Begin with what America does well

The American health system combines biomedical science, entrepreneurial capital, large-scale manufacturing, advanced hospitals, professional specialization, public programs, and employer-sponsored coverage. NIH directs most of a nearly $48 billion budget through almost 50,000 grants supporting more than 300,000 researchers at more than 2,500 institutions. NIH

Its public foundation is equally important. Medicare, Medicaid, CHIP, VA, IHS, public hospitals, and safety-net organizations provide coverage and care at enormous scale; community health centers alone served 32.4 million people through 139.4 million visits. HRSA America’s productive food and logistics system offers another strong base for healthier choices. Consumers spent $2.58 trillion on food in 2024. USDA Food Dollar

The next advantage is connection. A therapy has limited effect if it cannot be obtained; a test does not improve health if nobody acts on it; and a discharge is incomplete without medicine, equipment, transport, and follow-up. The final mile is how excellence becomes a human result.

The improvement opportunity is measurable

CMS projects health spending to grow 5.4% annually from 2025 through 2034, faster than projected GDP growth of 4.1%, reaching 20.6% of GDP. CBO projects major federal health-program outlays rising from $1.9 trillion in 2026 to $3.1 trillion in 2036. CMS, CBO Growth reflects aging, valuable new therapies, skilled labor, resilience, and better access as well as inefficiency.

The goal is therefore not indiscriminate spending reduction. It is to improve four connected dimensions:

• Spending and price: what care costs and who pays.
• Use and appropriateness: whether services benefit the person.
• Administration and capacity: how much work and infrastructure delivery requires.
• Access and distribution: who receives the benefit and who remains excluded.

A repeated test is a coordination problem; a medication injury is a safety problem; an unnecessary procedure is a clinical and incentive problem; and a low-margin injectable shortage is a manufacturing problem. The often-cited $760–$935 billion waste estimate and CAQH’s roughly $440 billion administrative-workflow estimate are useful taxonomies, not cash pools AI can automatically recover. JAMA, CAQH

The strongest opportunity sits between institutions

Certified EHR technology supports more than 96% of hospitals and 78% of office-based practices, but digitization is not the same as a shared plan. Clinical facts, benefits, pharmacy activity, results, images, bills, evidence, and preferences remain divided across systems. ONC/ASTP

Agentic AI can preserve a task across those boundaries, identify what is missing, direct the next action, verify completion, and escalate exceptions. The productivity test is not how much content it generates; it is whether verified work, delay, error, or risk disappears and the recovered capacity reaches patients and professionals.

Patients and families are participants, not endpoints

People do far more than receive care. They recognize symptoms, provide histories, choose among uncertain options, travel, schedule, pay, administer medicines, monitor recovery, and coordinate between organizations. Families often become an unpaid operational layer, particularly for children, older adults, people with disabilities, serious illness, or multiple chronic conditions.

A better system recognizes patient and caregiver time as a real resource. It asks how many calls, forms, visits, miles, repeated histories, and hours of uncertainty a pathway requires. It measures whether a person understood the plan, obtained the medicine, reached the appropriate professional, and knew where to seek help. It provides accessible human service for people who cannot or do not want to use a digital channel.

Employers and public programs are healthcare operators

Employers are not merely premium payers. In 2025, 67% of covered workers were enrolled in self-funded plans, including 80% at organizations with at least 200 workers. KFF In these arrangements, the employer may bear claims risk while a third-party administrator operates the network and claims, a PBM manages pharmacy benefits, and brokers, consultants, and stop-loss carriers influence the design. These organizations are natural buyers of navigation, benefit explanation, administrative simplification, and specialty-drug coordination. They also need strict separation between health data and employment decisions.

Government programs also operate distinct care and administrative environments. Medicaid combines federal and state financing and covers populations with different needs across states. Medicare operates through traditional coverage, private Medicare Advantage plans, drug benefits, contractors, and providers. VA integrates direct delivery with community care. IHS and tribal health programs operate under particular sovereign, facility, workforce, and data conditions. Correctional reentry programs must coordinate coverage, medicine, behavioral health, and community services at a high-risk transition. An agent designed for one environment should not be copied into another without adapting authority, policy, population, and evidence.

