Welcome to Part III of this series on payment models for healthcare AI. After a brief hiatus (medical board exams, summer travel, moving), I’ll be tackling a massive backlog of topics over the coming weeks.

First up – finally finishing this series... To receive these posts directly in your inbox, subscribe below!

In this article, we’ll define five performance-based pricing models, explore the strengths and weaknesses of each approach, and discuss various examples of how these models are deployed in the wild.

If you haven’t read previous posts in the series, check out:

• Part I, which explains why payment models are holding healthcare AI back

• Part II, explores five usage-based pricing models

The five performance-based pricing models discussed in this piece are:

• Performance-based Linear

• Shared Savings

• Hybrid Shared Savings

• Discounted Fixed Costs

• Dynamic Reimbursements

In the final installment of the series (Part IV) of this series, I’ll be wrapping up with additional considerations pertaining to the current landscape for healthcare AI pricing models.

Performance-Based

Performance-based pricing models are those that reward a vendor for achieving a given metric by allocating a share of upside delivered by the product or service.

As discussed earlier in the series, there is growing skepticism that healthcare and life science organizations will be willing and able to pay exorbitant prices for AI products or services unless there is a clear return on investment (ROI). Irrespective of modality, this dynamic poses challenges for AI vendors who face high cost structures at this point in the cycle.

Over the last few decades, many in the healthcare industry have championed value-based care as a panacea to rising costs of care and deteriorating outcomes. It’s been a long journey, and it appears performance-based models are now proliferating at a rapid clip. In 2022, 41.3% of payer payments were made through advanced payment models, with 24.5% flowing through risk-based models specifically, up +20% from 2021. There is now near-ubiquity of awareness for performance-based models, with a non-trivial percent of organizations actively engaged. In the midst of the AI hype cycle and in light of the tech sector’s ambitious plans for healthcare AI, there's a compelling case for AI vendors to adopt performance-based pricing strategies that more closely align vendor incentives with customer and end user needs.

In healthcare, performance-based models are typically measured against specific benchmarks, including:

• Clinical: Measurable improvements in specific health outcomes for a defined patient population (e.g., reducing the average A1c of a cohort of patients with Type 2 Diabetes)

• Financial: Quantifiable reductions in healthcare costs for targeted patient groups or conditions (e.g., reducing the total cost of care for oncology patients at the end of life)

• Engagement: Trackable increases in patient-provider interactions or program participation, often used as early indicators of potential clinical or financial improvements (e.g., tracking interaction frequency with congestive heart failure patients). Note: engagement-based benchmarks are often earlier proxies for downstream clinical and financial outcomes, but can be helpful for early-stage companies that do not have sufficient data to claim actuarial-grade improvements in clinical or financial metrics.

• Scientific: Verifiable enhancements in the efficacy or safety of medical interventions through laboratory-based metrics (e.g., reducing the immunogenicity of a development candidate by some defined assay)

Generating actuarial-grade evidence to demonstrate ROI against the aforementioned benchmarks remains one of the greatest challenges for companies operating under performance-based frameworks. Like their usage-based counterparts, performance-based models face a unique set of challenges:

• Negative working capital dynamics

• Unpredictable revenues

• Vulnerability to externalities

• Complex, individualized contract structures

• Costly adjudication processes

• Attribution difficulties

Before diving into the five types of performance-based models, let’s outline each of these challenges in a bit more detail. To skip to the pricing models, scroll below.

Negative working capital dynamics: Under performance-based models, AI vendors must go live with a customer before receiving some or all of a given payment, which means they have to invest substantial resources prior to receiving a performance-based payment. Depending on the structure of the performance benchmark, this process could take weeks, months, or even years for a given outcome to materialize, during which time the vendor must fund its own operations including deployment of the intervention under analysis. Fortunately, there are alternative structures that can mitigate these negative working capital dynamics discussed later in the piece.

Unpredictable Revenues: Many performance-based models are defined as paying a percent of value generated (e.g., financial savings); therefore, it can be difficult to forecast these numbers, especially in the early years of a company generating initial data underpinning a product or service. Actuarial analysis that maps a given AI intervention to downstream clinical and financial ROI is part of the answer, but as we know: all models are wrong, only some are useful. Greater predictability in performance-based arrangements comes with time and experience, deeper understanding of customer phenotypes, and mapping of levers for intervention. Of course, this dynamic can also deliver favorable surprises, such as when the delivered ROI is better than forecasted, yielding greater value capture for the vendor and customer.

