New AI-based software for interpreting echocardiograms sometimes performs better than trained professionals. A bionic pancreas with embedded AI results in better insulin control. 

AI is already affecting health care delivery, and the choices policymakers make about payment will define its future trajectory, says LDI Senior Fellow Amol Navathe. 

AI is already more scalable and varied in its actions than human services. But without policies that create financial incentives to improve care, we risk driving up costs without improved outcomes or stifling innovation, writes Navathe in a recent perspective piece in Health Affairs. Below, Navathe discusses how the current system is poorly fit for paying for AI and how drug pricing may offer some insights on how to pay for AI products. 

What motivated you to write this piece?

Navathe: AI is fast-evolving and entering all fields of medicine, and new technologies are at the forefront of discussions around health spending. When I was Vice Chair of MedPAC, we began considering how disruptive AI technologies are to how Medicare sets prices. Sita and I  wanted to join the ongoing policy conversation about how AI fundamentally upends our payment system’s focus on inputs such as time and skill.

The variability and the scalability of AI make it a poor match with current payment policy for clinical care. Can you give a few examples?

Navathe: Compared with human services, AI is more scalable—meaning that the incremental/marginal cost of providing one additional service is often small (sometimes even tiny) compared to a human performing it. Thus, once implemented, AI products can be used with large volumes of patients altogether. AI is also more variable in the actions it can perform, sometimes out of the patient’s view, and can open new categories of services that human clinicians do not provide or may not need to oversee. The variability necessitates a broader contemplation of payment schemes. 

For example, prescription digital therapeutics (PDTs) are FDA-authorized, software-based interventions designed to treat, manage, or prevent medical diseases and behavioral conditions, such as cognitive behavioral therapy delivered via apps. These are large-scale platforms in which each new user does not increase operating costs for either the prescribing clinician or the manufacturer. Prices are set by the manufacturer based on the perceived value of the product rather than the actual cost of delivery, which is quite low. And PDTs, because they are delivered on phones or tablets, do not fit into the traditional durable medical equipment pricing paradigm (think, oxygen tanks or “CPAP” machine for sleep apnea).

In these many situations where AI increases productivity without necessarily improving accuracy or outcomes, or when AI can perform services autonomously, current payment paradigms that reimburse based on human labor inputs like time and skill just don’t fit well.  This may lead to overspending and overuse, which must be balanced with the health benefits of access to the technologies.

Your piece discusses whether prescription drug pricing offers a pricing model for AI. What did you conclude?

Navathe: Prescription drugs offer a useful starting point for understanding the complexities of pricing AI technologies. The key common feature is that once the molecule is developed, the marginal cost of manufacturing and delivering prescription drugs is relatively low. For AI, the marginal cost is near zero. Prices set by manufacturers presumably reflect the time and resources dedicated to the research and development of these products, just like AI. However, they are also set to reflect what the market can bear, to maximize profits. To limit unnecessary product use, payers use cost-sharing to encourage clinicians and patients to use prescription drugs only when necessary.

This is where the similarities end, however. AI technologies are more difficult to manage uniformly. Since AI can adapt and change, setting fixed prices or payment rates that reflect AI’s health value becomes difficult. It has therefore become even more important that AI technologies meet a clinical benefit standard set by regulators such as the FDA or CMS that ensures they meaningfully improve health care delivery.

What are the shortcomings of the current payment mechanisms used for AI?

Navathe: Currently, hospitals are paid a fixed price under the Outpatient Prospective Payment System (OPPS) and the Inpatient Prospective Payment System (IPPS) based on the health care services delivered, including AI-assisted services. Technologies also receive supplemental payment through the transitional pass-through payment system for outpatient services, or the new technology add-on payment (NTAP) for inpatient services. One requirement for NTAP is that the new technology demonstrates substantial additional costs, which may incentivize developers to set high prices. However, these payment designations are temporary and do not weigh clinical improvement against cost, the central challenge in developing appropriate reimbursement pathways for AI technology.

Another route is to garner a new code on the Medicare Physician Fee Schedule. This is where the reference point for labor inputs looms large, and we do not yet have a good system for valuing new technologies. Once again, the clinical benefit standard is likely a critical factor to incorporate down the road.

How should federal officials consider care and/or cost effectiveness for AI technology in making coverage and payment decisions?

Navathe: Policymakers and thought leaders are grappling with this right now. There are custom pathways that CMS could use, like Coverage with Evidence Development, a Medicare policy that provides coverage for a promising but unproven service only if patients participate in approved clinical trials or registries. That may be the only current way to factor in clinical effectiveness. However, these have traditionally been used infrequently because they are administratively complex and can take time to reach a true coverage and payment decision. “Getting a CPT code” is another mechanism, but it does not rely on a clinical effectiveness standard. Given the constraints on Medicare and CMS by law, Congress may need to grant CMS new authority to factor in clinical effectiveness in AI technology coverage and payment decisions.

---

“Aligning AI Payment Policy With Desired Outcomes Rather Than Inputs May Require Customized Pathway” appeared in the January 2026 issue of Health Affairs. Authors include Sita K. Kottilil and Amol S. Navathe.

---

Author

---