Products have lifecycles, not just launch dates

A medicine moves from biological hypothesis through experiments, candidate selection, preclinical work, clinical trials, regulatory review, manufacturing, distribution, prescribing, coverage, dispensing, adherence, safety monitoring, and continuing evidence. A medical device has a parallel lifecycle involving design controls, validation, procurement, training, maintenance, utilization, complaints, recalls, and replacement. Success at one stage does not guarantee patient benefit at the end.

Only about 12% of drugs entering clinical trials are ultimately approved, according to CBO’s review. Estimated development cost per successful drug varies widely depending on method and assumptions. CBO AI may reduce search and experimental cycles, improve trial feasibility, and help regulatory teams maintain evidence. But a generated molecule is not a medicine, a matched record is not an enrolled participant, and a faster submission is not proof of safety or effectiveness. The unit of value is qualified learning that changes a real decision.

1. Make care easier to access and navigate

Coverage does not always translate into usable care. People encounter unavailable appointments, inaccurate network directories, prior authorization, travel, language barriers, inaccessible interfaces, uncertain prices, and fragmented service channels. Rural communities and many low-income or underserved areas face fewer professionals and facilities. Behavioral health, maternity care, home services, specialty care, dental care, and long-term support can be particularly difficult to obtain.

HRSA projects a shortage of 141,160 physicians in 2038, including 70,610 in primary care, with much larger relative gaps in nonmetro areas. HRSA Software cannot create a clinician or facility, but it can use existing capacity better: prepare referrals, find the correct destination, coordinate telehealth and travel, assemble records, monitor waits, and route exceptions to regional teams.

2. Return time to patients and professionals

Administrative work is not inherently wasteful. Documentation, eligibility, privacy, payment accuracy, safety, and appeals serve legitimate purposes. The opportunity is to remove duplicate work and adversarial reconciliation. Patients should not spend hours determining who owns the next step. Clinicians should not reconstruct a chart from scattered records. Staff should not repeatedly enter the same data or poll a portal for status.

The best early agents will work quietly: retrieving records, preparing information, drafting from evidence, moving a process through approved tools, and presenting exceptions rather than adding alerts. Their value should be measured in net time after review and correction—not gross minutes saved in one task while work moves elsewhere.

3. Strengthen prevention and chronic-care continuity

CDC reports that people with chronic and mental health conditions account for roughly nine of every ten dollars in annual U.S. health expenditure. This does not mean 90% is preventable. It means a projected $6 trillion health economy must produce health over time, not only during encounters. CDC

Cardiovascular disease, diabetes, kidney disease, cancer, mental illness, substance use, dementia, respiratory disease, and multimorbidity require ongoing medication, monitoring, behavior support, and adaptation. The operational failures are often familiar: a measurement is not reviewed, a medicine is not filled, a laboratory test is overdue, a referral is incomplete, or a change in symptoms does not reach the right professional. A longitudinal agent can maintain that state and make missing action visible while clinicians retain medical authority.

4. Improve affordability and financial clarity

Households experience healthcare cost through premiums, taxes, wages, deductibles, copayments, uncovered services, medicine prices, travel, and time away from work. Even an insured person may not know the likely cost or whether a professional is truly in network until after receiving care.

Agents can explain benefits, retrieve current criteria, estimate patient responsibility from available data, compare covered options, find assistance, and reconcile authorization, claim, and bill. They cannot by themselves change market power, benefit generosity, drug pricing, or public eligibility law. Their role is to make existing rights usable and policy reform easier to administer.

5. Improve safety and complete every handoff

Healthcare is safer when a signal becomes accountable action. In an HHS OIG clinical review of Medicare hospital stays in October 2018, one quarter of patients experienced an adverse event or temporary harm event, and reviewers considered 43% of those events preventable. Medication events were the largest category. The finding applies to that sampled population and period, but it illustrates the continuing opportunity. HHS OIG

A predictive alert alone is not a safety system. The complete loop identifies the responsible person, provides relevant evidence and uncertainty, proposes an approved response, observes whether the response occurred, reassesses the situation, and escalates if necessary. The same design applies to abnormal tests, deteriorating patients, medication discrepancies, infection signals, recalls, and post-discharge risk.

