AI care coordination and the future of patient engagement.

What is the problem Mila Health is trying to solve?

Large health organizations run on a complex web of software: EMRs, practice management systems, digital front doors, revenue cycle tools, and more. Each one handles its own swim lane well. The trouble lives in the space between them. Anything that falls outside a single system, or sits in the middle of several, tends to become manual work.

Preparing a patient for a procedure, following up afterward, reminding a member to schedule an annual wellness visit: today, most of that still depends on someone picking up a phone. Verma frames the cost of that manual coordination as a $334 billion problem across US healthcare. His view is that a large majority of this work can be handled by AI agents that are personalized to the patient, configured to the local setting, and guided by the provider.

Why not just build it on a large language model?

One of the sharper points in the conversation is the gap between a demo and a system that holds up at scale. Verma's caution to other founders is that it is possible to wrap a large language model into an agent that talks and texts and looks impressive in a proof of concept. Making it consistent, safe, controllable, and integrated into an enterprise stack is a different order of engineering.

That distinction is the foundation of how Mila is built. Rather than designing a product around a model, the company designed around the layered reality of how a health system or payer adopts technology, where every location carries its own requirements and every physician's office has its own way of working.

How does Mila fit without a rip-and-replace?

Verma points to two trends he sees accelerating. The first is that healthcare leaders no longer question whether AI agents should reach out to patients around the clock, in dozens of languages, to guide them through a care plan. That debate has largely settled. The second is a firm expectation about how it should happen: leaders want it to fit seamlessly into their existing software, with no complex integration project, and with return on investment measured in weeks rather than years.

Mila's answer is to give each provider a set of tools that connects up to the larger organization, integrates into existing software, and operates safely and compliantly, without asking anyone to adopt a new system to learn.

What does this look like in practice?

Verma shared a payer example from the conversation. Mila began with a roughly 30,000-member outreach program focused on setting members up for annual wellness visits with an in-network primary care provider. The payer's prior call center saw response rates in the range of 15 to 18 percent. In the conversation, Verma described reaching response rates above 50 percent, with the program now expanding toward 1.2 million members.

Just as telling is the texture of those calls. When Mila reaches an older patient who does not speak English and guides them through their care journey, the conversations sometimes run 15 to 20 minutes. People talk about their kids and their lives. For Verma, that combination of measurable reach and genuine human experience is the point.

The three questions Mila keeps asking

Across the discussion, Verma returned to three pillars that shape the work. Can the agent be consistent, safe, and configurable? Some of this is available through large language models today, and some of it is not, which is where much of Mila's engineering focus sits. Can it work simply, at scale? The goal is engagement that embeds across an organization without a months-long integration for every use case. Is it actually working? That means listening to the conversations themselves, asking whether patients pick up, whether Mila is empathetic, whether it brings a human into the loop when needed, and whether each interaction could be done better.