AI x Care Delivery: Healthcare Services as Software
By Andy Slavitt, Dr. Meera Mani, and Josh Loria
Generative AI uptake in healthcare has been a stunning contrast to other technologies, including EHRs, with 66% of physicians utilizing generative AI tools in 2024 – nearly double the number that used them in 2023. This exponential adoption in an industry notorious for slow adoption of new paradigms and limited labor productivity gains for decades indicates that generative AI has the potential to rapidly rewire healthcare in the coming 2-3 years. By providing healthcare services via software, which has an exponentially lower cost to serve, AI can help shift care delivery from being cost prohibitive to financially sustainable. This is especially true in the government funded programs of Medicare and Medicaid. Everything from how the healthcare system interacts with patients, the resources providers use to make clinical decisions, and what “value-based care” actually means – is being impacted.
This is why Town Hall companies, collectively serving 4 million Americans across the country, are embracing this technology. These companies are using AI to accelerate their growth and do so in a much more capital-efficient way – deploying multiple AI use cases across clinical and administrative areas. Here, we describe how gen AI is reshaping care delivery – and how care providers who implement it improve patient satisfaction, access to care, and margin.
How generative AI is reshaping care delivery
Today, the technology and processes that power care providers are centered around data storage and administration. Technology, including EHRs, e-prescribing, and revenue cycle management tools, has provided modest but highly variable benefits in health outcomes and productivity. Gen AI can change this status quo.
1. Ambient documentation platforms are becoming the “co-pilot” for provider <> consumer interaction
What it is: As ambient listening and documentation of healthcare conversations becomes widely adopted across languages, these vendors are beginning to take on pre-, during-, and post-visit activities for both clinical and non-clinical interactions.
Before visits, medical co-pilots can prepare chart summaries highlighting relevant context for the visit – and can assist in hand-off between different team members (e.g., nurse shift changeover in inpatient settings)
During visits, they make it easy for providers to accurately document clinical conditions, address patient issues, and close care gaps – all while bringing back greater eye contact and engagement between the provider and the consumer
After visits, these tools audit charts, provide accurate documentation for timely billing and authorization of future care, and handle follow-up care interactions with patients
One Town Hall portfolio company, Cityblock Health, is freeing up valuable care team time by utilizing AI to automated assessment completion and member outreach. Another, Thyme Care, is using AI to conduct quality assurance and quality control (QA/QC) on care team interactions, ensuring that members see a consistent experience that maps closely to best practice clinical guidelines.
Why it matters: Ambient data capture is already creating a meaningful productivity boost – scribes save physicians up to an hour a day. Now, co-pilots can expand into adjacent areas – reducing delays and inefficiencies in the care process.
2. Agents, agents, agents (+ robotic process automation)
What it is: Robotic process automation (RPA) handles rules-based, repetitive activities – like copying data from clinical systems into care management platforms or billing systems. AI agents, on the other hand, can learn and infer from context – and independently complete complex tasks like drafting tailored prior authorization appeal letters citing clinical evidence from chart notes.
Patient-facing voice and chat agents are already taking on common administrative activities such as scheduling and sub-clinical activities like remote patient monitoring; other agents are automating interactions with payors to handle routine activities in key areas like claims submission. We expect to see the scope of activities widen as many of the underlying foundation models improve, the depth of training data for common activities grows, and compute cost falls.
As an example, Zing Health, a special needs insurance plan focusing on diabetics in under-served communities, is using AI voice agents to conduct health risk assessments (HRAs) directly with new members, providing flexibility to the members (e.g., the HRA can be done at any time of day) and operational leverage for Zing.
Why it matters: For the last 3 decades, software has helped knowledge workers automate their work to deliver services more efficiently. With agentic AI, that paradigm is flipped on its head – rather than software helping humans complete a task, software is completing previously human-dependent tasks itself. Instead of talking about “software as a service,” we’re talking about “services as software.”
3. The “AI doctor”
What it is: The question many have is when can the co-pilot fly at least part of the flight. As AI models train on real-world data, as the interactions become more human, and as humans show increasing comfort interacting with avatars, many of the tasks clinicians currently perform will, in fact, be completed using technology. Bill Gates, in a recent interview, predicted that we won’t need humans “for most things” including “great medical advice” in the near future. While anything is possible in 10 years, we are better positioned to look at what we see as possible over the next 2 years. Town Hall and our portfolio companies are focused on the 150 million Americans who lack access to basic primary care, specialists when they need them, and high quality care management programs to manage chronic disease; it is against that backdrop that we see technology allowing greater care access for all Americans.
