Where Health AI Is Actually Landing

AI is getting funded fastest where it attacks labor cost and recovers cash, and slowest where it has to prove a clinical outcome. Everything below is evidence of that claim or an exception to it.

Photo by Irwan on Unsplash

1. Clinical AI / Clinical Decision Support

What it is. Diagnostic assistance, treatment recommendations, and risk prediction at the point of care. Imaging triage, deterioration alerts, evidence retrieval.

Who it touches. Radiologists, ED and ICU physicians, hospitalists, quality teams. Patients benefit but don’t interact with the software.

What it delivers. Speed and standardization in diagnosis. Catch the PE or stroke faster; surface the right literature without the three-tab browser search.

Who buys it. CMO and CMIO for imaging AI; radiology leadership for triage tools. OpenEvidence is the outlier, free to physicians, ad-supported, no procurement cycle.

What AI is doing. Computer vision dominates: 1,450+ FDA-authorized AI devices, 76% in radiology. GE HealthCare leads with 120 authorizations. Aidoc got the first foundation-model device clearance (CARE1, rib fracture) in February 2025. LLMs power evidence retrieval. Predictive models run sepsis and deterioration alerts in the EHR background.

Key vendors. Aidoc, Viz.ai (imaging triage); GE HealthCare, Siemens Healthineers (embedded scanner AI); OpenEvidence ($250M Series D, $12B valuation, January 2026–40%+ of US physicians using it daily); Epic (embedded predictive models).

How proven it is. Split three ways. Imaging triage: proven, FDA-cleared, decade-long evidence base. Evidence retrieval: explosive adoption, thin outcomes scrutiny. Predictive CDS: the credibility gap. Epic’s sepsis model was marketed at AUROC 0.76–0.83; prospective measurement across 38,000 hospitalizations at Michigan Medicine returned 0.63, sensitivity of 33%. The benchmark in the deck is not the number that runs in your hospital.

2. Administrative Automation

What it is. Software handling the paperwork around a care encounter. Scheduling, intake, prior authorization, referrals.

Who it touches. Front-desk and back-office staff, prior-auth nurses, schedulers, patients (experiencing delays), payers (receiving the requests).

What it delivers. Labor cost reduction and speed. Every prior auth submitted without a human touching a keyboard is a measurable saving.

Who buys it. Practice administrators and health-system operations leaders. Physicians create internal pressure. They’re the ones waiting on approvals. Displaced administrative staff is the adoption tension.

What AI is doing. NLP reads clinical documentation, ML interprets payer policy, RPA/agents navigate portals. Optum’s Digital Auth Complete connects to 250+ payer systems; Availity’s AuthAI returns a recommendation in under 90 seconds… framed as a recommendation, not an autonomous denial, which is the key liability distinction. Conversational voice AI handles inbound scheduling. CMS-0057 mandates payer FHIR APIs; the WISeR pilot brings AI-enabled prior auth to Medicare starting January 2026.

Key vendors. Cohere Health, Waystar, Availity (prior auth); Notable, AKASA (broader admin); Hippocratic AI (patient-facing voice).

How proven it is. Among the most mature AI categories in the stack. Labor math is auditable and regulatory mandates create durable demand. Still human-in-the-loop on provider-side review, which contains the liability footprint.

3. Revenue Cycle Management

What it is. The financial lifecycle of a claim. Coding, scrubbing, submission, denial management, appeals, payment integrity.

Who it touches. Coders, billers, CDI specialists, revenue cycle directors, CFOs. Physicians indirectly, since documentation quality determines coding accuracy.

What it delivers. Cash recovery and leakage prevention. The clearest dollar-for-dollar ROI in health tech. A hospital losing $10M annually in preventable denials will pay almost anything to get it back.

Who buys it. Hospital CFO and VP of Revenue Cycle. CDI nurses are the clinical influencers. Their documentation habits determine whether claims get paid.

