oriented health policy think tank — argues that demographic diversity in AI medical device training data doesn't guarantee reliable performance for individual outlier patients, and proposes a voluntary tool called Digital Similarity Analysis (DSA) that would compare a patient's medical image against a device's training data before use to flag potential generalization failures. The proposal sidesteps mandatory disclosure requirements in favor of a market-driven approach, which reflects Paragon's policy orientation but doesn't diminish the underlying technical concern it raises.

AI Radiotherapy Planning Tool Clears High Bar in International Trial — The ARCHERY trial, led by UCL and the London School of Hygiene & Tropical Medicine across hospitals in India, South Africa, Jordan, and Malaysia, found that AI-based radiotherapy planning met international quality standards in more than 95% of cervical cancer cases and 85% of prostate cancer cases. The technology cuts planning time from days or weeks to roughly an hour — a meaningful capability gap-closer in low- and middle-income countries where only 10% of patients who need radiotherapy currently receive it.

AI Eats the World — Benedict Evans, May 2026 — Benedict Evans has released his latest AI landscape presentation — a twice-yearly independent attempt to take stock of where the technology actually stands across capital deployment, adoption, and market structure. The deck draws on data from Gallup, Bain, a16z, and others to map the gap between AI investment and real-world deployment — and raises the questions Evans thinks leaders should be asking as the platform shift plays out.

Data Centers and Local Economies in the Age of AI — A new NBER working paper from researchers at the University of Chicago and Yale uses county-level data from 1995–2020 to measure the local economic effects of data center expansion, finding positive effects on employment, wages, adjusted gross income, and business establishments — with construction employment spiking during the build-out phase and data-processing employment growing persistently over time. The tradeoff is real: data centers are also associated with measurably higher local electricity prices and house prices, a policy-relevant tension as counties increasingly compete to attract AI infrastructure investment.

Randomized Trial Finds Conversational AI Reduces Anxiety and Depression in Students — A randomized clinical trial of roughly 1,000 Israeli students, published in JAMA Network Open, found that the Kai.ai conversational platform significantly reduced anxiety and depression compared to a waitlist control — and outperformed human-led group therapy on anxiety specifically, with 58% of participants with clinical anxiety moving into the healthy range.

The Uncritical Adoption of AI in Science Is Alarming — We Urgently Need Guard Rails — A Comment in Nature by researchers from Yale and Princeton synthesizes growing empirical evidence that LLM-assisted scientific papers are narrowing inquiry and declining in quality: an analysis of 41.3 million papers found AI adoption pushes researchers toward known problems rather than new ones, while an audit of nearly 7,000 Organization Science submissions found LLM-assisted papers were less likely to be accepted. The authors raise an underexplored risk: as AI automates entry-level scientific tasks, early-career researchers may never develop the tacit knowledge needed to responsibly supervise AI-driven research.

Are 'AI Co-Scientist' Tools Actually Useful for Scientists? — Following high-profile Nature papers from Google and FutureHouse describing AI systems for autonomous scientific discovery, STAT's Brittany Trang surveyed working scientists and found a persistent gap between the hype and practice: tools still require humans to define the problem, constrain the search space, and verify results — and current models are unreliable at literature synthesis, frequently missing key papers, hallucinating citations, and misrepresenting findings. A deeper structural problem: AI trained on scientific literature inherits its biases, including the systematic absence of negative results, which skews what the tools know about any given research area.

The Hard Road to Healthcare Automation, and Why This Time Is Different — Writing from his experience at Olive AI, Geoffrey Martin argues that first-generation RPA failed in healthcare because it couldn't scale reliability across fragmented workflows and constantly shifting payer systems. His practical guidance for this generation: start with operational strategy and governance before technology, operationalize a few high-impact workflows reliably before expanding, and demand vendor transparency into how automation decisions are made.

Gen AI Could Fix Performance Reviews—or Make Them Even Worse — Most organizations are using AI to produce more polished performance review narratives — which makes weak underlying evidence sound more credible without actually improving it. Boston University professor Chrysanthos Dellarocas argues the better use is directing AI to surface verifiable behavioral evidence: the specific decisions, influence patterns, and pivotal moments embedded in project documents and communications, rather than generating more compelling descriptions of abstract traits.

Why Are Healthcare Companies Trapped in an AI 'Pilot Purgatory'? — Komodo Health co-founders Arif Nathoo and Web Sun diagnose a common failure mode in healthcare AI adoption: companies run proof-of-concept pilots to prove the technology works, but never define what operational change they're trying to achieve — leaving 95% of AI pilots generating zero measurable return, per a 2025 MIT analysis. The fix is conceptually simple: design pilots around changing how a team or business operates, not around demonstrating that the tool functions.

👉 May 29, 1:00–2:00 PM ET — AI in Biomedical Research and Healthcare: Ethical Considerations — In-person, Fred Hutch Cancer Center, Seattle, WA

👉 Jun. 1–5 — Stanford Health AI Week

👉 June 2 — CHAI Webinar: The Cost of Catching Up — Why AI Governance Can't Wait Until Deployment — Virtual

👉 Jun. 8–10 — Fortune Brainstorm Tech, Aspen, CO

👉 Jun. 15–18 — Databricks Data + AI Summit 2026, San Francisco + Virtual

👉 June 24 — CTRC 14th Annual Digital Health Summit: Bridging the Gap — Digital Health and AI for Smarter Care, Stronger Teams, and Sustainable Healthcare — Virtual

👉 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 (Abstract submissions due May 29)

And finally, if you like what you are reading, please share this newsletter with your networks and encourage them to sign up. ✍️ 🆙 And/or, give me a shout out on LinkedIn.

Till next time,

BC