Healthcare Employs 1 in 9 American Workers, Anthropic Just Published a Plan for AI Wiping Out the Back Office & VC Is Quietly Building the Machinery That Decides Who Gets Redeployed
Video Preview
🎧 Part I Podcast free on Spotify.
🎧 Part II Podcast episode for paid subscribers only. Also available on Spotify.
To listen to paid episodes in Apple or Spotify, link your Substack subscription via the show settings on those platforms (instructions inside the Substack app under Subscriptions → Podcast).
Abstract
Anthropic published its Economic Policy Framework in June 2026, a tiered plan for AI-driven labor disruption pegged to 5% unemployment, 10% unemployment, and unprecedented unemployment. Healthcare shows up in it twice: once as a victim, once as the escape hatch.
Healthcare and social assistance employs roughly 22M Americans and has carried US job growth for two-plus years. The sector splits cleanly into an exposed half (admin, RCM, coding, prior auth, scheduling, roughly $500B to $1T in annual spend) and a protected half (hands-on care, where shortages are projected to get worse, not better).
The framework explicitly proposes enhanced benefits for displaced workers who opt into healthcare, elder care, and child care roles, and Tier 3 public investment in human-facing work. Translation: healthcare is the designated labor sink for the rest of the economy.
VC is funding both sides of the trade at once: automation of the back office (ambient scribes, autonomous coding, prior auth AI, agentic RCM) and the redeployment machinery (staffing marketplaces, credentialing rails, training-to-placement programs, licensing reform plays).
Sleeper themes: occupational licensing reform as a federally funded tailwind, services-as-software margin math, the contract labor hangover at health systems, and why Baumol’s cost disease finally has a challenger.
Table of Contents
One industry, two labor markets
What Anthropic actually put on paper
The trillion dollar back office is the kill zone
The hands-on half is protected, sort of
Healthcare as the designated labor sink
Where the venture money is actually pointed
Licensing reform is the sleeper trade
What breaks first
How this probably ends
One industry, two labor markets
Start with the number everyone in healthcare knows and nobody outside it believes: healthcare and social assistance employs somewhere around 22 million people in the US, more than retail, more than manufacturing, more than anything else. For most of 2024 and 2025, healthcare was responsible for a third or more of all net job creation in the country, month after month, while tech shed headcount and white collar hiring went sideways. Strip healthcare out of the payroll reports over the last two years and the labor market story reads very differently. The sector has been the shock absorber for the entire economy, quietly soaking up workers while everyone argued about whether the soft landing was real.
So when Anthropic drops a formal policy framework for AI-driven labor market disruption, healthcare people should read it twice, because the sector sits on both sides of the ledger. Healthcare contains one of the largest concentrations of automatable knowledge work in the economy (the administrative apparatus, which by most estimates runs somewhere between 15% and 30% of the roughly $4.9 trillion in national health expenditure) and simultaneously one of the largest concentrations of work that AI cannot do, at least not in this product cycle (lifting a post-op patient, starting an IV, sitting with a family in the ICU, wiping someone after a stroke). Same industry, same hospital building even, but two completely different labor markets with opposite exposure profiles. The badge readers at a large health system clock in coders and CNAs through the same door, and AI is coming for exactly one of them.
That bifurcation is the whole essay. Everything else, the policy framework, the VC funding flows, the licensing arbitrage, the staffing marketplace economics, hangs off the fact that healthcare admin is a target and healthcare delivery is a destination.
What Anthropic actually put on paper
The framework itself is worth reading in full, but the short version for people who price risk for a living: Anthropic organized its policy recommendations into three tiers calibrated to the unemployment rate. Tier 1 assumes roughly 5% unemployment with elevated churn underneath the headline number. Tier 2 assumes recession-level disruption around 10%. Tier 3 contemplates unemployment exceeding historical peaks while the economy simultaneously posts record output, which is a sentence that would have gotten an economist laughed out of a seminar five years ago and now anchors a policy doc from a frontier lab.
The Tier 1 toolkit is the interesting one for near-term operators: universal pre-distributive capital accounts (building on the federal accounts already seeded at birth, expanded to incumbent workers in exposed occupations, fundable with equity in AI companies), wage insurance for workers who take pay cuts after displacement, retention tax credits for firms that redeploy rather than RIF, employer-connected sectoral training grants, and federal money for states that loosen occupational licensing. Tier 2 centers on expanded UI with automatic triggers, sector-specific transition support, and a basic needs relief program with a detail that should make every healthcare workforce person sit up: enhanced benefit levels for people who opt into roles addressing labor shortages in healthcare, education, child and elder care, infrastructure, and public safety. Tier 3 gets into new tax bases (token taxes, compute levies, digital dividends), sovereign wealth structures, and substantially expanded public investment in human- and community-facing work, with the framework explicitly citing federal health workforce shortage projections as the rationale.