6. Make medicine and device supply more resilient

Advanced care depends on ordinary products being available at the right moment. FDA was tracking 102 drug shortages as of July 31, 2024, including 71 sterile injectables commonly used in hospitals and cancer treatment. GAO FDA has identified low profitability, weak market reward for mature quality systems, and slow recovery after disruption as important causes. FDA

Medical-device supply deserves the same attention. FDA’s June 2026 shortage list included products used in oxygenation, breast biopsy, neurosurgery, and endoscopic vessel harvesting. A listing does not establish patient harm, but it identifies dependencies that operational leaders need to manage. FDA

Agents can connect demand, inventory, manufacturing quality, logistics, substitution rules, utilization, expiration, and local clinical priority. They can identify risk earlier and coordinate conservation or replenishment. They cannot create a missing production line or supplier. Resilience still requires purchasing contracts, quality incentives, redundancy, capital, and transparent risk sharing.

7. Accelerate qualified scientific and operational learning

Biomedical evidence now exceeds what any individual can continuously absorb. Research teams must synthesize literature, integrate complex data, design experiments, activate sites, recruit participants, monitor trials, maintain regulatory documents, and learn from post-market experience. AI can improve each step, especially when it operates through specialized scientific tools and preserves provenance.

The standard is not the number of hypotheses or documents generated. It is the rate of reproducible learning: better experiments, earlier failure of weak candidates, more feasible protocols, representative enrollment, cleaner evidence, faster identification of safety signals, and more effective products reaching appropriate patients. Scientific judgment, participant protection, regulatory accountability, and manufacturing quality remain human and institutional responsibilities.

These seven priorities share a practical pattern. Information exists somewhere, but the responsible participant does not receive it in time, the next action is unclear, or nobody verifies completion. This is the opening for governed agentic AI.

Nine properties of a useful healthcare agent

The nine properties are a bounded goal, permissioned context, a plan, approved tools, state over time, verification, visible uncertainty, escalation, and accountability. Together they distinguish a system that completed an authorized task from one that merely produced a plausible response.

Authority should match consequence. A scheduler can make reversible changes under clear rules; a clinical-treatment or coverage decision affects rights, health, and finances and requires stronger evidence and an accountable professional or lawful process.

Clinical operations before autonomous medicine

Clinical judgment includes diagnosis, treatment, consent, uncertainty, and preference-sensitive decisions. AI may organize evidence or propose alternatives, but a qualified professional remains responsible. Clinical operations—collecting records, exposing prerequisites, routing tasks, monitoring completion, and escalating exceptions—are a better initial domain for agents.

Frontier research has demonstrated agents working in sandboxed EHR environments, but capability is not proof of readiness for unsupervised care. Nature Near-term deployment should therefore scale broadly across low-risk coordination before expanding consequential authority.

From assistance to verified completion

Authority progresses from retrieving and summarizing, to drafting for review, to executing reversible actions, to completing preapproved workflow steps, and finally to consequential decisions. Most organizations should concentrate on the first three levels.

Deeper authority requires prospective evidence for the exact workflow and population. Technical accuracy alone is insufficient; evaluation must include access, net work, completion, outcomes, equity, and economic return.

People need continuity across life stages

People encounter different risks and institutions across life. A reusable agentic layer can coordinate those transitions without treating maternal care, behavioral health, chronic disease, disability, and serious illness as interchangeable.

Every pathway still needs clinical and community expertise and a real service capable of responding. An agent can expose a missing maternity unit, therapist, medicine, or home-care service; it cannot replace one.

Give patients and families a coordinated front door

With permission, a personal health agent can assemble a sourced longitudinal record, explain benefits, prepare visits, find appropriate appointments, coordinate referrals, monitor results and medicines, estimate financial exposure, and help correct errors or appeal decisions. Success means fewer hours navigating care, faster access, completed follow-up, and greater confidence—not more digital activity.

Trust requires transparent incentives, data minimization, portability, accessible human service, and no hidden commercial steering. The agent may support navigation and financial protection, but it should not independently diagnose or make adverse eligibility and coverage decisions.

