Could the VA Be the Key to Lowering the Cost of American Health Care?

Key Takeaways

Introduction

The Congressional Budget Office’s estimate that net interest on the national debt will total $952 billion in 2025¹ has given renewed urgency to calls for government spending reductions.² Health care is the largest category of federal expenditures, consuming 27 percent of last year’s federal budget,³ and an ideal candidate for smart reforms. Many other large and wealthy nations have significantly lower health care spending⁴ than the United States, often with superior results.⁵ The desire to reduce health care costs has been the subject of innumerable proposals, but the potential role of the Department of Veterans Affairs (VA) in achieving their savings goals is consistently absent. This omission, however, may become less common as artificial intelligence (AI) advances. The VA could provide a unique environment for resolving technical, operational, and strategic deployment issues that might limit AI within other health systems.

AI is one of the most promising tools for reducing America’s health care spending because, alongside its capacity to improve diagnosis and treatment decisions, it has the long-term potential to empower less expensive clinicians (e.g. nurse practitioners and physician assistants) to function at the level of more expensive staff as well as eliminate human labor in some cases.⁶ However, cost savings are not an inevitable result of increasing the clinical use of AI. First, a major reduction in health care spending requires a project to map AI solutions to the nation’s health care cost drivers and evaluate each individual mapping to determine in that instance whether AI has a prospect to reducing costs significantly. The mapping must consider not only AI solutions that are already in the market but also which solutions have a paradigm that can be reasonably extrapolated to a new context. Second, the same project’s candidates for AI-enabled cost savings must be operationalized so that they may be:

• validated or disproved,
• improved where necessary and possible, and
• emulated in deployment by other health systems.

This latter issue is particularly complicated with respect to AI. AI functionalities, and the outcomes they produce, are affected by their manner of deployment. Learning how to adjust protocols to maximize AI effectiveness in a clinical setting and identify when AI training data is not sufficiently aligned to serve local populations will be critical issues in realizing AI’s capacity for cost reductions. Thus, a clinical setting is necessary to validate (or disprove) candidates for AI savings. Not only does this setting need to pursue AI savings experiments iteratively—given the possibility of suboptimal deployment and the need for training data adjustment—but the health system must not have financial priorities in conflict with cost reduction. Moreover, the health care AI market is currently a collection of point solutions rather than a re-imagining of clinical workflows for the AI era. Likewise, AI must not be treated like a panacea free from the kinds of marketing hype and quality issues that have been found in non-AI medical devices.

This paper will demonstrate why the VA is a superb candidate to test how (and to what extent) AI can reduce health care spending. The VA, and specifically its Veterans Health Administration (VHA), operates at a national scale, controls clinical protocols, possesses data on both patients and claims, and faces no financial conflict in reducing expenditures because it pays for the care it delivers. These structural advantages make it an ideal platform for demonstrating the savings potential of health care AI tools and generating evidence that could extend these savings across the U.S. health care system.

The State of U.S. Health Care Expenditures

The Centers for Medicare and Medicaid Services (CMS) reported that U.S. health expenditures grew by 7.5 percent in 2023 (the most recent annual reporting) as compared to 2022.⁷ This increase raised the nation’s spending hundreds of billions of dollars from $4.53 trillion to $4.87 trillion, pushing its share of gross domestic product (GDP) from 17.4 percent to 17.6 percent.⁸ For comparison, CMS estimated health care spending to be 6.9 percent of GDP in 1970,⁹ and the nation’s current per capita spending is almost twice that of comparable nations.¹⁰

A major contributing factor to large health care costs in the United States is high medical labor costs. The average full-time primary care physician (PCPs) made $265,000 in 2022.¹¹ When specialists were considered along with PCPs, the average American physician earned $363,000 annually.¹² When compared to Europe, average U.S. physician earnings were 120 percent higher than the average in Germany, 189 percent higher than in the United Kingdom, and 278 percent higher than in France.¹³ Data does not suggest U.S. physicians earn more than foreign doctors because they see more patients. In fact, the average American physician sees fewer patients per week than do physicians in Germany, France, and Spain (though they do see significantly more than physicians see in the United Kingdom) based on survey results from the medical information service MedScape.¹⁴