What AI is doing. Generative AI drafts appeal letters. Waystar reports cutting appeal-package creation from hours to 16 minutes. Predictive models flag high-risk claims before submission. NLP surfaces missed diagnosis codes from unstructured notes. Agents navigate payer portals autonomously. SmarterDx reads the full chart to surface documentation evidence for conditions that weren’t coded or weren’t adequately supported.

Key vendors. Waystar (reports $15.5B in prevented denials); SmarterDx/Smarter Technologies (60+ health systems, claimed 5:1 ROI); FinThrive, Experian Health. PE consolidation is the signal: New Mountain Capital is combining SmarterDx, Thoughtful.ai, and others into what is reportedly a ~$30B holding company.

How proven it is. The most mature and highest-conviction AI vertical in healthcare. ROI is direct and auditable, denied claims are a line item with a number attached. When PE rolls up a sector, they’ve already run the numbers.

4. Ambient Clinical Intelligence / AI Medical Scribing

What it is. Real-time capture of the physician-patient conversation, auto-generating structured clinical notes, codes, and increasingly orders.

Who it touches. Physicians primarily, nurses increasingly. Patients are in the room. Downstream coders receive the generated note.

What it delivers. Time savings and burnout reduction. The secondary pitch i.e. where the real money is, is that a better note produces better coding and more defensible billing.

Who buys it. CMIO and CMO own the clinical rationale; CFO owns the ROI case. KLAS rankings function as the de facto procurement filter. The physician is both end user and make-or-break adoption variable.

What AI is doing. ASR captures the conversation; LLMs summarize it into structured documentation mapped to note templates. Advanced implementations layer “contextual reasoning” not just transcription but inference about what the note should contain given the full chart context.

Key vendors. Abridge (~$5.3B valuation, $316M Series E extension April 2026, 150+ health systems, ~$117M contracted ARR, ~$2,500/clinician/year); Microsoft Nuance DAX Copilot (enterprise installed base, M365 integration); Ambience Healthcare ($243M Series C, $1.25B valuation, July 2025); Suki, Nabla, Commure/Augmedix.

How proven it is. Real adoption, modest measured ROI. The largest controlled study (JAMA, April 2026, 8,581 clinicians, five health systems) found 13–16 minutes saved per day “modest” per the authors. Burnout improved. Incremental billing revenue averaged $167/clinician/month. Power users saved more (21–27 minutes), but only 32% of adopters used it that consistently. Abridge’s valuation prices in expansion into coding and RCM that is not yet proven. The moat is Epic integration any scribe without it faces a permanent adoption ceiling.

5. Population Health Management

What it is. The analytic backbone of value-based care. Risk stratification, care-gap identification, chronic disease management, and risk-adjustment coding across a defined population.

Who it touches. Care managers, population health directors, ACO leadership, payers, primary care physicians, high-risk patients.

What it delivers. Cost avoidance and quality-score improvement. Find the CHF patient trending toward readmission before they cost $80K, and intervene for $200.

Who buys it. ACO executives and IDN population health leaders in risk-bearing contracts. Payer medical-economics teams. The ROI only materializes if the organization is actually in a contract that pays for better outcomes which is still not most of US healthcare.

What AI is doing. Predictive risk models score patients by likelihood of near-term high-cost events. NLP surfaces documented-but-uncoded conditions, a direct HCC risk-adjustment revenue mechanism. Chart-retrieval automation replaces manual nurse review. Natural-language “copilot” interfaces let care teams query population data without SQL. Agentic AI is beginning to automate care-management documentation.

Key vendors. Innovaccer ($275M Series F, ~$3.45B valuation, January 2025; 80M+ patient records; 7 of 10 largest US health systems); Cedar Gate, Arcadia, Health Catalyst; Epic’s population health module.

How proven it is. Mature on analytics, emerging on AI-driven action. The infrastructure has existed for a decade; the AI layer converting a risk score into autonomous care-team action is newer and thinner on evidence. ROI is real but contract-dependent. CMMI’s ACCESS Model (rates live February 2026) is a genuine tailwind — it creates more risk-bearing arrangements that actually pay for this work.