Return clinical attention to the relationship

Care teams spend substantial time reconstructing context, documenting visits, managing inboxes, completing forms, reconciling medicines, and chasing referrals. Ambient documentation provides an encouraging entry point: randomized and pragmatic studies found reduced documentation time or work exhaustion in some settings, with variable results and continued need for clinician review. NEJM AI randomized trial, pragmatic trial

A longitudinal practice agent can prepare the chart, draft sourced documentation, prioritize inbox work, reconcile medicines, assemble authorization evidence, and maintain referrals and abnormal results until completion. The governing rule is one review, one responsibility, and no hidden work: AI should remove net work rather than create longer notes, correction burden, or new alerts.

Primary, specialty, pharmacy, behavioral, dental, vision, rehabilitation, and home-care workflows require different evidence and controls. In every case, diagnosis, treatment, consent, and relationship remain with accountable professionals and patients.

Help hospitals operate as one real-time system

Hospitals coordinate staffed beds, tests, consultations, procedures, medicines, supplies, transport, environmental services, discharge, and external care capacity under uncertainty. Emergency-department boarding shows how a delay anywhere can affect the whole system and patient safety. AHRQ

An operations agent can maintain the shared state of admissions, beds, surgery, diagnostics, pharmacy, equipment, and post-acute acceptance. In surgery, it can identify missing prerequisites and cancellation risk. In laboratories and imaging, it can connect detection to accountable follow-up. The objective is safe care, fair prioritization, and sustainable workload—not throughput alone.

At discharge, the agent can verify medicines, equipment, education, transport, follow-up, and acceptance by the next team. Physical infrastructure—utilities, sterile processing, maintenance, and downtime capability—remains part of safe capacity and still requires trained people and investment.

Make coverage easier to understand and use

A governed payer or provider agent can retrieve current benefits, assemble an authorization packet, process bounded claims, explain status, and prepare an appeal. It should approve clearly qualifying requests before escalating exceptions and should measure time to appropriate care, accuracy, rework, affordability, and member experience.

Automation must not become an opaque denial channel. CMS guidance states that algorithms may support Medicare Advantage coverage decisions but cannot replace required individualized assessment. CMS Adverse action requires current policy, individual facts, an accountable decision-maker, a specific reason, and meaningful recourse.

Connect pharmacy access to therapeutic success

A prescription creates value only when the medicine is appropriate, covered, affordable, available, safely dispensed, understood, and monitored. An agent can reconcile the medicine list, assemble coverage evidence, locate supply, coordinate support and administration, and confirm the first safe dose and follow-up.

Pharmacists retain dispensing accountability, clinical intervention, substitution judgment, and counseling. Automation should reduce calls, status chasing, and duplicate reconciliation while making benefit and channel incentives visible.

Improve the medicine and device lifecycle

Scientific and development agents can synthesize evidence, use qualified tools, record experiments, test protocol feasibility, support trial operations, and maintain regulatory state. AI is already being explored across drug development, but value must be measured through reproducible learning, representative trials, qualified evidence, and effective products—not generated hypotheses or documents. Nature Medicine

Post-market agents can connect reports, clinical data, product lots, complaints, and exposure for expert review. Device and supply agents can link procurement, training, maintenance, quality, inventory, shortages, recalls, and replacement. Scientists, investigators, participants, regulators, quality teams, pharmacists, and clinicians retain authority.

Turn research and data infrastructure into a learning network

EHR, claims, pharmacy, laboratory, imaging, device, registry, and research systems each hold part of the picture. The goal is not to centralize everything, but to make the minimum necessary context available for an authorized purpose with provenance, explicit delegation, and revocation.

As agents gain the ability to act, identity, least-privilege tools, transaction limits, monitoring, downtime capability, and independent evaluation become essential. The Change Healthcare cyberattack demonstrated how disruption at one platform can affect national claims and cash flow. GAO

Fifteen priority workflows

The full model evaluates 28 workflows. This reader-oriented portfolio keeps the 15 most representative categories and the measure that should govern each one.

Three journeys show how the pieces connect

Chronic care. With consent, an agent prepares visits, reconciles medicines, follows tests and referrals, and makes meaningful exceptions visible to the clinician. The patient receives one understandable plan and less navigation work; success is better control with fewer crises and fewer unnecessary contacts.