This cost burden Americans bear for higher-expense physicians is inflated further by rapid consolidation among health systems across the country.¹⁵ A 2023 study from Harvard and the National Bureau of Economic Research found physician services provided by large, consolidated health systems were priced 12-26 percent higher than at independent physician practices, while hospital services averaged 31 percent higher in consolidated systems than in independent hospitals.¹⁶

Collectively these costs become even more problematic when viewed against a January 2025 report from the U.S. Treasury Department that showed the nation has an estimated $36.22 trillion gross national debt,¹⁷ which does not include a far higher amount in liabilities and unfunded obligations from commitments through programs like Social Security and Medicare. America’s outsized health care costs only exacerbate its deficit spending and debt-servicing expenses. This cost trajectory has been historically characterized by the Social Security Advisory Board as “unsustainable,” with the added warning that it represents “perhaps the most significant threat to the long-term economic security of workers and retirees.”¹⁸ Alongside this macroeconomic effect is the microeconomic: High health care costs inflate health insurance premiums for businesses and individuals. In the past 20 years, the cost of health insurance coverage for a single employee has risen 135 percent from $3,481 annually (2003) to $8,182 (2023)¹⁹ —which was approximately double the rate of inflation during that period.²⁰ Ultimately, the entire premium burden—including the employer share—is borne by the employee, as the employer share represents forgone wage compensation for the employee.²¹

AI as Cost Reduction Tool

In accord with Executive Order 13859²² promoting sustained government agency investment in AI research and development (R&D) for economic and national security,²³ the government should initiate a demonstration project that attempts a major health care cost reduction via AI. Regardless of the executive order, there is an additional motivation behind the administration: AI left on its own will become one more incremental expense in an already costly American health system.

AI itself is well-positioned to remedy industry overspending in ways competing tools of cost control (e.g. value-based care) cannot. Its range of capabilities include:

• reasoning and decision-making that can exceed human performance,
• “agentic” AI tools that can independently start tasks based on user-supplied objectives and choose how best to complete them,
• complex pattern detection that can identify the early manifestations of disease within medical images,
• statistical models that can predict future disease risk,
• devices whose functionality can improve after deployment in a health system based on data derived from ongoing use,
• natural language processing capable of understanding written or verbal information,
• faster execution than a clinician performing the same task, and
• task scaling at lower marginal costs than a clinician performing the same tasks. (Regarding the application of AI to clinical tasks see the article “Translating AI for the Clinician.”²⁴)

These capabilities have already been applied in diverse areas across health care, including medical diagnosis, radiology, neurology, cardiology, gastroenterology, anesthesiology, hematology, pathology, and surgery.²⁵ When such capabilities both satisfy clinical quality standards and provide greater productivity (such as reading thousands of medical scans in a day through a remote server rather than a few dozen a day by an on-site radiologist), there is a sound basis for investigating an AI system’s cost savings potential.

Unfortunately, these AI capabilities, while individually valuable, are unlikely to lower the cost of health care within a piecemeal implementation. Significant obstacles exist between provider-side savings and the transmission of those savings to Americans.²⁶

Foremost are the pre-determined payment rates that health insurance plans negotiate with health care providers. This network payment model does not dynamically adjust payments downward based on localized efficiencies within the larger group of contracted providers. Likewise, individual providers will not advocate for less compensation in circumstances where AI reduces their labor costs. The same financial interests prevent providers from purposely investing in technology that would reduce their payment per service. Additionally, standardized out-of-pocket fees a patient pays for in-network medical care do not encourage provider shopping and the downward pressure such shopping can exert on prices. Negotiating future billing increases to cover the technology costs for AI-assisted services, instead, is the most probable scenario for providers. They will seek increased compensation for both the costs of AI systems and the associated overhead costs involved in its deployment. Thus, if payment systems just add AI overhead fees, the new technology is more likely to raise prices than reduce them.

A Health Organization for Exploring AI-Facilitated Cost Reductions

Considering the obstacles to health care savings, the effort to reduce American health care costs through AI requires a project dedicated to that goal. Likewise, the project needs an appropriate health system to serve as a platform for operationalizing and validating AI-based cost reductions. It is actually the platform, not the AI scheme, that is more important for this project to be a success.

