Sector: Healthcare · Framework v0.1 · Vintage 2026-07-10
Healthcare AI: Ambient documentation is the first at-scale deployment
Horizon: 0-15y (now / near / decade)
1. Unit definition
This proposal treats "healthcare AI" as the application of foundation-model-era artificial intelligence to the delivery, administration, and financing of health services (with US health systems as the evidentiary anchor because that is where at-scale deployment is currently observable, and with explicit disclosure of what does and does not travel to non-US, non-EHR-integrated settings).
The boundary is deliberately drawn to include ambient clinical documentation, medical coding & billing, prior authorization, imaging read, clinical decision support, and administrative back-office (the highest-exposure task clusters in the framework's exposure model). It excludes drug discovery (a different capital and regulatory regime with a longer diffusion clock) and consumer wellness applications (a different distribution and evidence problem).
The unit spans a $10.4T global health economy [16], of which the US is roughly 17-18% of GDP [16]. Sector-level projections range widely and are labelled as projections throughout, per the framework's evidence standard.
| Baseline metric | Value | Source |
|---|---|---|
| Global health spending (WHO, 2022 basis, projected forward) | ~$10.4 trillion | [16] |
| US health spending share of GDP | ~17-18% | [16] |
| Projected US healthcare AI savings by 2030 (order-of-magnitude, methodology-dependent) | ~$150 billion / year | [2] |
| Adoption gap: organisations using AI / seeing enterprise-level EBIT impact | 88% / 39% | [2] |
2. Exposure assessment
Healthcare is high exposure. The exposed tasks are cognitive, language-heavy, and routine-shaped: documentation, coding, prior authorisation, structured summarisation of imaging reads, and administrative correspondence. The exposure profile is uneven across the sector: administration and documentation are dominantly exposed, diagnosis is meaningfully exposed but liability-gated, and hands-on care is low-exposure this decade absent robotics maturity.
| Task cluster | Exposure read | Source |
|---|---|---|
| Clinical documentation (notes, discharge summaries, letters) | Very high: language-heavy, structured output, currently the largest observable deployment surface (ambient documentation). The at-scale case: Abridge across 150+ US health systems [1]. | [1] [2] |
| Medical coding & billing, prior authorisation | Very high: rule-based transformation of language into structured codes, exactly the shape foundation models handle well. | [2] |
| Triage & routing (nurse-line, admin front door) | High: language-heavy, but liability-gated where clinical judgement enters. | [2] |
| Imaging read: structured report generation | High: augmentation not replacement; the read is generated from images plus prior report style. | [2] |
| Diagnosis (differential, primary care) | Meaningfully exposed, but heavily friction-gated by liability and reimbursement. The framework predicts augmentation this decade, not replacement. | [2] |
| Care coordination, chronic-disease follow-up | Medium: language + workflow, but crosses institutional and payer boundaries where the workflow-redesign lag bites hardest. | [2] |
| Hands-on care (bedside, procedures) | Low this decade: a robotics-maturity story, not a foundation-model one. | [2] |
3. Readiness assessment
Readiness in healthcare is bimodal: the same clinical-documentation AI means something structurally different in a US health system with Epic and structured clinical data than in a district hospital with paper records. This bimodality is the single most important shape-of-outcome fact in this proposal.
Strong: Epic and Cerner-class EHRs are the substrate that lets ambient documentation land in the chart, not just in a transcript file. This is why Abridge scales to 150+ health systems and analogous tools stall elsewhere.
Weak to absent: paper records or fragmented digital records mean the exposure is theoretical until a structured-data substrate exists. Framework's readiness score would floor here.
Uneven: clinician upskilling is happening ad hoc, not through a national programme; large systems have AI officers, small practices do not.
Broadly weak, but AI-readiness workforce policy is a live national-strategy item in ~12+ new national AI plans published in 2024 alone.
Strong: vertical AI has attracted large valuations (Abridge $5.3B [1], Harvey $11B [1], Legora $5.55B [1]) and healthcare AI is a named priority in hyperscaler capex [5][6].
The sovereign-AI layer is starting to matter here (UK £500M Sovereign AI Fund [7], UAE MGX $49B [8]), but AI-readiness capital lags AI-infrastructure capital by roughly an order of magnitude.
Adequate at the encounter level (EHR), still weak longitudinally (patient across providers).
Weak: the substrate problem again; without structured intake, the model has nothing to condition on.
Emerging: health-system AI committees, FDA's AI/ML action plan, some outcome tracking, but no shared evaluation standard for foundation-model clinical tools yet.
Cross-national institutional gap is real; readiness indices (IMF AI Preparedness [3], Oxford Insights Government AI Readiness [4]) exist to name this but do not by themselves fix it.
4. Friction map
Healthcare's friction stack is the sector's defining feature, and the framework's own explanation for the 88%/39% adoption gap. Friction here is not just a brake on adoption; it shapes who captures the value.