Surgery. The agent maintains readiness across authorization, testing, preparation, staff, room, device, bed, transport, and follow-up. It identifies missing dependencies early and supports safe discharge; the surgical team retains clinical authority.

Therapy development. Scientific, trial, regulatory, manufacturing, and safety agents maintain traceable work across the product lifecycle. Accountable researchers and regulators approve consequential decisions; success is faster, more reproducible learning and timely access to a safe, effective product.

Across all three, the agent coordinates information and action while people and institutions retain judgment, physical care, resources, and accountability.

Incumbents are economic control points, not background infrastructure

The market sits inside a larger corporate and technical structure. Insurers may own PBMs, pharmacies, and provider assets; health systems own practices and ambulatory sites; distributors combine logistics, purchasing, oncology, and prescription technology; and EHR, clearinghouse, cloud, CRO, and diagnostic platforms control critical handoffs. CVS and McKesson illustrate how care, financing, pharmacy, distribution, and technology can converge inside large enterprises. CVS Health, McKesson

For buyers, the durable questions are straightforward: who owns the workflow and outcome; what the product may read, recommend, or transact; how it is paid; what evidence supports it; and whether data, logs, and operating state remain portable. Integration creates distribution and value, but it can also create steering, dependency, or conflicts that governance must make visible.

The startup opportunity is at the broken edge

The strongest startups will own a measurable handoff rather than sell intelligence in isolation. Durable advantages come from domain expertise, permissioned workflow data, trusted distribution, reliable integrations, outcome measurement, security, regulatory capability, and model or vendor portability.

High-value spaces remain in patient-controlled navigation, disability and caregiver support, Medicaid case completion, rural transport and referral networks, post-acute transitions, supply resilience, shared operations for small practices, independent agent evaluation, and identity, consent, delegation, and action provenance. The companion company diligence register provides the broader research view.

Build on the strengths of America’s food system

America combines abundant agriculture, advanced food science, efficient logistics, widespread distribution, and substantial choice. The opportunity is to make nutritious, culturally appropriate food easier to afford and use. In 2024, 13.7% of U.S. households experienced food insecurity; in 5.4%, eating patterns were disrupted and intake fell at times. USDA A small NIH randomized trial also found that adults consumed more calories and gained weight during a two-week ultra-processed diet phase, a useful controlled result that should not be generalized to every processed food or lifetime outcome. PubMed

With consent, agents can simplify nutrition-benefit enrollment, personalize options around budget and culture, support healthier institutional procurement, forecast demand, and verify that food support was fulfilled. AI improves coordination; income, benefit adequacy, product design, retail access, transport, and public policy remain the structural levers.

Support behavior without moralizing

Tobacco, alcohol, activity, sleep, stress, and social connection occur within addiction, products, environments, relationships, pain, and economic pressure. A useful agent helps a person choose a goal, understand options, obtain covered support, and reach a professional when risk rises.

Public-health agents can identify service gaps, support multilingual outreach, and measure reach. They should not create hidden behavioral profiles for employment, insurance selection, marketing, or law enforcement.

Strengthen vaccine evidence and trust

Vaccines have benefits, contraindications, common reactions, rare serious adverse events, manufacturing requirements, and continuing surveillance. The correct unit of analysis is the specific vaccine, population, and outcome. VAERS is an early-warning system and generally cannot determine causality by itself; signals require comparison data and expert study. FDA

AI can enrich reports, link product and timing, analyze active-surveillance data, prioritize expert review, check contraindications, and communicate known benefits, risks, and uncertainty. Trust comes from open methods, timely investigation, correction, accessible explanation, and respect for informed choice.

Treat prevention as a coordinated service

Agents can connect public health, primary care, schools, employers, pharmacies, community organizations, and households around a clear outcome: identify an evidence-based need, arrange the service, resolve practical barriers, monitor the result, and return the plan to the accountable professional.

This is a positive agenda—using America’s food, scientific, public-health, and delivery capabilities to help people remain healthy, functional, and independent for longer.

A cautious, driver-based scenario

The companion model applies adoption, performance, realization, outcome, cost, transfer, overlap, and evidence assumptions to 28 workflows instead of multiplying the projected $6.017 trillion 2026 total by a general savings percentage. Inputs are available in the updated workbook and methodology.