A basic framework (to be expanded and improved during actual practice) can uncover areas where the technology can lower costs—although it requires arduous analysis, problem solving, and creativity. At the highest level of abstraction, it would encompass the following activities:

1. Set a health care cost reduction goal (e.g., 15 percent) that must be achieved for the health system within a specific time frame without a material decrease in existing quality metrics.
   1. This cost reduction goal should not be combined with secondary goals because the utilization of supplemental goals risks the health system being able to fail at its primary goal but continue the project through favorable secondary metric performance.
   2. The goal helps prioritize which AI initiative proposals to implement.
   3. The goal and its timeline determine whether the return on investment for a particular AI investment is acceptable.
2. Categorize and rank health system spending by patient procedure or service and administrative expenditure.
   1. This categorization requires a volume of data that is statistically relevant and likely to resemble broad health care utilization trends across the country.
   2. This categorization requires data covering sufficient years to establish cost trajectories.
   3. This categorization requires data from facilities across the country to reflect regional variations in population health.
   4. This categorization requires a geographic distribution of spending that is sufficient to reflect regional variations in population health.
3. Review the list of spending for opportunities and map AI capabilities to spending categories where the technology can be applied and meaningfully reduce costs.
   1. This review requires the participation of health system clinicians and leading AI manufacturers.
   2. This review requires the participation of leading AI manufacturers and AI experts.
      1. AI manufacturers can discuss existing AI solution capabilities and areas where these capabilities may be extended to new health care functions and contexts.
      2. The extension of AI capabilities may require a regulatory sandbox that temporarily waives specific (not all) regulatory obligations for a promising class of products or services so its results may be evaluated in consideration of future evidence-based rulemaking for the category.²⁷
      3. Opportunities for the extension of AI capabilities to reduce specific categories of spending should be shared with the larger AI manufacturer and startup community to encourage collaboration.
   3. This review should include a concrete estimate for an AI solution’s cost reduction effect.
4. Establish the implementation considerations for each AI cost reduction opportunity.
   1. Implementation considerations include any patient risks, technology costs, data on AI performance, and changes to internal protocols and processes.
   2. Implementation considerations extend past deployment to the monitoring of AI performance and its achievement of cost reduction objectives.

With respect to the clinicians’ and AI manufacturers’ cost reduction estimates, the group should seek opportunities where AI can:

• reduce the existing labor effort (and attendant costs) expended by a primary care physician or specialist to perform a task;
• empower lower-cost labor (e.g., nurse practitioners, physician assistants, technicians) to perform a task ordinarily performed by a primary care physician or specialist;
• reduce or eliminate administrative labor currently performed by clinicians;
• eliminate the labor ordinarily performed by a health care worker through autonomous function performed inside or outside a clinical setting*;

• improve diagnosis (or patterns suggestive of a diagnosis) and/or treatment determinations to reduce waste due to misdiagnosis or suboptimal therapy and medication choice; and
• improve patient-dependent adherence to recommended medication usage, annual exams, vaccine/immunization schedules, etc.

For a health system²⁸ to serve as the platform for the spending reduction project, seven criteria are desirable. It should:

• Serve the patient volume equivalent of 1 percent (or more) of the U.S. population (3.4 million patients) annually,
• Serve patients residing across every U.S. state,
• Own and control its patients’ medical claims,
• Own and control its patients’ electronic health record (EHR) data,
• Control the protocols detailing patient care directions or how a procedure is performed by a clinician,
• Control administrative processes implemented in its organization, and
• Lack organizational goals and incentives that conflict with the goal of reducing health care costs

The first two criteria reflect a central concern regarding health care AI: its ability to “generalize”—that is to say, to ensure that its usefulness is not limited to the original data that trained the AI system. An AI system that can generalize will perform its functionality successfully for new patient populations where an AI is deployed.