Ambient documentation has largely stayed under the SaMD line by not producing clinical decisions; the moment a tool suggests a differential or a dose, the regulatory bar shifts. This is why the current deployment surface is documentation, not diagnosis.
Case-law allocating fault for AI-influenced clinical error is still forming. Until liability is priced by insurers, the augmentation model dominates; replacement is uninsurable.
AI-specific CPT / procedure codes are the throttle on adoption at the practice level: no code, no revenue line, no procurement case for anyone but the largest systems.
Nurse and physician associations shape what tasks a tool can even be permitted to touch; low task-substitution risk in the near term, meaningful task-recomposition risk in the decade.
Care coordination is the exposed task with the highest workflow-redesign cost because the workflow spans institutional and payer boundaries.
5. Channel analysis (two clocks)
Five channels, each across three horizons: Now (0-2y), Near (2-5y), Decade (5-15y). The framework's two-clocks discipline: the investment clock is fast and visible; the diffusion clock is slow and decisive.
Ambient documentation cuts clinician documentation time in measurable ways in the US systems that have deployed it; task augmentation, not net role loss, at the frontline. Admin / coding roles are the first with observable task-composition change.
WEF-class projections (net across the whole economy: +170M new / -92M displaced roles by 2030 [9]) hide healthcare's distributional profile: admin restructures, clinical roles re-compose, hands-on care is stable. Never report the net; netting hides the story per the framework.
Access-effects in low-doctor-density countries are the largest decade-scale labour story: extending clinical reach where practitioners are scarce, without displacing them. Least-funded of the decade-scale stories.
Vertical healthcare AI attracts brand-name capital (Abridge $5.3B, Harvey adjacent-vertical $11B, Legora $5.55B [1]); hyperscaler capex (~$600-725B in 2026 [5][6]) is the upstream flow supplying the compute layer this all rides on.
Sovereign AI capital (UAE MGX $49B [8], UK £500M Sovereign AI Fund [7]) starts to shape non-US readiness. Investment concentrates further where readiness is high (US, UK, Singapore-class), potentially widening the bimodality.
Capital-into-readiness (data substrate, workforce, institutional capacity) is the highest-leverage flow for the decade-scale story, and structurally under-supplied by market mechanisms alone.
Measured productivity gains are still emerging. Observed at the tool level (documentation-time reduction), not yet at the sector-productivity level. Per the framework: separate observed from forecast.
Penn Wharton projects generative AI's whole-economy GDP contribution at ~+1.5% by 2035 [10]; Goldman-style ranges span +1.5% to +15% depending on diffusion-speed assumptions [10]. The gap between these is the Solow-paradox / two-clocks disagreement, not a measurement error.
Sector-specific productivity read requires BLS output-per-hour in health services to move: a specific, observable, currently-untracked-in-headline signal. Called out in section 8.
Vertical AI incumbents (Abridge) sit above the foundation-model layer with proprietary data and EHR-integration workflows the base model cannot replicate. The 'wrapper' layer above them is being repriced as base-model capability moves up [12].
Foundation-model providers move into vertical workflow layers; incumbent health-system IT vendors (Epic, Oracle Health) either integrate or get disintermediated. Consolidation likely in the tooling / middleware tier.
The load-bearing question: does the value accrue to foundation-model providers, to vertical AI incumbents, or to the health systems themselves? Currently distributed, decade-scale outcome depends on data-rights and workflow-lock-in mechanics.
Fiscal effect is marginal but detectable: data-centre property-tax windfalls flow to specific US counties [11], not to health-system budgets. Public-service effect: uneven across state Medicaid systems.
Reimbursement policy is the near-term public-sphere lever. CMS and state-Medicaid decisions on AI-associated CPT codes will shape which providers can adopt at all.
The decade-scale public-sphere question is access: whether healthcare AI extends care to under-served populations (the largest decade-scale upside) or accelerates a two-tier system (the largest downside). Both scenarios are live.
6. Scenarios (decade-scale)
Scenario ranges, not point estimates, per the framework's evidence standard. Each scenario names the load-bearing assumption that separates it from the others.
Workflow redesign takes 15-30 years, matching the electricity precedent; measured productivity gains lag deployment by a decade or more.
Capex-heavy, revenue-light decade; foundation-model providers and hyperscalers absorb losses; a 2000-telecom-fiber-glut outcome for some capital, but the infrastructure enables the next decade. Clinical impact is real but slow to show in BLS output-per-hour or CMS spend curves.
Diffusion follows a faster curve than any prior general-purpose technology because the interface (natural language) removes the training barrier, and healthcare's structured workflows lend themselves to redesign.
Measured productivity gains within 5 years in documentation-heavy specialties; reimbursement codes and liability frameworks form on a matching timeline; access-extension in non-US settings begins in the late 2020s. Upper bound of the Goldman +1.5% to +15% GDP range.