Vulnerability to Externalities: A highly performant AI product backed by a compendium of robust research will fail to deliver ROI if it is not successfully integrated into technological and operational workflows. To be more specific, if a system delivers an intervention to an end user (e.g., a clinician) who cannot act upon that information at the point of receipt, that individual is almost better off without the tool altogether. While methods development will continue to drive step-function improvements in our AI capabilities, implementation science will remain a critical component of capturing the utility of these novel capabilities in the wild. Furthermore, under certain pricing models, customers may be highly incentivized to cap the value generated by an AI solution so as to mitigate exorbitant costs under a performance-based model. It is essential that customers and vendors collaborate on contract structures that disincentivize perverse behaviors that could inhibit a solution from maximizing value delivery.

Snowflake-like Contract Structures: You know what they say. When you’ve seen one performance-based contract, you’ve seen one performance-based contract. It is rare to see two contracts that look the same (i.e., have the same performance measures, shared savings arrangements, payout structures, timelines, or attribution processes). This dynamic can be messy for early-stage companies just trying to get paid, and difficult for investors to value on a forward-looking basis.

Costly Adjudication: Related to the aforementioned point, a plethora of contract structures can be difficult to manage, and therefore costly to adjudicate at the end of a performance period. Unsurprisingly, there is a massive consulting sector that profits from these difficulties! There are green shoots though, as new companies are popping up to address these challenges and make it easier to manage, assess, and ultimately adjudicate performance-based arrangements between multiple parties (e.g., Arbital Health). It begs the question of whether there is an opportunity for these new solutions to build capabilities to assess AI ROI, and how that maps to ongoing initiatives such as assurance labs.

Attribution, Attribution, Attribution: Like any intervention in healthcare, it’s difficult (and often contentious) to prove causality between an intervention and an outcome. For example, an AI company may argue that its technology was solely responsible for a reduction in surgical cancellation rates, but the health system may argue that were a number of other operational and technological initiatives that also contributed. Any claims made that a given AI intervention produced an outcome will be heavily scrutinized under performance-based models, and companies must be prepared to defend these claims with robust data and statistical analyses.

As we explore each of the five performance-based pricing models for healthcare AI, we'll see how they attempt to address these challenges while creating value for both vendors and customers. Each model ultimately explores tradeoffs among balancing risk, capturing value, and aligning incentives.

Performance-based Linear

Definition: flat rate per unit of value generated against determined criteria.

Under Performance-based Linear models, AI vendors do not receive payment unless they have successfully completed an agreed upon task or delivered on an agreed upon outcome. This pricing model works better for specific premium features, but can be challenging for subsidizing the entirety of a business; however, when tied to specific outcomes, this pricing model closely aligns value generated.

Example: A biopharma company pays per successful enrollment for an AI algorithm that can identify patients eligible for a given trial.

Below is a non-exhaustive list of the advantages and disadvantages of performance-based linear pricing models. Note that attribution challenges, negative working capital dynamics, and vulnerability to externalities are key disadvantages of this model from the vendor perspective.

Shared Savings

Definition: charge a share of financial returns generated by a solution.

Ah, shared savings – a pricing model familiar to anyone in value-based care delivery. Under this model, the financial savings generated by a given AI intervention are split via an agreed upon ratio between the AI vendor and the customer. For customers, shared savings models are a fantastic way to explore a novel solution without having to make a direct financial investment upfront; instead, customers invest time and team resources to stand up the initial collaboration. Shared savings models can be wildly performant in the first few years where an AI vendor can capture value from “low hanging fruit”; however, as the customer evolves and invests further in technology across the organization, the “savings to be shared” can dwindle.

I predict this phenomenon will play out in the clinical documentation space. As the quality of clinical documentation improves and standardizes with AI scribes, value capture facilitated via downstream initiatives spanning coding optimization, clinical documentation integrity, and denials management may dwindle. As an analogy, imagine searching for coins in between couch cushions after you’ve just redesigned the couch to be one continuous piece. Naturally, the essential question underpinning this prediction is: over what time horizon does the industry adopt AI scribes?

