Company

Building a World Model for Healthcare

Jonathan Kolstad, Co-founder and CEO, Knit Health

·

November 16, 2026

Today, we're announcing Knit Health's $35 million Series A, led by NEA, with participation from Frist Cressey Ventures, Town Hall Ventures, and SignalFire and continued support from Uncork, Moxxie, the Aligned Fund and Coalition Operators.

Medicine has become extraordinarily sophisticated in what it can do for patients. Advances in medical science, new diagnostic and therapeutic technologies, and increasingly capable AI agents are expanding those capabilities further. There is, however, a problem that receives far less attention. As we develop more sophisticated and specific ways to treat patients, how do we figure out which of them to use, for whom, and when?

Consider how healthcare operates today. We have built an extraordinarily specialized system, with clinicians, facilities, and technologies capable of treating an enormous range of conditions. Yet the mechanisms for connecting patients to these capabilities remain surprisingly primitive. Referrals depend heavily on professional relationships. Scheduling relies on rigid rules and appointment availability. Clinical pathways reduce complex decisions to simplified categories and scores that often fail to capture what matters for an individual patient.

As a result, we routinely fail to make full use of the capabilities we already have. Patients face long waits and often receive care poorly suited to their needs, while clinicians spend time on work that others could do more effectively. Consider a surgeon whose clinic is filled with patients who don't need surgery, while patients who could benefit from an operation wait weeks or months to be seen. Or a patient who could benefit from an innovative new therapy but never reaches a specialist familiar with it. In both cases, the expertise and treatments exist, but we struggle to connect them to the patients who need them.

This problem will become even more consequential as AI agents begin to participate directly in delivering care. They will be able to communicate with patients, review medical records, coordinate appointments, and carry out increasingly complex tasks, such as autonomously ordering diagnostics. But to be genuinely useful, AI needs more than textbook medical knowledge or the ability to execute individual tasks. It needs to understand the clinical context in which it operates; what a particular patient needs at a particular moment, how its actions might affect that patient's trajectory, and how those actions fit into the broader delivery of care.

We founded Knit to solve these problems. We envision a world in which clinical intelligence is no longer scarce, and every patient can be connected to the care best suited to their needs, at the moment they need it. We believe that making this intelligence available at scale will not only change how patients access care, but how healthcare itself is delivered.

Learning how healthcare actually works

When we think about intelligence in healthcare as the ability to reason dynamically about a patient and navigate them through diagnosis, treatment, and ongoing care, the approach to AI becomes fundamentally different. It has much more in common with the physical world models used to navigate self-driving cars, guide robots, and understand complex physical processes than with models designed primarily to understand language. That is why we’re building the first world model of healthcare. Our world model, Med-BEAM, has learned how patients move through the healthcare system: how their conditions evolve, what clinicians actually do, what happens after different interventions, and how those patterns vary across patients and settings. Embedded in those patient journeys is an extraordinary accumulation of human knowledge and experience. Clinicians make complex judgments every day, drawing on scientific evidence, years of practice, and an understanding of individual patients that cannot be fully captured in the medical literature or guidelines.

This is a radically different approach from the large language models that have driven much of the recent excitement around AI. LLMs are remarkably good at understanding and generating language, and completing language-based tasks. They can partially synthesize medical information, answer questions, summarize records, and help clinicians communicate. But knowing what the medical literature or guidelines say about a condition is not the same as understanding what actually happens when a real patient is treated by a real care team in a real healthcare system.

By combining an understanding of how healthcare actually works with an unprecedented scale of clinical data, Med-BEAM represents a fundamentally different approach to AI in medicine. On benchmarks predicting real clinical events and care decisions, a clinical Turing test, Knit's world model performs approximately 4 times better than leading language models. That level of performance lets AI do more than answer clinical questions. Knowing what the literature says about a condition is not the same as knowing what this patient needs next. Med-BEAM can help a patient reach the right specialist in one visit instead of three, starting treatment months sooner and freeing up appointments for other patients.

Med-BEAM accomplishes this remarkable performance 100 times faster and at 250 times lower cost. The speed and cost advantages are just as important as accuracy. For AI to become integral to how healthcare is organized and delivered, clinical intelligence needs to be available continuously, for every patient, wherever decisions are made. Making that possible at the speed, scale, and cost required by healthcare opens up possibilities that would be impractical with today's general-purpose models.

