The Short Answer
Tsinghua's Agent Hospital has run 42 AI agents across 21 clinical departments since April 2025, and the 93% everyone quotes is a 2024 simulation result, not a clinical one. Clinical AI still has not transformed care delivery because the constraint is data: records are fragmented across institutions, and 72% of hospitals report information gaps. With ibl.ai you own all the code and the data, so agents read your systems of record in place.
A model that reasons well over a complete record and a model that reasons well over a third of one are the same model with very different clinical value.
What is Tsinghua's Agent Hospital, and is it actually new?
It is a virtual medical facility staffed by AI agents — and no, it is not new, which is the more useful fact about it.
Tsinghua University's Institute for AI Industry Research published the framework paper in May 2024, began internal testing late that year, and inaugurated the facility in April 2025.
It runs 42 AI "doctors" across 21 clinical departments, supporting diagnosis for more than a thousand conditions.
The widely-quoted 93% needs precision, because it is routinely reported as a clinical result. It comes from that 2024 framework paper, measured on a subset of MedQA in simulation — a benchmark score for the approach, not a measurement of the facility treating patients.
It circulates periodically as a just-launched breakthrough.
It has been operating since April 2025, and that longevity is what makes it evidentiary rather than promotional: if agent-based diagnostic reasoning at that level has been demonstrated since 2024, and clinical AI still has not changed care delivery at most health systems, the reasoning was not the bottleneck.
Why isn't one institution's record enough for clinical AI?
Because patients do not live inside one institution, and the record follows the patient rather than the building.
Research on interoperability finds 72% of hospitals report information gaps — patient records with missing or incomplete information.
A typical adult's medical history is distributed across a primary care practice, one or more hospital systems, an imaging center, a lab network, a pharmacy chain and several specialists — each with a separate EHR instance, separate identifiers, and a separate release process.
This is the gap between a benchmark and a bedside. A diagnostic benchmark presents a complete, curated case. Clinical practice presents whatever happens to be in the local chart, with the rest of the patient's history somewhere else entirely.
An agent scoring 93% on complete cases is not 93% accurate on a patient whose record it can only partly see. It is reasoning confidently over an incomplete history, which is a well-understood way to be wrong in medicine and one that clinicians are trained to guard against in themselves.
Why doesn't building a central data warehouse solve healthcare AI's data problem?
Because it creates a second problem that is harder than the first.
The instinct is to consolidate: extract from every source into one repository the AI can query. In healthcare that runs into three obstacles at once.
Every copy is a new compliance surface. PHI duplicated into a warehouse is PHI to secure, audit, retain and eventually destroy under the same obligations as the original — multiplied by each copy.
Copies go stale in a domain where staleness is clinical risk. A medication list extracted last night does not include this morning's change. In many workflows a confidently stale answer is worse than no answer.
Consolidation across organizations is a legal project, not an engineering one. Extracting another health system's records into your warehouse requires agreements that in most cases do not exist and would take years to establish.
The alternative is reading in place: role-scoped, read-only access to each system of record, with the answer assembled at query time. Nothing is copied, nothing goes stale, and each source keeps its own governance.
This is the same architecture argument that decides whether banking AI reaches production, arriving in a domain where the consequences are clinical rather than financial.
What does a clinical AI deployment need beyond the model?
Four capabilities, none of which improves with a better model.
- Read-in-place access across systems, role-scoped and read-only, so the agent assembles from sources rather than from a copy.
- Identity resolution across institutions that never agreed on a patient identifier — the genuinely hard problem, and the one that determines whether cross-institutional context is possible at all.
- Provenance on every retrieved fact. A clinician needs to see which system an assertion came from and as of when. An answer without provenance cannot be safely acted on, however well-phrased.
- Governance enforced beneath the model. Role-based access tied to the existing identity provider, complete audit trails, and PHI handling enforced server-side rather than requested in a system prompt.
How does ibl.ai deploy for health systems?
By putting the platform inside the institution's perimeter, where the data already is.
With ibl.ai you own all the code and the data.
The platform runs on the health system's own infrastructure with full source code access — so PHI never transits a third party's cloud unless the institution chooses it — is model-agnostic across any LLM, is usage-based with no per-seat pricing, and deploys anywhere from private cloud to on-premise or a fully air-gapped network.
Agents read from systems of record in place under role-scoped permissions enforced server-side, every access is audited, and access control binds to the institution's existing identity provider.
For a compliance officer, the relevant property is that all of this is inspectable: you can read the code that handles PHI rather than accept an assurance about it.
ibl.ai is family-owned and operated from New York, NY.
Related reading: why only 15% of banking AI reaches production — the same three architectural gaps in a differently regulated industry.
Sources: the framework and the 93%-on-MedQA simulation result from the May 2024 Agent Hospital paper; the April 2025 inauguration from TechNode; the information-gap figure via Salesforce's healthcare interoperability overview.