This example reveals that when deployed in isolation without other fee structures such as monthly recurring fees, shared savings models can pose challenges for companies seeking predictability in revenue forecasting. There are many ways to mitigate these risks through customer profiling, use case refinement, and close ongoing monitoring – all of which require close partnership with customers, a robust customer success function, and actionable interventions should course correction be required. Data generated through customer profiling activities can be utilized to forecast projected savings and value capture on an account basis.

Here is a great example ROI calculator from Abridge, which empowers customers to enter various metrics to estimate outcomes.

Example: AI denials auditing for providers that captures 20% of additional revenues recouped from health plans.

Below is a non-exhaustive list of the advantages and disadvantages of shared savings pricing models. Note that attribution challenges, negative working capital dynamics, and customer operational efficiency are key disadvantages of this model from a vendor perspective.

Hybrid Shared Savings

Definition: recurring subscription fee plus a share of financial returns generated by a solution

Hybrid Shared Savings models mitigate some of the negative working capital risks of a pure shared savings model while allowing for uncapped upside. Under these models, companies capture some percentage of the deal value upfront in the form of a recurring fee, offering cash flow to fund deployment and ongoing resourcing of the account, while also maintaining an incentive to deliver strong value under the shared savings arrangement.

Companies deploying this model must navigate deal terms carefully to ensure that there is enough upfront value capture (in the form of recurring fees) to fund operations, while ensuring there is enough skin in the game for the customer to feel like they’re paying for high performance.

Example: AI claims auditing for providers charges a base fee of $100k per hospital per year plus 10% of revenues recouped from health plans.

The relative advantages and disadvantages of hybrid shared savings pricing models borrow from the performance-based linear and shared savings models described previously.

Discounted Fixed Costs

Definition: recurring cost with discounts if certain clinical, financial, scientific, or engagement-based outcomes are not met.

Discounted Fixed Costs mitigate all risks associated with negative working capital issues for the vendor given the upfront payment model; however, this model doesn’t incentivize strong alignment between the AI vendor and customer. In fact, the customer is incentivized to argue why the AI vendor did not achieve specific outcomes such that discounts can be applied (think about conversations you had with your landlord moving out of that fun group living situation in your early 20’s). Nevertheless, it is critically important that the clinical, financial, scientific, and/or engagement-based outcomes are directly influenced by the AI solution, not one or two-times removed.

Consider the following example: a value-based care company pays based on a discounted fixed cost model for an AI algorithm to detect certain opportunities for intervention in a population (e.g., identifying patients with heart failure not on therapy). The customer recoups a percentage of the fees if the intended clinical or financial outcomes are not met during adjudication (e.g., reduction in admissions for heart failure patients). In this case, being on appropriate heart failure therapy significantly reduces but does not guarantee that a heart failure patient will not be admitted.

Below is a non-exhaustive list of the advantages and disadvantages of discounted fixed cost pricing models. Note that the payback requirement, challenges in margin forecasting, and potential for disputes over outcome assessment are key disadvantages of this model from the vendor perspective.

Dynamic Reimbursements

Definition: higher per unit or percentage costs if certain positive outcomes are demonstrated.

Dynamic Reimbursement models charge a unit fee based on achieving certain performance-based benchmarks akin to the Performance-based linear model discussed earlier in the series, but also includes an opportunity for the AI vendor to charge more based on performance. The escalation of the fee schedule could be linear, logarithmic, or exponential depending on the use case and how value accrues for the N+1 positive result.

Example: An AI vendor that is focused on detecting fraud, waste, and abuse charges more after certain thresholds are achieved, such as identifying $100k, $250k, and $500k worth of charges.

Below is a non-exhaustive list of the advantages and disadvantages of Dynamic Reimbursement pricing models. Note that the challenged attribution, difficulties predicting revenues, and implementation and monitoring costs are key disadvantages of this model from the vendor perspective.

Performance-based pricing models for healthcare AI offer a range of options to align incentives between vendors and customers. Each model comes with its own set of advantages and challenges, and the choice of model depends on factors such as the specific use case, risk tolerance, and benchmark-based goals of both parties.

In the final installment (Part IV) of this series, we'll explore additional considerations and key takeaways on healthcare AI pricing models, tying together the insights from all parts of the series.

Stay tuned for the conclusion, and don't forget to subscribe to receive the full series directly in your inbox!

Thanks for reading!

Morgan

Huge thanks to my pals Nikhil Krishnan and Gaurav Singal for their feedback on this piece.