Who pays for healthcare AI? In this four-part series, I explore various business model architectures, pricing frameworks, and key questions for the healthcare and life sciences industries in the era of large-scale AI.

This first post covers:

• How current payment models are hindering healthcare AI adoption

• Why business model innovation is required for AI companies to capture sufficient value

• A quick history of healthcare AI reimbursement from CPT to SaMD to NTAP

• Two business model architectures for healthcare AI: Usage-based and Performance-based, and a preview of ten pricing models

The second and third posts will describe Usage-based and Performance-based business models respectively, articulating five pricing models within each type.

The fourth post will walk through additional considerations of AI pricing models including ROI, the great FFS vs. VBC debate, and key questions for the industry to answer.

Current payment models are hindering healthcare AI adoption

I recently came across a shocking statistic: the average single-hospital IT budget totals just $7.8 million annually, which includes software, medical devices, imaging equipment, hardware and networking components, cybersecurity, and salaries for IT personnel.

This figure confirmed a gut feeling I’ve had for a while: current software budgets for providers are unlikely to support ten $10 billion dollar software companies ($1 billion in revenue, assuming 10x ARR multiples) today… or at least not without significant business model innovation that facilitates greater value capture by tapping into wallet share outside of the current IT line item.

The math is clear: with 6,129 hospitals in the US, a software company would have to garner an average contract value of ~$160,000 at every hospital to eclipse $1 billion of revenue (this is theoretical – of course, contract values would vary by customer size and budget). Many were shocked to learn that Epic, the leading electronic health record in the US, generates just $4.6 billion of revenue, despite wielding power and influence whose impact extends far beyond what any number can capture.

Of course, these figures vary wildly based on the setting: academic vs. community, metropolitan vs. rural, or single hospital vs. health system. The hospitals with the highest IT expenses below underscore the variance in these figures.

Regardless, the point stands – based on these data, there is a ~$47.8 billion provider IT TAM in the US today, ~10% of which is spent on the Epic EHR, and a separate, small fraction of which is spent on other software products today.

$7.8m annual IT expense / hospital x 6,129 hospitals = $47.8 billion annual provider IT spend

All told, it’s an unsurprising reality that providers cite cost as the biggest barrier to AI adoption, especially considering that the software segment of healthcare IT budgets remain immensely under-resourced relative to the magnitude of the opportunity in AI. Of course, payer and biopharma IT budgets for software and computation are growing; however, many opportunities for AI – spanning RCM to novel biomarkers – often require provider alignment and buy-in to achieve success. This point relates to a fundamental truth – the best healthcare and life sciences businesses unlock opportunities across multiple stakeholders.

All hope is not lost. There are several ways the healthcare sector can increase addressable spending and therefore the TAM for software and AI across providers and beyond. First, each prospective customer can simply spend more (tough ask, but providers appear to be trending this way based on a recent 2023 survey by HFMA and McKinsey report on healthcare profit pools). Next, we can re-architect operational expenses previously allocated to other modalities, such as human labor and other services, to redirect spending towards AI. More on the specific business and pricing models that enable this shift later in the series.

So the question remains: now that AI has taken center stage, with 70% of healthcare organizations having moved to the cloud, how can AI companies capture the value of their technology in such a “resource-constrained” industry? What business models maximize AI value capture while simultaneously aligning with the financial, clinical, scientific ROI delivered by a given intervention?

Quick History of AI Reimbursement in Healthcare

Payment for AI in healthcare (i.e., provider/payer world) was largely experimental (e.g. pilot programs, grants, etc.) until the late-2010’s. Historically, AI business models have been bundled (e.g., a software product leverages AI, but charges via a typical SaaS or hardware pricing model).

Development of more robust reimbursement mechanisms largely matured in parallel alongside the key ingredients for healthcare AI: growth in compute, curation of digitized data, and scaling of model sizes. As more AI solutions came online, a unified approach to characterizing these solutions became essential. In 2018, the US FDA pathway for Software-as-Medical-Device (SaMD) became a cornerstone for regulatory oversight of AI devices across product lifecycles.

In 2020, US Centers for Medicare & Medicaid Services (CMS) announced the first AI-specific CPT code, which would allow for direct billing and reimbursement of AI products. The code, 92229, was for IDx-DR, a cloud-based algorithm developed by Digital Diagnostics that detects diabetic retinopathy using images captured via a retinal camera.

In the same year, CMS announced that AI devices would be eligible for New Technology Add-On Payment (NTAP), a program initially introduced in 2001 focused on supporting timely access to innovative therapies for the Medicare population. For technologies accepted to the NTAP program, CMS provides an additional payment to hospitals above the standard MS-DRG payment amount. Viz.ai’s large vessel occlusion algorithm was the first AI-based product to be accepted by NTAP for a three-year period. In the arrangement, NTAP adds a maximum of $1,040 to a hospital’s payment for managing a stroke episode for the next three years, which is meant to cover the cost of operating room use, nursing, supplies, as well as laboratory and imaging services. To be accepted to NTAP, the technology must exhibit:

• Newness: the technology is less than three years old (excluding FDA 510k clearances predicated on other technologies)

• Cost: the technology is not adequately covered under the existing MS-DRG

• Substantial clinical improvement: the technology must demonstrate to CMS that it provides a substantial clinical advantage over other available technologies, typically via improved patient outcomes.

As of the 2024 fee schedule, applicants for NTAP must receive FDA marketing authorization as an AI-enabled device under the software-as-medical-device pathway prior to applying. Below is a graph showing the historical number of NTAP applications by type (device, drug, diagnostic) and approvals.

The news that AI companies could receive NTAP designation was initially met with enthusiasm – historically, government-funded payers design healthcare payment policies and the private payers follow. However, reimbursement for AI hasn’t been this straightforward. The program's brief duration (specifically, a maximum of three years for reimbursement) prevents the NTAP from completely resolving the payment model issue. For now, NTAP appears to be a modest regulatory-grade incentive structure (falling short of a “Meaningful Use Moment”) for driving short-term adoption of AI products in healthcare.

Business Model Architectures for Healthcare AI

Business models for healthcare AI can be divided into two architectures: 1) usage-based and 2) performance-based.

Thus far, most of the business and pricing models deployed by healthcare AI companies have been usage-based and have not differed materially from traditional cloud companies (save for couple historical examples mentioned previously and later in the series). This is due in large part to the fact that AI products have been delivered within cloud modalities: SaaS and APIs. These business models – from seat-based pricing to API calls – are now familiar to the CXOs of healthcare and life sciences organizations.

I expect a high percentage of healthcare AI companies to continue leveraging these well-trodden cloud business models, such as those detailed in this Bessemer report, with a few modifications for the healthcare and life sciences industries. In addition to SaaS and APIs, healthcare AI companies also utilize tech-enabled services frameworks as a delivery modality, with unit-level value creation measured via API calls, per seat fees, and per completed task fees. On the other hand, the healthcare industry is moving in the direction pay-for-performance, which represents another pricing model for AI technologies.

In Part II of this series, I’ll explore ten pricing frameworks for healthcare AI spanning usage- and performance-based models, highlighting real-world examples and the relative advantages and disadvantages of various structures:

Subscribe to receive Part II for free directly to your inbox:

Thanks for reading!

Morgan

N.B. There are many other barriers to large-scale AI adoption in healthcare not discussed in this series including validation and monitoring, safety, and technical infrastructure, and more.

Huge thanks to my pals Nikhil Krishnan and Gaurav Singal for reading drafts and giving tons of thoughtful feedback on this piece.