Anthropic also committed $350M of its own money, $200M to an Economic Futures Research Fund for policy trials and $150M to a national fellowship program, and stated outright that it is willing to help fund responses that private firms don’t traditionally finance. Take the sincerity at whatever discount rate seems appropriate. The substance is what matters here, and the substance keeps pointing at the same place. When the framework’s authors went looking for where displaced labor should go, they landed on care work. Twice. By name.
The trillion dollar back office is the kill zone
Now the exposed half. US administrative spend in healthcare is famously hard to pin down because nobody agrees on the denominator, but the credible range runs from roughly $500B (billing and insurance related costs alone) to north of $1T if you count the full apparatus of scheduling, credentialing, eligibility, utilization management, quality reporting, and the army of people whose job is arguing with the army of people on the other side of the claim. McKinsey’s pre-AI estimate put avoidable admin waste around $265B a year. CAQH has tracked the transactional slice for two decades and keeps finding double-digit billions in savings just from moving manual transactions to electronic ones, a transition the industry has somehow stretched across an entire human generation.
The headcount behind that spend is the kill zone. Medical records specialists and coders, somewhere around 190k per BLS, plus the much larger uncounted population doing coding-adjacent work inside RCM vendors and payer ops. Prior auth teams, where AMA surveys consistently show physician practices burning 12-plus hours of staff time per physician per week on roughly 40 PAs. Patient access and scheduling. Claims examiners at payers. Denial management. HIM departments. Release of information. Credentialing committees that take 90 to 120 days to verify what a database query confirms in milliseconds. The honest tally of US workers whose daily output is reading a document, applying a rule, and producing another document probably runs in the low millions within healthcare alone, and that job description is uncomfortably close to the technical definition of what a frontier model does for twenty bucks a month.
The tells are already in the data. The framework opens by citing research showing entry-level workers in AI-exposed occupations seeing weaker employment growth, and healthcare admin is a canonical entry-level white collar on-ramp, especially in secondary metros where the regional health system and the regional payer are two of the top five employers. Autonomous coding vendors are publishing case studies with 85% to 95% of charts coded without human touch in high-volume specialties like radiology and pathology. Ambient documentation has gone from conference demo to standard of care procurement line item in about thirty months. None of this shows up as mass layoffs yet. It shows up as attrition that doesn’t get backfilled, req freezes, and outsourced functions getting re-insourced as software. Which is exactly the churn-under-the-headline pattern Tier 1 of the framework describes.
The hands-on half is protected, sort of
The other half of the building is a different planet. HRSA projects a shortage of 187k FTE physicians by 2037, concentrated brutally in primary care and rural markets. BLS projects roughly 194k RN openings every year through the early 2030s, driven mostly by retirements and an aging patient base rather than sector growth. Home health and personal care aides are projected to add more absolute jobs than any other occupation in America, over 800k this decade, at a median wage around $33k, which tells you everything about why those jobs go unfilled. Demand for hands-on care scales with the 65-plus population, which grows about 3% a year regardless of what GPT-7 can do, because the boomers did not consult the AI timeline before being born between 1946 and 1964.
The protection is real but it deserves an honest stress test, because sort of is doing work in that subtitle. AI does not replace the nurse, but it changes the ratio math. If ambient monitoring, predictive deterioration models, AI-drafted care plans, and documentation automation let one RN safely cover eight med-surg patients instead of five, that is functionally a 35% to 40% capacity expansion without a single new grad, and every CFO staring at a labor line that runs 50% to 60% of total operating expense has already done that arithmetic. Hippocratic AI raised at a $1.64B valuation marketing AI clinical agents at $9 an hour for the phone-based slices of nursing work, post-discharge calls, med adherence checks, pre-op instructions. Those slices were never the core of the job, but they were billable FTE hours, and they are leaving the labor market first. The pattern to expect in clinical roles is not displacement, it is task decomposition: the licensed human keeps the physical and judgment-heavy core while the communicative and documentary shell gets peeled off and sold back to the health system as software. Wages for the core should hold or rise. Total FTE demand per unit of care delivered probably flattens even as absolute demand grows. Both things will be true at once and both sides of the political argument will cherry-pick their half.