6. Patient Engagement & Care Navigation

What it is. Patient-facing tools extending the care team’s reach outside the visit. Scheduling and intake agents, post-discharge follow-up, billing engagement, digital therapeutics, remote monitoring.

Who it touches. Patients directly. Call-center and care-navigation staff being augmented or displaced. Billing departments. Care managers.

What it delivers. Scale and access. Handling outreach, intake, billing, and monitoring at volumes no human staff could sustain, especially during ongoing workforce shortages. Digital therapeutics pitch clinical improvement in enrolled patients.

Who buys it. Health systems, payers, employers (for chronic-disease management). The patient is almost never the paying customer. Which is the structural problem this category has been fighting since inception. The buyer is an employer or plan underwriting cost-of-care reduction, not patient experience.

What AI is doing. Hippocratic AI’s voice agents handle non-diagnostic outreach with explicit guardrails against diagnosis or prescribing. Cedar’s Kora voice agent targets patient financial communication (30% call-volume reduction claimed). RPM platforms layer predictive models on device data streams. CPT 2026 added new short-duration RPM codes a meaningful reimbursement signal.

Key vendors. Hippocratic AI ($126M Series C, $3.5B valuation, November 2025; $404M total raised; 115M+ reported patient interactions in six countries); Cedar (50M+ patients, $10B+ in payments processed); Hinge Health (IPO May 2025, ~$2.6B valuation, FY2024 revenue $390M, down ~60% from 2021 private mark); Omada Health (IPO June 2025, ~$1.1B valuation, FY2024 revenue $170M).

How proven it is. Commercial traction real; durable clinical outcomes thin. Hinge and Omada’s IPO prices say it plainly. Both priced well below 2021 peaks. Vendor outcomes exist but independent RCTs show engagement improvements that often don’t persist past one month. The business model problem hasn’t been solved: building something genuinely useful for patients without a buyer whose budget it fits is a constraint technology doesn’t fix.

7. Payer / Utilization Management

What it is. Software payers use to review claims, conduct medical-necessity review, manage authorizations, and detect fraud, waste, and abuse.

Who it touches. Payer medical directors and UM nurses. Claims adjudicators. Providers on the receiving end of denials. Patients whose care is approved or not.

What it delivers. Medical cost containment. Every unnecessary admission not authorized falls to the payer’s bottom line. FWA recovery is a secondary mechanism.

Who buys it. Health plan CMO and CFO. Legal and compliance are the most important influencers right now this is the vertical where AI deployment has created the most litigation exposure in the sector.

What AI is doing. Predictive models flag claims for human review. NLP maps clinical documentation against coverage policies. Automation routes clear-case approvals and escalates borderline decisions. Payment integrity AI (Machinify) scans paid claims for improper-payment patterns and generates recoupment requests.

Key vendors. Optum’s nH Predict (NaviHealth acquisition); Availity, Cohere Health (prior-auth side); Machinify (payment integrity); Innovaccer’s Galaxy (payer risk and UM analytics).

How proven it is. Operationally deployed at scale and the most legally exposed vertical in the stack. UnitedHealth’s nH Predict faces an active class action alleging that 90%+ of the model’s denials are reversed on appeal, while fewer than 0.2% of patients ever appeal. The February 2025 ruling dismissed five of seven counts but allowed breach-of-contract and good-faith claims to proceed; 2026 brought broad discovery. Cigna and Humana face parallel suits. The structural lesson: full automation is safe in administrative plumbing and radioactive where it overrides clinical judgment. The payer captures the margin and bears the downside.

8. Supply Chain & Operational AI

What it is. Software optimizing hospital operations, inventory, nurse and physician scheduling, bed and capacity management, patient flow.