[M]odel performance is heavily tied to details particular to the dataset the model was developed on—even slight deviations from the training conditions can result in wildly different performance. For example, when researchers trained a model to diagnose pneumonia from chest X-rays using data from one health system, but evaluated on data from an external health system, they found the model performed significantly worse than it did internally (Zech and others, 2018). The model failed to generalize (i.e., predict accurately) due to the shifts between the training conditions (health system one) and the deployment/testing conditions (health system two). These shifts are very common when moving a model from the training phase to deployment and can take a variety of forms, including changes in patient demographics, disease prevalence, measurement timing, equipment, treatment patterns, and more. Beyond contributing to poor performance, failing to account for shifts can also lead to dangerous decisions in practice: the system can fail to diagnose severely ill patients or recommend harmful treatments.²⁹

The advantage that a large-scale health system provides to the cost reduction project is that its patients are more likely to reflect the nation’s patient demographics and regional health trends and illuminate instances where AI technology does not satisfactorily generalize.

A health organization that is both provider and insurer provides access to medical claims and the EHR system with important patient data corresponding to those claims. The former allows for real-time cost tracking of savings initiatives related to AI, and the latter provides documentation of outcomes (and related issues) resulting from AI interventions. The EHR provides an additional value for the project: It has historical data that AI tools can analyze to identify the most cost-effective therapies and medications.

Just as important as billing and EHR data access is control over the organization’s protocols and workflow integration. A protocol is a formal code of rules and procedures that govern an instance of patient care. A health care system has a vast body of protocols to deal with all the care it is expected to provide—from addressing patient events (such as a cardiac arrest or stroke) to operational procedures such as scheduling prioritization for surgery or medical imaging. AI, when applied in the context of a medical service, is governed by a protocol. This protocol may extend beyond direct technology interaction with a patient to include staff training as well as oversight and audits. Likewise, the protocol may operate alongside multiple competing protocols pertaining to the physician and other medical devices and, thus, be integrated within a larger workflow. Having access to protocols, and the prospect to iteratively modify them, is a prerequisite for actualizing AI opportunities for AI-facilitated expenditure reductions, as the introduction of AI may dictate an adaptation to the protocol. Similarly, there is the matter of administrative processes. Like protocols, AI may require administrative process modification in order to perform its functionality and achieve the intended cost savings goal.

Finally, there is the matter of organizational goals and incentives. Organizational behavior is heavily influenced by these two conditions. A prerequisite for AI to lower health care costs is an organization that benefits from lower costs. In contrast, a hospital system whose performance, including for its key executives, is evaluated primarily on growing revenue may not have the adequate incentives to make the necessary investments and efforts to realize the full savings potential of AI. This is not to discount the possible contributions of for-profit health systems in the project of AI health care savings but to recognize that there are conditions that may not be aligned with the effort.

The VA as AI Testbed

The VA operates the VHA to finance and provide health care services to military veterans and qualifying dependents. The VHA may be the nation’s most promising—if not the ideal—candidate for a health care cost reduction demonstration project using AI.

The VA already has a chief AI director³⁰ and also hosts the National AI Institute, which “aims to improve the health and well-being of our Veterans and empower our workforce by harnessing the potential of AI and emerging technologies to deliver safe, effective, and trustworthy solutions.”³¹ The agency has already made an effort, through the Center for Care and Payment Innovation (CCPI), to effect data-driven care transformation and cost reduction.³² Moreover, the VA mission states that the agency will take actions to support public health alongside its other pursuits.³³ Finally, the VA meets key project requirements regarding scale, geographic breadth, deployment and protocol control, data access, and organizational incentives.

