In enterprise healthcare, there are many “Jobs to be Done” (JTBD) for AI. To this end, we present the first part of the 6th episode of the Digital Health Go-to-Market Playbook series–Commercializing AI in Healthcare.
This piece, Part A, uses Clay Christensen’s Jobs to be Done lens, along with an assessment of viable product wedges and business models, to share what we see as the most promising applications of AI in enterprise healthcare. We also describe which use cases are more conducive to the application of generative AI in the form of large language models (LLMs) versus traditional machine learning (ML).
In Part B, we summarize interviews with key decision makers at leading payor and provider enterprises on what entrepreneurs should consider as they bring their AI products to market.
Enterprise healthcare tasks are well-suited to AI for two reasons:
Broadly speaking, we classify the opportunities for AI across two dimensions: the complexity of the task and the cost of a mistake. Enterprise healthcare tasks are highly complex and unforgiving to errors. Building specialist AIs to perform healthcare tasks offers the most challenging technical problems in the field today, as well as the greatest opportunity for impact.
Within the healthcare tasks, we further subdivide along two axes: clinical vs. non-clinical, and consumer- vs. professional-facing. Each box in the map below represents what we believe is a viable wedge for a company building opportunity, where taking an AI-native approach will confer a competitive edge, and where a credible business model exists.
In addition to the general factors mentioned above, what specifically makes a JTBD worthy to be on this opportunity map? We applied the following criteria to make that determination:
A top pain point for healthcare enterprises these days is workforce retention, especially among workers who require significant upfront and ongoing training (the average hospital turned over 100% of their workforce in the last 5 years, with labor costs growing approximately 21% over the past three years). These are areas where hiring “AI staff members” (versus buying software and mandating employees to change their workflow to use it; see third criterion below) has high potential for uptake. In addition, there tends to be a wide gradient of labor challenges across different domains; e.g. the revenue cycle department likely has more labor pressure than the marketing department.
Relatedly, areas in which humans are prone to error or are generally slow and inefficient (even when supported by software products) are most likely to benefit from AI approaches. For instance, in prior authorizations, a recent AMA survey found physicians and their staff spend approximately 14 hours per week completing PAs, while the GAO estimates that federal agencies made upwards of $128 billion worth of incorrect payments from Medicare and Medicaid to providers in 2022.
Healthcare enterprises are more likely to adopt AI if its cost benefit is at least an order of magnitude (and ideally more!) better than the status quo. Therefore, we’re likely to see a stronger opportunity in areas that have a low penetration of existing software tools, where AI cost benefit is being compared to human labor, versus software. For example, medical scribing is one of the areas where there appears to be high uptake of AI solutions because humans currently perform the majority of scribing tasks.
Healthcare is one of the only industries that has at least one established AI regulatory framework (e.g. FDA’s Software as a Medical Device, or SaMD, pathway, and the FDA 510K pathway). In the areas not covered by the existing regulatory regime, non-clinical products may feel safer to C-suites than clinical products. In the realm of clinical use cases, founders will have to account for the need and preference of the enterprise to have “humans in the loop” versus a fully autonomous system.
Some of the less-well-funded areas for health systems, such as social needs coordination, might find it harder to justify “hiring” AI than areas of higher, more specialized labor spend, or areas in which an AI product could simply ride existing revenue or reimbursement rails–either for substitute human services or software tools.
Note that some JTBDs did not make it onto the map based on these criteria, per the note in the graphic above (e.g. marketing, conflict of interest management, scientific journal/grant writing). That said, those task areas may rise in importance over time as technology, financial rails, and regulatory frameworks develop.
Per our healthcare AI thesis, we believe that the best builders of enterprise healthcare AI solutions will both understand how to exploit the latest advancements in AI, and, more importantly, how to commercialize a product with a durable go-to-market strategy. While each of the JTBDs described above are viable product wedges, we also believe that the winning companies will have multiple integrated products that perform a wide breadth of tasks, since healthcare enterprises are continuing to consolidate their vendor relationships and relying on each partner to cover a large surface area of use cases.
The healthcare industry desperately needs entrepreneurs building solutions to the scalability and cost structure problems that can uniquely be addressed by AI. Stay tuned for Part B, in which we hear from the buyers of enterprise AI solutions and discuss defensibility, pricing, and packaging of these AI solutions.
Jay Rughani is a partner at Andreessen Horowitz, where he invests in AI applications across healthcare and life sciences.
Daisy Wolf is an investing partner on the Bio + Health team, focused on consumer health, the intersection of healthcare and fintech, and healthcare software.
Vijay Pande is the founding general partner of the Bio + Health team at Andreessen Horowitz, focused on the cross-section of biology and computer science.
Julie Yoo is a general partner on the Bio + Health team at Andreessen Horowitz, focused on transforming how we access, pay for, and experience healthcare.