Who it touches. Nurse managers, operations and throughput leaders, supply-chain teams, command-center staff.

What it delivers. Cost reduction and capacity unlock, fewer contract-labor hours, shorter wait times, higher effective bed utilization without capital build.

Who buys it. COO and CNO sign the contract; finance is involved because contract-labor is one of the largest controllable P&L line items. CMS minimum-staffing rules (April 2024) make capacity tooling close to mandatory for compliance.

What AI is doing. Predictive demand and acuity forecasting models project patient volume by unit by hour. Automated shift-matching fills scheduling gaps without 11pm phone calls. Capacity prediction models sequence patient movement ED boarding, OR turnover, discharge likelihood. Emerging agentic workflows route bed requests, transport, and supply orders without human dispatch.

Key vendors. LeanTaaS (iQueue for ORs and infusion), Qventus (capacity and care progression), TeleTracking (patient flow), QGenda (workforce scheduling); Epic command center from inside the EHR.

How proven it is. Proven in pockets with strong tailwinds. Stanford Health Care cut chemotherapy-infusion wait times 31–40% with predictive load-leveling; Ascension reportedly cut contract-labor dependency 15% in six months. The mechanism is simple enough that the math is auditable. The ceiling is change management the AI can forecast accurately and the nurse manager still has to trust the model enough to staff to it.

9. Health Information Exchange & Interoperability

What it is. The infrastructure enabling health data to move. FHIR APIs, data normalization, longitudinal records, and the TEFCA “network of networks.”

Who it touches. Every other vertical depends on it. Directly: health-system IT, payer data teams, public health agencies, every developer building a clinical or operational AI application.

What it delivers. Data liquidity, the precondition for every other AI use case. A scribe needs the chart. A population health model needs the claims. A CDS tool needs the labs. None of it works if the data can’t move or can’t be trusted. AI’s specific contribution is semantic normalization: making the same concept coded 50 different ways mean the same thing everywhere.

Who buys it. Health-system CIOs and CTOs buy integration software. Payers buy normalization and record linkage. The network itself is a public good, which is precisely why it was chronically underfunded and required regulatory mandates to move.

What AI is doing. Patient matching resolving that “John Smith, DOB 3/15/1962” at Hospital A is the same person as “John Smith, DOB 03–15–62” at Clinic B. Terminology mapping to standard ontologies (SNOMED, LOINC, RxNorm). Generative AI is appearing in FHIR-layer tooling, allowing natural-language queries against health data.

Key vendors. Health Gorilla, Datavant, Particle Health (connectivity and normalization); Redox (integration infrastructure); Innovaccer’s Gravity (AI-on-data layer); the five TEFCA QHINs as the regulatory backbone.

How proven it is. Regulatory-driven and foundational; the AI intelligence layer is nascent. TEFCA hit ~500M records exchanged by 2026. HTI-1 raised EHR certification to USCDI v3 (effective January 1, 2026); CMS payer-API requirements land January 1, 2027. The mandates are creating real data flow for the first time. Value accrues as enabling infrastructure — difficult to monetize directly, easy to underinvest in. The beneficiary is the ecosystem.

The Pattern

Map these nine on one axis, does the buyer capture the value directly?, and the capital allocation becomes legible.

RCM, ambient scribing, administrative automation, and operational AI all share the same property: a CFO can underwrite the check before the outcome is proven. Labor cost is the line item, cash recovery is the metric, and the contract pays for itself. That’s where the confident capital went in 2025.

Patient engagement, population health, and predictive CDS all require either a clinical outcome or a risk-bearing contract structure that monetizes that outcome. Both are real, both are slower. The buyer and the beneficiary keep not being the same person.

Drug discovery and interoperability are patient-capital plays. Infrastructure takes a decade; so does an approved drug.

The $14.2B that moved in 2025 is a specific, legible bet: the plumbing gets monetized before the outcomes do, and then the outcomes follow. Whether that sequencing holds is the question the next fund cycle will answer.