The Difference Between Renting Intelligence and Building the Agentic Web
Dear readers,
When we tell the market that Isaree refuses to build thin wrappers around closed cloud models, the critique is immediate: Why not just use the best available API?
Why insist on on-device, open-source or open weight based agents? You are being dogmatic about technology.
We are not dogmatic about technology. We are dogmatic about freedom.
In healthcare, building a wrapper is a renter’s game. You rent the intelligence, you rent the infrastructure, and your margin becomes your vendor's opportunity. More importantly, you surrender data sovereignty, you accept latency, and you build a system that fundamentally cannot pass the strict regulatory certification required by the EU AI Act.
A wrapper creates dependency.
Isaree creates ownership.
Freedom as the Foundation of Science
Freedom is not a brand value we chose for marketing. It is the foundation of scientific work itself. The entire edifice of modern medicine rests on open inquiry, peer review, reproducibility, and the ability of any researcher to challenge prevailing consensus with new evidence. Remove that freedom, and you do not just slow progress, you alter the very nature of how medical knowledge is produced.
We are now approaching a fork in the road. There are two possible futures.
In the first, clinical AI is built on top of centralized, proprietary foundation models. These can be fine-tuned for specific intended uses, and they are. But the foundation beneath them is mutable. When the vendor updates the underlying model, your fine-tuned behavior shifts. Your certification is invalidated. Your reproducibility is gone. Run the same patient case through the same API endpoint six months apart and you may get a different clinical recommendation, not because the evidence changed, but because the model weights did.
This is the epistemological shift. Medical knowledge is no longer validated through open scientific discourse and reproducible methods. It is mediated by systems whose reasoning you cannot inspect, whose training data cannot be fully specified or audited, and whose behavior can drift over time in ways that are hard to detect or control. Clinicians become consumers of conclusions they can neither reproduce nor challenge.
In the second future, intelligence is distributed. Research agents and personalized clinical assistants run directly on the clinician's device — private, fast, under their full control. Clinical agents serving diagnostic or therapeutic functions are diverse, independently certified, and deployed across a hybrid infrastructure that brings compute as close as possible to the point of care.
The scientific method remains intact because the tools of inquiry remain transparent, auditable, and owned by the people doing the work. Knowledge is not monopolized. It is shared, contested, and refined — as science demands.Isaree exists to make the second future inevitable.
The Agentic Inflection Point
We are not alone in this conviction. At the Morgan Stanley Technology, Media & Telecom Conference, Nvidia CEO Jensen Huang made a statement that perfectly captures the shift we are driving in healthcare:
"There will be no software in the future that's not agentic. How could you have software that's dumb? And so, it is absolutely true that every software company will become an agentic company. They're going to simultaneously use open models that they download themselves and fine tune."
Huang expanded on this beginning 2026, predicting that in ten years, we will run on a 100-to-1 ratio of agents to humans, 7.5 million agents working alongside 75,000 employees. He noted that some models will be rented, but crucially, some models must be owned.
Translate this to the hospital environment. The future is not a clinician typing into a monolithic Electronic Medical Record (EMR) system. The future is a clinician managing a team of 10 to 50 specialized, on-device agents. One agent handles transcription. Another retrieves longitudinal patient memory. A third cross-references current medications against the latest clinical guidelines.
You cannot run 50 specialized agents efficiently if they all require round-trips to a centralized cloud server. The latency, the compute cost, and the privacy risks break the model entirely.
The Illusion of the Wrapper Moat
You have likely seen the pattern. A new AI tool appears in your inbox promising to "transform your clinical workflow." You try it. It feels like ChatGPT with a medical skin. That is because it is.
Most healthcare AI companies today are simply placing a thin interface on top of a general-purpose cloud model they do not control. They cannot guarantee where your patient data goes. They cannot promise the tool will behave the same way next month when the underlying model updates without warning. And they certainly cannot certify it as a medical device.
In healthcare, the moat is not the foundational model. The model is becoming commoditized, a compression algorithm for applicable knowledge, much like the MP3 was for audio.
Germany invented the MP3 algorithm, but failed to capture its value because they did not build the platform. Spotify and Apple Music built the platforms that distributed and monetized the MP3.