Readiness rather than exposure is the binding constraint; capital flows to high-readiness geographies; low-resource-setting adoption requires public / sovereign capital that does not fully materialise.
US and comparable systems restructure administration and augment diagnostics this decade; low-doctor-density countries see uneven, project-scale pilots but no substrate build-out; the largest access-extension upside is deferred to the following decade or foregone. Considered by the authors of this proposal the modal scenario absent explicit policy intervention.
7. Implications
- The highest-leverage lever is reimbursement. AI-specific CPT/procedure codes shape which providers can adopt at all.
- Structured-data-substrate investment (EHR, longitudinal patient records, portability standards) is the pre-condition every framework dimension depends on; without it, adoption capital lands unevenly.
- For non-US governments: the AI-readiness gap between national AI strategies and institutional capacity to execute them is the specific gap sovereign-AI capital should be sized against.
- The survival trait at the vertical layer is proprietary data, domain expertise, or workflow integration the base model cannot replicate. Undifferentiated 'wrapper' positioning is being repriced.
- The next tier of deployment surface after ambient documentation is medical coding & prior authorisation: same shape (language → structured output), same substrate (EHR).
- Regulatory-boundary discipline matters: staying under the SaMD line by not producing clinical decisions is what let ambient documentation scale before diagnostic tools; that boundary is durable, not a temporary loophole.
- Task composition changes before headcount does; the framework predicts augmentation not replacement across the decade for clinical roles, with meaningful task-recomposition for admin and coding.
- The skills that hold value are the ones the model does not do: judgement under ambiguity, patient relationship, cross-provider coordination.
- Scope-of-practice negotiations and reimbursement-code decisions are the concrete points where worker organisations shape outcomes.
- Two clocks discipline: capex is real GDP today; the productivity payoff is a forecast. Position accordingly by clock, not by narrative.
- The 2026 generative-AI market spread of $121B-$395B [5] is a methodology spread, not a forecast disagreement. Do not treat point estimates as data.
- The bimodal scenario is the modal outcome absent policy intervention; healthcare-AI returns will concentrate in high-readiness jurisdictions and in vertical incumbents with EHR-integration moats.
8. Signals to watch
Falsifiability section: the observable indicators that would confirm or kill this proposal's thesis. If none of these move, the proposal is not being tested.
A CPT code specific to ambient documentation, or a Category III code progressing to Category I, would move this signal materially and accelerate small-practice adoption.
A first appellate ruling apportioning liability between clinician, health system, and AI vendor would let insurers price the risk and unblock diagnostic-tool deployment.
Deployment past 20+ non-US health systems (currently near-zero at scale) would falsify or delay the bimodal scenario.
ARR growth continuing at Cursor-class rates ($0 → $2B in under three years [1]) would confirm the vertical incumbent thesis; a plateau would signal foundation-model absorption.
Two consecutive quarters of measured productivity gain above trend in health services is the ultimate two-clocks scoreboard.
Sustained authorisation-count growth without a high-profile safety event would signal the friction stack loosening on the diagnostic side.
A material downward revision would shift the entire sector's cost curve and the plausibility of the Goldman-optimistic scenario.
Upward revisions above +1.5% GDP by 2035 [10] would shift the sector's decade-scale scenario weighting toward the optimistic case.
9. Sources & data vintage
- AI market landscape: Impact Suite Module One (Abridge, Harvey, Legora, Cursor deployment / valuation figures) (ValueAdd VC)
- Sector Impact Framework v0.1: the exposure × readiness × friction model and the healthcare micro-example (Penn Wharton Budget Model)
- AI Preparedness Index (IMF)
- Government AI Readiness Index 2025 (Oxford Insights)
- Gartner: Worldwide AI spending to grow 47% in 2026 ($2.59T; genAI range $121B-$395B methodology spread) (Gartner)
- Hyperscaler capex 2026: >$600B, +36% YoY (IEEE ComSoc)
- UK Sovereign AI Fund: £500M (Wikipedia)
- Sovereign AI Index: MGX $49B, UAE/Japan concentration (CNAS)
- AI and the global economy: WEF (170M new / 92M displaced by 2030 net projection) (WEF)
- Projected impact of generative AI on future productivity growth (+1.5% by 2035 baseline; Goldman-style range references) (Penn Wharton Budget Model)
- The AI Economy: sundaypyjamas.org North American Economy learning kit (hyperscaler capex, rural data-centre property-tax windfalls, Solow paradox framing) (Goldman Sachs (referenced))
- AI Company Rankings 2026 (vertical AI concentration, wrapper repricing signal) (TLDL)
- Global health spending / US share of GDP context (Impact Suite economic backdrop) (Penn Wharton (order-of-magnitude reference))