Healthcare as the designated labor sink
Here is the part that deserves more attention than it has gotten. Read the framework’s Tier 2 basic needs relief design again: enhanced monthly payments for displaced workers who opt into shortage roles in healthcare and elder care. Then read Tier 3’s call for directing AI-driven surplus toward public investment in human-facing work, justified explicitly by federal health workforce shortage projections. This is a frontier AI lab, in a formal policy document, proposing to use the safety net as a routing mechanism that channels displaced paralegals, claims adjusters, and junior analysts into CNA programs and home care agencies. Healthcare is not just a sector that will experience the AI transition. In the emerging policy consensus, healthcare is the landing zone for everyone else’s transition.
The logic is sound as far as it goes. Care work is the largest pool of unfilled, hard-to-automate, socially necessary labor in the country, and it sits right where displaced workers live, in every county, not just the metros. The friction is also obvious to anyone who has run a healthcare workforce P&L. The jobs pay $33k to $45k at the entry tiers, often below the wage the displaced worker just lost. The work is physically punishing with injury rates that embarrass construction. Turnover in home care runs 60% to 80% annually at many agencies, which means the sector already churns through its entire frontline workforce roughly every 18 months and still can’t fill shifts. Pouring subsidized labor supply into that funnel without fixing the wage structure, the reimbursement rates underneath the wage structure (looking at you, state Medicaid HCBS rates), and the working conditions is a recipe for expensive churn at federal scale. The framework’s wage insurance proposal partially addresses the pay gap, time-limited top-ups for workers who take lower-paying jobs after displacement, and the US evidence on wage insurance is genuinely decent, with studies showing it roughly pays for itself through avoided UI and recovered income tax. But time-limited is the operative phrase, and home care does not become a career when the supplement expires in year two.
So the sink exists, the plumbing into the sink is half-built, and the question of whether the sink drains or overflows is, weirdly, now a venture-fundable question.
Where the venture money is actually pointed
Follow the money and the strategy reveals itself, because VC is not waiting for the tiers to trigger. Digital health funding bottomed around $10B in 2024 after the 2021 sugar high of $29B, and the composition shifted hard: AI-labeled companies took roughly 37% of dollars in 2024 and crossed 60% in 2025. Within that, the capital is split across the two sides of the labor trade with almost suspicious symmetry.
Side one is the automation book. Ambient documentation (Abridge at a $5.3B valuation, Ambience, Microsoft’s DAX franchise) attacking the two hours of pajama-time charting per clinician day. Autonomous coding (Nym, CodaMetrix, Fathom) going after the medical records workforce directly. Prior auth automation on both sides of the trench, vendors selling submission AI to providers and adjudication AI to payers, which is the healthcare equivalent of an arms dealer working both sides and an outcome the industry fully deserves. Agentic RCM platforms promising to collapse the cost-to-collect from 3% to 4% of NPR down toward 1%, which at a $5B revenue health system is a nine-figure annual prize. Denial appeals written by models, answered by models, in a closed loop where the only humans involved are the ones whose surgery is pending. The services-as-software thesis underneath all of it is simple: US healthcare admin is a multi-hundred-billion dollar labor market, software gross margins applied to labor budgets is the largest TAM expansion trick in venture history, and General Catalyst was explicit enough about it to buy an actual health system (Summa, roughly $485M) as a deployment vehicle, which is either visionary vertical integration or the most expensive product demo ever staged, ask again in 2030.
Side two is the redeployment book, and it is smaller, scrappier, and arguably the better risk-adjusted bet. Per-diem staffing marketplaces (Clipboard Health, ShiftKey, IntelyCare) built liquidity in exactly the segment, CNAs and LPNs in post-acute, where the labor sink needs throughput. Nurse hiring platforms like Incredible Health compress time-to-fill from 90 days to under 30, which is job matching infrastructure, a phrase that appears verbatim in the framework’s Tier 1 table next to a federal funding recommendation. Training-to-placement companies (Stepful, CareAcademy, the apprenticeship plays) sell exactly the employer-connected sectoral training the framework wants to fund with grants, and the evidence base it cites, the WorkAdvance-style sectoral studies showing significant income gains, is the same evidence those companies put in their fundraising decks. Credentialing and license-verification rails (Medallion, Certiverse-style plays, primary source verification APIs) are the picks and shovels for moving a licensed workforce across state lines and across employers fast. The pattern: side one shrinks the exposed half of healthcare labor, side two industrializes the intake pipe into the protected half, and several large funds hold positions in both, collecting a toll on the worker coming and going. Calling that solving the labor market is generous. Arbitraging it with a policy tailwind is closer. But the uncomfortable truth is that the arbitrage and the solution are the same physical infrastructure, and nobody else is building it at speed.