Right now, the healthcare AI market is obsessed with the "MP3", the massive, generalized cloud models. But the real value lies in the platform that distributes, certifies, and orchestrates specialized intelligence.
Clinicians as Creators, Not Consumers
This brings us back to ownership.
If a hospital relies on a wrapper, it is renting its operational intelligence. When the internet goes down, the hospital stops. When the vendor changes its pricing, the hospital pays.
Isaree is building the operating system for the Agentic Web in healthcare. Our platform allows clinicians to build, select, and download their own specialized AI agents directly onto their devices. This enables 100% offline use. It guarantees that sensitive patient data never leaves the institutional boundary.
But the most critical aspect of ownership is commercialization. We believe that the physicians of the future must be both users and creators of AI. On the Isaree Community Hub, a clinician can build a highly specialized agent using an open-source model, fine-tune it on their proprietary clinical data, and then share or commercialize that agent with other hospitals.
You cannot commercialize a wrapper.
You cannot certify a wrapper as a medical device when the underlying model changes unpredictably.
What Becomes Possible When You Own the Intelligence
Skeptics say on-device AI cannot match cloud performance. They are looking at last year's benchmarks.
With the advent of 2nm chip technology, we are witnessing a 180-degree shift in the market — comparable to the transition from analog to digital cameras. Within five years, 10B parameter models will run flawlessly on the phone in your pocket. But the real unlock is not just inference. It is training.
We are approaching a future where clinicians can train and fine-tune models directly on their local data, in full private mode, without a single byte leaving their device or institutional network. Consider what this means in practice.
A cardiologist at a university hospital has reviewed more than 12,000 echocardiograms over the course of her career. Her pattern recognition is exceptional—yet it remains locked in her own experience. With on-device training, she can fine-tune a specialized AI agent using her annotated cases, effectively encoding her clinical judgment into a model. That agent becomes her personal tool, or an institutional asset. It runs directly on her tablet during rounds, adapts continuously through her feedback, and improves in real time. No cloud infrastructure. No external vendors. No data-sharing agreement
A rural GP practice with three physicians manages a population of 4,000 patients with complex chronic conditions. They fine-tune a longitudinal care agent on their own patient histories, medication interactions, seasonal patterns, local health determinants that no generic model would ever capture. The agent becomes an institutional asset that grows more valuable with every consultation.
A neurology consortium deploys a network of 70 specialized clinical agents, with around half adapted to their own local patient cohorts to better match regional demographics and reduce bias. Orchestrated through a coordinating agent, this distributed system achieves stronger local performance than a single centralized model trained on pooled or proxy data, especially for site-specific workflows and documentation styles. Such an agentic network would be extremely difficult to reproduce with a centralized model alone, because the necessary data is both highly sensitive and too context-specific for a general-purpose system to learn from safely. Running on-device or within each hospital’s secure perimeter, it becomes a powerful clinical tool that the group can describe in their next paper and offer as a configurable agent bundle to other centers via the Isaree Community Hub.
These are not distant fantasies. The hardware is arriving. The architecture is ready. The only question is whether clinicians will own this future or rent it from someone else's data center.
The compute cost drops to near zero, breaking the linear relationship between AI growth and cloud expenses. Wrapper companies will burn venture capital on API calls while clinicians on Isaree build, train, and commercialize intelligence that is truly theirs.
The Choice
We are building the infrastructure for this inevitable future. We orchestrate small, highly efficient, certified open-source models at the edge. We decouple clinical intelligence from rigid legacy software. We give clinicians the tools to become creators, not just consumers, of the AI that shapes their practice.
Build With Us
If this resonates — if you want to stop reading about the future and start building it, we are making that possible right now. We are organizing our first large-scale Healthcare Hackathon together with 15 university hospitals across Germany. Clinicians, researchers, and developers working side by side to build agents that solve real clinical problems on real infrastructure. This week, we had our first conversations to expand the hackathon to Switzerland and the Netherlands.
This is not a pitch event. It is a building event. Come build an agent. Own what you create. Share it with others who face the same challenges.
We are dogmatic because we know that the future of healthcare AI is not locked in a distant data center.
It is owned by the clinician and the institution, trained on their own data, running at the exact space and time of demand.That is not dogma. That is freedom.
Wish you a great week,
Bart de Witte
CEO Isaree GmbH