We see world models and language models as complementary. LLMs provide increasingly capable interfaces for patients, clinicians, and AI agents, while Knit's world model provides a dynamic understanding of the clinical environment in which they operate. This understanding is also essential for safety. As AI agents take on greater responsibility in healthcare, they need to understand not just what actions they can take, but how those actions may affect patients and their trajectories through care. Combining the capabilities of language models with the clinical intelligence of a world model can make AI agents both more capable and safer. The combination is far more than the sum of their parts.

Knit is developing causal AI for healthcare

Building our world model requires an unusual combination of expertise. Our team brings together scientists and clinicians at the forefront of AI, precision health, causal inference, and human behavior with people who have spent their careers studying and working in healthcare. This combination is essential because healthcare is not simply a collection of medical facts or a sequence of clinical events. It is a complex system in which patient biology, clinical decisions, organizational constraints, and human behavior interact.

Learning to predict what happens to patients is an important first step. The more difficult problem is understanding how outcomes change when different decisions are made. This is one of the central scientific challenges we're working on. Med-BEAM presents a unique model that will be able to understand not only patient trajectories, but also the capabilities of different clinicians and care settings, and ultimately how different ways of delivering care can change outcomes.

A new model for healthcare

The immediate opportunity is to make much better use of the resources healthcare already has. Patients can be matched to clinicians and services better suited to their needs. Hospitals can anticipate how patients will progress through care and plan accordingly. Health systems can allocate scarce clinical resources more effectively, reducing unnecessary delays, interventions, and administrative work.

More than two centuries ago, Adam Smith observed that the division of labor is limited by the extent of the market. Specialization drives productivity, but how much specialization an economy can support depends on its ability to connect specialized capabilities with demand. Healthcare has developed extraordinary specialization, with clinicians possessing remarkable expertise in treating particular conditions in particular types of patients. But the benefits of that specialization can only be realized when patients reach the right experts. We envision a future in which every patient can be matched quickly and effectively to the clinicians best suited to their individual needs, with specialized care coordinated around the patient rather than leaving patients to navigate the complexity themselves, as they must today.

But the potential goes well beyond allocating existing resources better. The way healthcare is organized today is not inevitable. Hospitals, departments, referral networks, and clinical roles developed partly in response to the difficulty of coordinating complex care. We bring services together within organizations because coordinating them independently is hard. We rely on standardized pathways because evaluating every alternative for every patient is beyond the capacity of clinicians and existing information systems.

As those constraints change, so will the organization of care. Clinicians will be able to focus on the patients and conditions where their expertise is most valuable, developing even greater specialization and skill. Care teams will be assembled around the needs of individual patients rather than fixed organizational boundaries. Services will be coordinated across institutions and settings. AI agents will take on new responsibilities alongside physicians, nurses, and other healthcare professionals.

When we can understand the needs of patients, the capabilities of clinicians and technologies, and the consequences of different ways of delivering care, we can begin to rethink how healthcare itself is delivered: developing new forms of specialization, combining human and AI capabilities in new ways that are not practical today, and organizing care around what works best for patients rather than the constraints of today’s systems. Some of the most important innovations may be forms of care we cannot yet imagine.

We see the real opportunity for AI in healthcare as fundamentally changing what healthcare can accomplish. It is not simply to automate the work we do today, or even to allocate existing resources more efficiently. It is to make possible a different, much more productive way of understanding and organizing care. And the ultimate measure of that productivity is not how many appointments we schedule or how much administrative work we eliminate. It is how much more we can do to improve human health.

Where we're going

To realize this vision, we’re starting with applications in patient navigation, specialist routing, scheduling, and hospital operations. These are important problems in their own right, and they allow us to bring the world model into practical use.

But we don’t see Knit as a collection of applications. We’re building a foundation of clinical intelligence that health systems, technology companies, and AI agents can use to understand patients, anticipate how care will unfold, and evaluate different ways of delivering it.

Nor do we believe Knit needs to own every application or workflow. Quite the opposite: by providing a shared foundation of clinical intelligence, we can enable health systems, technology companies, and developers to build applications without each having to develop that intelligence independently. We are creating an ecosystem in which independent developers, clinicians, and organizations can build on a shared understanding of how healthcare works.

Our Series A gives us the resources to accelerate this work, expand our scientific and engineering teams, and bring our technology into more clinical settings. I’m grateful to our investors for sharing our ambition, and especially to the team at Knit.