China Built an AI-Agent Eye Clinic. Here Is What Happened in the Real World

AI is already widely used in medical imaging and diagnostic support, but most systems still perform individual tasks within a traditional clinical workflow. Researchers at Tsinghua University in China wanted to test a more ambitious model: a clinic where multiple AI agents support patients across the entire care journey.

The result was the AI-Agent Augmented Tsinghua Eye Clinic (AI-TEC). A prototype was deployed at Beijing Tsinghua Changgung Hospital in November 2025, and the first lessons from its real-world implementation were published in Nature Medicine in September 2026.

The project offers an early look at what the researchers describe as a transition from AI-assisted care to AI-native healthcare. More importantly, it shows what happens when an advanced AI system leaves a controlled research environment and meets real patients, hospital data and clinical workflows.

The idea: AI across the entire patient journey

AI-TEC was designed differently from conventional medical AI tools.

Instead of developing one algorithm for one task, the researchers created specialized AI agents for different stages of ophthalmic care. The system supports pre-consultation, triage, diagnosis, clinical decision support, patient education and follow-up.

Information moves between the agents as the patient progresses through care. The aim is to create a connected clinical pathway where AI becomes part of the infrastructure linking patients, physicians, clinical data and decisions. Tsinghua University describes the AI-TEC model and its implementation here.

Deployment in a real hospital quickly exposed challenges that conventional model testing does not necessarily capture.

The first problem came from the data

Real-world hospital data seemed like a natural resource for improving the system. The AI-TEC experience showed why that process is not straightforward.

Routine medical records can contain incomplete information, inconsistencies and unreliable clinical labels. The researchers found that post-training on raw clinical data could actually reduce model performance.

The team responded by introducing clinician-verified labels and expert review before the data were used to refine the models. Model performance subsequently improved, demonstrating that large volumes of real-world data alone are not enough. Clinical expertise remains essential for ensuring that the data used to train medical AI are reliable.

The experience highlights an important principle for healthcare AI: data quality can matter more than data volume.

The next challenge was making AI work for physicians

Another problem emerged when AI became part of everyday clinical practice. A technically strong system still has to fit naturally into the way physicians work. Additional screens, clicks or interruptions can turn a useful tool into another burden during a busy consultation.

The AI-TEC team therefore treated workflow integration and clinician engagement as part of the development process. Physician feedback was used to refine how AI fitted into existing clinical workflows rather than expecting clinicians to adapt their work around the technology.

Unlike the data-quality problem, workflow integration does not have a simple technical fix. The system continues to evolve as researchers study how clinicians interact with it.

The lesson is clear: strong model performance has limited clinical value if physicians do not use the system consistently.

From model performance to real clinical value

AI-TEC remains an early real-world implementation, and questions around scalability, governance, safety and patient outcomes still require further study. Its experience, however, highlights an important shift in how medical AI needs to be evaluated.

Technical metrics can show whether an algorithm performs its intended task, but they cannot fully capture what happens when the technology enters routine care. AI-TEC demonstrated how data quality, physician adoption and workflow integration can influence performance and ultimately determine whether an AI system delivers clinical value.

As AI expands into medical imaging, diagnosis, patient identification, trial matching and clinical decision support, generating evidence in real clinical settings will become increasingly important. Prospective evaluation can help determine whether strong technical performance translates into reliable use, better clinical decisions and meaningful benefits for patients.

Developing a powerful AI model may be the beginning. Demonstrating that it works safely, consistently and effectively in clinical practice is the next challenge, and clinical research will have an increasingly important role in generating that evidence.