Licensing reform is the sleeper trade
Buried in Tier 1, easy to skim past: federal funding for states that make it easier to enter licensed occupations, with interstate recognition and shortened qualification paths called out specifically, and the framework citing the Kleiner and Soltas welfare analysis showing licensing reduces employment and raises prices without offsetting gains. Roughly 24% of US workers hold a license or certification, and healthcare is the most licensed sector in the economy by a mile. This is the single highest-leverage and least-discussed lever in the entire document for healthcare specifically.
The proof of concept already exists. The Nurse Licensure Compact covers 40-plus jurisdictions and is the only reason the travel nurse market could scale at all during COVID. Physician licensure remains state-fragmented despite the IMLC, telehealth licensure flexibility mostly snapped back after the PHE ended, and scope-of-practice fights between physician societies and NP/PA associations consume more state legislative calories than almost any other healthcare issue. Now run the framework’s logic forward: federal grant money conditioned on licensing liberalization, arriving at the exact moment AI clinical decision support gives regulators political cover to argue that an NP with a model achieves physician-equivalent outcomes for defined condition sets. That is not a hypothetical, the outcomes literature on NP-led primary care was already strong before the models showed up. Every incremental scope expansion reprices labor across the delivery system: an NP at $130k doing work previously gated to a $280k physician, an MA with AI support doing intake work previously gated to an RN, a CHW with a protocol doing what required an LPN. The companies positioned for this are the ones selling the supervision, protocol, and documentation layer that makes expanded scope defensible to a state board and an actuary, and the trade gets a federal subsidy if the framework’s recommendation becomes appropriations language. Watch state legislative sessions in 2027 the way credit people watch Fed minutes.
What breaks first
The optimistic version of this transition assumes the sequencing cooperates, and sequencing never cooperates. A few specific fragilities. First, the regional employment problem: in dozens of metros, the health system is the largest employer and its back office is the largest white collar employer within it. Automating 30% of admin headcount at the anchor institution of a mid-sized city is not a rounding error in the national statistics, but it is a local recession, and the framework’s own measurement section concedes that government statistical infrastructure cannot currently see that disruption at useful speed or granularity. Second, the UI delivery problem: the framework is blunt that state UI systems buckled under pandemic volumes, took months to implement extensions, and leaked billions in improper payments, and Tier 2 depends entirely on those same rails scaling on demand. Anyone who watched state systems melt down in April 2020 should price meaningful execution risk into every tier. Third, the margin capture problem: when AI takes out admin cost, the savings get divided among the software vendor, the provider, the payer, and theoretically the premium payer, and the historical base rate for healthcare efficiency gains reaching premiums is approximately never. If automation savings accrue to vendor valuations and payer MLR management while the displaced coder gets a wage-insured job at 70% of prior comp, the political economy of the whole framework gets ugly fast, and the framework’s authors clearly know it, which is why the document keeps returning to equity-sharing mechanisms and capital accounts funded with AI company stock. Fourth, Baumol: healthcare costs have outrun the economy for fifty years partly because care labor productivity could not grow, and AI is the first genuine challenger to that dynamic in living memory, but Baumol cuts both ways. If productivity finally rises in the automatable half while wages in the hands-on half have to climb to attract the redeployed workforce, the net effect on national health expenditure is genuinely ambiguous, and anyone selling certainty on that number is selling something else.
How this probably ends
A reasonable base case, stated plainly. The back office hollows out over five to eight years, faster than incumbents claim and slower than the demo videos imply, mostly through attrition rather than layoffs because health systems are politically incapable of mass RIFs in their own communities. Clinical labor splits: licensed hands-on roles hold wage power and gain AI leverage, while the communicative shell of clinical work gets absorbed by software. The care economy absorbs a real share of displaced workers from inside and outside healthcare, but only after wage floors get dragged up by some combination of Medicaid rate reform, wage insurance, and pure scarcity, and the absorption is uglier and slower than any white paper projects. Licensing reform turns out to be the policy that mattered most and got covered least. VC does not solve the labor market, because that was never the mandate, but it does end up owning the matching engines, training pipelines, credentialing rails, and automation layers through which the entire transition physically flows, which is a more durable position than solving anything.
The framework’s authors wrote that the economy has real capacity to adapt but adaptation is not automatic. For healthcare the sharper version is this: the sector spent fifty years building a trillion dollar administrative immune system around a clinical core it chronically understaffs, and AI is about to attack the immune system while policy herds millions of workers toward the core. Whether that ends as the great rebalancing of American healthcare labor or as an expensively subsidized churn machine depends on reimbursement rates set in fifty state capitols, UI mainframes written in COBOL, and a few hundred founders who mostly got into this to sell scribe software. Plan accordingly, and maybe be nice to the coders on the way out. Some of them will be taking care of everyone’s parents in ten years.