• Scale and geographic breadth: The VHA serves over 9.1 million veterans and operates 1,380 health care facilities across the United States,³⁴ which satisfies the project’s scale and geographic breadth criteria.
• Deployment and protocol control: At those facilities, the VHA can modify the protocols and processes surrounding the use of AI. These changes allow AI to be optimized to achieve project goals when initial implementation is promising but below either cost or quality expectations.
• Data access: The VHA effectively acts as both insurer and provider, which would provide access to EHR data and claims data. The combination of this data is critical for identifying opportunities for AI-enabled cost reduction as well as validating that the associated outcomes satisfy quality standards.
• Versatility: The VHA, as a federal agency, operates outside state mandates on AI as well as on insurance benefits. Such mandates could impair AI experimentation to produce health care cost reductions.
• Organizational incentives: Another advantage of the VHA is that the health system does not have a conflict of interest around reducing health care expenditures by virtue of its CCPI interest in lowering costs. Many nonprofit health systems, which outnumber for-profit hospitals in the United States,³⁵ may have misaligned incentives to lower costs because reduced revenue could threaten excess earnings that are often redirected into higher salaries or other cost-shift accounting practices.³⁶
• Transformation efforts already active: Under the new Secretary of Veterans Affairs, Doug Collins, the agency is exploring ways it can reduce wasteful spending.³⁷ Secretary Collins has stated, “We’re going to maintain VA’s mission-essential jobs, like doctors, nurses, claims processors, while phasing out nonessential roles.”³⁸

Other Project Considerations

Aside from the need for a formal government commitment to the demonstration project proposed in this paper, regulation and intellectual property practices can influence the project’s prospects for success.

Regulatory Policy Considerations for Project

A key policy consideration for the AI-driven effort to reduce health care spending is the potential for some medical services to be delivered entirely without clinicians. While AI will more commonly support lower-skilled clinical staff in performing higher-level tasks, there remains a theoretical—but important—possibility of full disintermediation in certain cases. Autonomous care, where a patient uses a self-service system without clinician involvement, is not entirely new: Pharmacies have long offered basic tools such as blood pressure machines and body mass index calculators for consumer use. What sets AI apart is the sophistication of its software and the expanded range of services it might deliver autonomously. Among all AI applications, fully autonomous medical services have the most significant potential for reducing health care costs. Accordingly, there should be a regulatory pathway—such as Food and Drug Administration (FDA) approval under the Software as a Medical Device framework³⁹—for AI systems to perform medical functions independently, provided they empirically meet three key criteria:

1. Accuracy levels (or patient outcomes, depending on the nature of the application) equal to or exceeding the average rate for clinicians performing the same function;
2. No amplification of health risks as compared to when the same function is performed by a clinician; and
3. Output communications that are comprehensible and actionable for the patient without the additional explanation from a physician.

A fuller discussion of beneficial AI regulatory principles and warnings against AI misregulation can be found in “Healthcare AI Regulation: Guidelines for Maintaining Public Safety and Innovation.”⁴⁰

Intellectual Property Needs

Decisions around intellectual property (IP) can either promote or constrain the rate of innovation within an industry. Positively, IP protections preserve the incentives for AI R&D investments, as a patent prevents competitors with more financial resources from replicating an invention and then leveraging mature distribution channels to dominate the market and starve an early-stage competitor. However, these same protections, if misused, can stifle advances and encourage the use of AI patents as legal weapons to bleed funds away from AI startups through patent litigation rather than protect the rights of innovators for their new products.

Although a previous Paragon paper⁴¹ addresses some important ways that the U.S. Patent and Trademark Office can promote robust AI invention and competition, the current discussion raises an additional issue concerning IP. As mentioned earlier, the mapping of AI functionality to health care costs may involve extrapolations of existing technology and models to new clinical contexts. For startups that manufacture such extrapolations for the health system testbed, the FDA should enact an accelerated approval path. The path should be consistent with its existing safety validation procedures. This accelerated process would be analogous to the current Breakthrough Devices Program,⁴² which is designed to help successful devices reach the market sooner.

Conclusion

AI provides an opportunity to address the high health care expenditures that have plagued the nation for decades. Reducing these expenditures may free up tens, if not hundreds, of billions of dollars for the public and private sectors and allow them to redirect these funds to more productive investments that would benefit the nation’s economy. Additionally, a successful cost reduction project would produce a paradigm that may be shared internationally, particularly in areas of low health care access where large medical bills can result in predatory debt bondage.⁴³

The health system resources associated with the VA have often been overlooked in discussions of American health care transformation and cost reduction. The rise of AI in health care, however, has brought new opportunities to the discussion along with needs that are best satisfied outside traditional commercial and academic health system settings. Given these circumstances, the VA may prove itself key in the quest to lower wasteful health care spending.

Footnotes