More and more sub-clinical and clinical activities will be directly delivered AI-to-patient as the AI<>patient interface improves, consumer comfort grows, and appropriate human-in-the-loop guardrails are developed. Expect to see this first deployed for narrow, protocol-defined conditions or pathways like diabetes and tasks like medication and prescription management. First line triage, diagnostics, and treatment applications, along with ongoing care management, are not far behind. Many jobs for AI are critical activities that physicians and their teams have not had adequate time to manage as panel sizes and waiting lists ballooned. Expect to see AI’s clinical scope expand meaningfully in the coming 12-24 months.
Why it matters: 60% of community health centers report difficulty obtaining new patient specialty visits for Medicaid patients; access and affordability challenges persist for the neediest Americans. The AI doctor can outperform the average human doctor at near zero marginal cost – available 24/7/365 – while acting on more data points and with greater longitudinal consistency. That means higher quality care for more patients in need.
As the use of gen AI in healthcare expands into co-pilot and services as software, it has the potential to fundamentally transform the healthcare stack. Some questions remain, though. For instance, what if the data needed to engage consumers in better preventive care increasingly resides outside the clinical system of record? Is it the optimal model for payor AI agents and provider AI agents to ‘interact’ with each other via natural language? What changes are needed to healthcare data labeling in order to maximally unlock productivity gains from gen AI tools?
The post AI care delivery organization
So, who will be successful at implementing the modern care delivery technology stack? We have seen several key characteristics contributing to success – (1) the willingness to experiment; (2) high levels of patient intimacy; and (3) data governance and security.
Experimentation
One important characteristic in leading AI-forward teams is a willingness to experiment. There are an immense number of administrative (e.g., coding / billing) and sub-clinical (e.g., pre-visit preparation) activities that lend themselves well to AI use cases – and the best organizations encourage AI fluency and usage for all team members in finding ways to improve even basic tasks.
Qualified Health allows healthcare partners to build, test, deploy, monitor, and scale AI applications. They provide the tools to train and enable healthcare staff to use LLMs in an intuitive way to reduce mundane and repetitive work. Health systems adopt a single product which ingests a system’s data from multiple sources and utilizes an AI processing layer to deliver insights across a wide range of use cases of the provider’s choosing – all while maintaining appropriate AI governance and protecting patient data privacy.
Patient Intimacy
AI can offer personalized care that is superior to the human-centric status quo in driving better patient outcomes and satisfaction. The best healthcare organizations are going to deploy tooling to improve patient interactions on the back of more context, more knowledge, and a richer understanding of a patient’s needs.
Arine, a medication intelligence platform, uses advanced AI models to enable plans and providers to serve the right intervention to the right member at the right time. Arine leverages generative AI to automate high-volume, low-complexity medication-related care management activities. At the same time, its adaptive AI continuously enhances clinical precision by incorporating new evidence and real-world insights into its knowledge graph — ensuring recommendations remain current, relevant, and highly targeted. As a result, Arine drives industry-leading uptake rates – over 50% of Arine’s recommendations get implemented – to improve patient outcomes and reduce overall total cost of care.
Data Governance, Privacy, and Security
US consumers are comfortable with gen AI in healthcare but only when closely managed – over 50% are neutral or comfortable with gen AI use for note taking and clinical decision support, but ~80% are uncomfortable with standalone AI use for treatment plans and prescriptions. And, according to United States of Care polling, 75% of voters want transparency into how AI developers and the health care sector is using AI within patients’ healthcare experiences.
Contrary to popular practice – doctors prompting personal instances of ChatGPT – the best organizations set up secure environments, adhere to a set of consistent practices on the labeling and use of data, teach providers how to prompt or query in a way that minimizes misleading answers, and maintain robust audit and explainability standards.
CIOs seldom have the influence over front line clinicians to teach, train, and police these practices. The best approach is to create environments that protect patient data by default, provide tools that are aligned to deliver answers consistent with how physicians use them, and provide a common reference point for data governance.
Over the past seven years, Town Hall has invested in companies with the ambition to transform care for those poorly served by the US health care system. We have closely watched policy and technology trends for opportunities to change the dynamic for millions of patients. We see generative AI similar to the way we see total cost of care payments and population health – as a mechanism for transformation. In the case of generative AI, thinking about important health care services as software can serve as an unlock to access to care, to deeper patient touch points, to more appropriate utilization, and to happier providers and patients.
And this innovation is within reach. The building blocks of AI are more accessible than ever. Healthcare technology should no longer be considered distinct from healthcare services – the two are becoming so interwoven that they will effectively become one-in-the-same. That’s the Healthcare Services as Software paradigm. Leading care delivery organizations are going to use AI and technology in every step of the care journey; leading healthcare technology companies are going to deliver practice capabilities that would have required too much labor to execute.
We have observed how this can work in early 2025 but expect the examples we are seeing will quickly be dwarfed by tens, then hundreds, then thousands of new uses that emerge over the course of this year and next. If pointed the right way, this will be the first technology advancement that feels good to physicians and carries real benefits for millions of patients.