The Year 5 figures apply a uniform 2.99% prudence adjustment to that horizon’s workflow-adoption assumptions; Years 1, 3, and 10 are unchanged. The $121.4 billion year-ten base case is an annual value pool in constant 2024 dollars. It describes the modeled net value created by mature agentic-AI adoption through four economic pathways:

• Less waste — $55.47B: potentially cashable value that can lower operating costs, premiums, and public-program pressure or be reinvested.
• More healthcare capacity — $37.67B: usable workforce and asset capacity that can help clinicians and facilities serve more patients, reduce overtime, and improve workforce retention.
• Fewer avoidable costs — $12.21B: avoided medical expenditure that can free household, employer, insurer, and government resources for other productive uses.
• More innovation and resilience — $16.08B: research, operational, and supply-chain value that can support R&D, digital infrastructure, reliable supply, and faster product development.

The total is not all cash and is not a spending reduction, vendor forecast, or government budget score. Capacity becomes economic value when it produces more care, shorter waits, lower overtime, retained workers, or other measurable output. If organizations reinvest 25%–50% of the potentially cashable portion, that would provide approximately $13.9–$27.7 billion annually for workforce development, clinics, technology, cybersecurity, and research. Such reinvestment can support new and redesigned jobs while strengthening care delivery and innovation. Administrative value emerges earliest; clinical, preventive, supply, and scientific value grows as evidence and physical response capacity mature.

The percentage column uses the projected $6.017 trillion 2026 denominator throughout. Because the CMS projection is in current dollars while the modeled value is stated in constant 2024 dollars, the percentages are scale comparisons—not inflation-adjusted savings rates.

For context, an NBER paper estimated $200–$360 billion in annual savings from wider AI adoption, while a 2026 systematic review found the economic evidence too limited and heterogeneous to support aggregate claims in the hundreds of billions. NBER, Health Policy The model should therefore be treated as an auditable portfolio scenario.

Apply the transfer test

Before counting value, ask: Who gained, who paid, and did the person receive a better result?

• Payment or coding changes are social value only when they reduce rework or improve appropriate care—not when money merely moves between parties.
• Avoided unnecessary care and safe transitions can create value; denied needed care or unpaid risk shifted to families creates harm.
• Better enrollment may increase public spending, and supplier revenue becomes social value only when the product improves an outcome after total cost.

Benefits accrue differently across the system

The model allocates $19.2 billion of year-ten base financial or service-capacity value to federal programs and $10.9 billion to state and local government. The combined $30.1 billion is not automatic deficit reduction: better access may appropriately increase care, and capacity may be reinvested in service, prevention, cybersecurity, or research.

AI can also help government administer payment and policy, but it should not be credited with the full value of reform. For example, GAO estimated $141 billion in ten-year Medicare savings from aligning certain site-of-service payments; AI may support implementation, but the value comes from policy. GAO

The human dividend is the real objective

Economic accounts omit caregiver work, pain, uncertainty, travel, time on hold, lost independence, interrupted education and careers, and reduced participation. Every implementation should therefore report:

• survival, safety, health, and function;
• access, continuity, and financial protection;
• patient and caregiver time, agency, privacy, and trust;
• workforce burden and usable capacity; and
• distribution across the populations material to the workflow.

The pathway is the unit of proof: did the person reach appropriate care, receive the medicine, understand the plan, and experience less burden? Low-cost, persistent coordination can broaden support while reserving scarce human attention for relationship, judgment, empathy, and physical care.

Start with the right problem

Agents fit informational and coordination failures; they are weak substitutes for missing professionals, medicines, facilities, housing, or food. Every proposal should answer four questions:

1. What bounded failure, population, and outcome are in scope?
2. Is there an authenticated and legitimate path to act?
3. Is evidence proportional to consequence, and who could be harmed?
4. Can the organization observe success, provide recourse, and stop safely?

Protect recourse and human access

Automation must not erase responsibility. Affected people should know when it was involved, receive a meaningful explanation, correct information, and reach someone authorized to resolve the issue through an accessible digital, telephone, or in-person channel.

Agents must not fabricate evidence or consent, conceal steering, impersonate professionals, or use health data for undisclosed employment, advertising, or risk selection. They should stop when identity is uncertain, instructions conflict, evidence is missing, or behavior leaves the evaluated boundary.

Govern identity, data, and tools

Each agent needs verifiable identity, explicit delegation, minimum necessary data, least-privilege tools, transaction limits, provenance, action receipts, monitoring, and rapid revocation. High-risk steps may require an independent check or human release. NIST AI RMF, NIST agent identity

Because some consumer health data sits outside traditional HIPAA relationships, purpose limitation, retention limits, and meaningful consent should apply even where legal coverage differs. HHS

Evaluate the human-and-AI system

Benchmarks do not measure whether the response system worked. Evaluation should compare the new workflow with a real baseline and cover:

• task completion, accuracy, correction, and exception work;
• access, safety, clinical, and patient-reported outcomes;
• total implementation, response cost, and retired work;
• performance across relevant populations plus cybersecurity and misuse; and
• rebound effects such as more messages, orders, or low-value use.

Evidence should become more independent, prospective, and multisite as authority and scale grow. Material changes require re-evaluation, incident review, and the ability to pause or retire the system. WHO guidance likewise emphasizes human rights, independent audit, and risk-proportionate governance. WHO These controls allow trustworthy systems to scale.

Year 1: prove that verified work disappears

Begin with a few bounded, observable workflows such as visit preparation, documentation, referrals, authorization packets, scheduling, medicine reconciliation, discharge prerequisites, inventory visibility, or research search. Establish the baseline for time, error, delay, workload, cost, and access before deployment.

Redesign the workflow, limit authority, compare it with usual practice, and ask one question: which verified work, delay, or risk disappeared, and where did the recovered capacity go?

Year 3: connect institutional workflows

Connect related workflows inside organizations and with selected partners: visits with referrals and pharmacy, surgery with beds and discharge, benefits with computable authorization rules, and supply with quality signals. Shared identity, consent, delegation, monitoring, and incident standards should replace isolated departmental bots. Smaller and safety-net providers need shared infrastructure so access to these capabilities is not limited by scale.

Year 5: redesign roles, payment, and regional operations

Convert recovered time into access, safer staffing, less overtime, or stronger continuity. Payers and public programs should share administrative gains, support prevention and home care, and reward resilient supply rather than unit price alone.

Design workforce change by task: professionals retain judgment, relationship, teaching, and physical care while agents reduce search, drafting, status chasing, and reconciliation. Organizations should state which tasks end, how people are trained or redeployed, and how work intensification will be prevented.

Year 10: build a governed learning health network

The long-term objective is a federated network—not one national agent. Authorized agents should retrieve necessary evidence, use interoperable tools, carry permissioned goals, keep referrals and transitions visible, support real-time operations, and create verifiable transaction records across institutions.

Coverage remains contestable, science remains reproducible, and public agencies can evaluate access, outcomes, cost, and distribution. Consequential authority still advances more slowly than capability; the network makes accountability portable rather than removing responsible people.

What each leader can do now

Federal and state governments should make policy computable, establish identity and incident standards, modernize administration with appeal rights, and fund independent evaluation.

Health systems and clinics should create cross-functional governance and begin with a small portfolio of documentation, referral, medication, discharge, capacity, and supply workflows.

Payers and employers should publish current criteria, automate approval before denial, protect health data from employment use, and share gains with members and providers.

Pharmacies, pharmaceutical, biotechnology, and medical-device companies should connect access, safety, manufacturing, supply, and evidence while measuring reproducibility and validated progress.

Universities and research organizations should provide permissioned research environments, workforce training, and independent evidence.

Technology companies should design for bounded roles, provenance, portability, safe failure, measurable outcomes, and responsible exit.

Patient, disability, caregiver, and community organizations should shape objectives, accessibility, consent, appeal, evaluation, and the distribution of gains.

A Health Outcomes and Capacity Compact

The transition needs a shared scorecard covering access and outcomes; safety and recourse; patient, caregiver, and workforce time; total cost and financial protection; and equity, accessibility, trust, and resilience. Productivity gains should support patient-facing time, sustainable staffing, prevention, cybersecurity, independent evaluation, and lower burden—not simply higher quotas or vendor surplus.