Dr. Mao Hongjing has spent nearly three decades trying to help the vast number of Chinese who spend long hours in bed every night sleeplessly staring at the ceiling.
As deputy head and chief physician of Hangzhou Seventh People’s Hospital, he has seen the same frustrating pattern repeat itself: patients arrive at his clinic after already cycling through a cocktail of inappropriate sleep medications, their underlying conditions — depression, anxiety, sleep apnea — never properly investigated.
China has a massive population suffering from insomnia, he said, yet only a small fraction ever receives standardized treatment. For him, the problem isn’t just awareness, it’s the math — there are never enough hours in a specialist’s day.
In August 2024, Mao decided to try something different. He partnered with Ant Group Co. Ltd. to build a digital version of himself — an artificial intelligence (AI) agent — deployed on the Alipay platform, capable of interviewing patients about their sleep, taking their medical history, and recommending personalized treatment plans. Two months later, it went live.
To date, more than 1 million users have consulted it. The agent handles intake and follow-up cognitive behavioral therapy, while Mao reserves his limited clinical time for what only he can do: diagnose patients and chart a course of treatment.
This virtual version of Mao offers a glimpse at the promise of AI agents for China’s long-strained health care system. China’s top hospitals and specialists are heavily concentrated in richer provinces on the eastern seaboard, leaving swaths of the population with far less access to expert care.
Proponents argue that AI agents could begin to close that gap, putting something close to specialist-level guidance within reach of anyone with a smartphone. But despite a number of anecdotal success stories like Mao’s, several obstacles remain to the mass deployment of AI agents in health care.
These challenges are important to understand because of the potential that AI agents have in alleviating some of the problems plaguing China’s public health care system. Chronically underfunded, geographically lopsided and often overwhelmed by patients, the system could make good use of the help promised by AI technology. But it is also an environment where the barriers to adoption remain deeply entrenched.
Accurate, to a point
The promise of AI agents is easy to articulate, but the reality is messier. Despite the enthusiasm surrounding anecdotes like Mao’s, multiple industry insiders told Caixin that current agents still fall short of expert-level human cognition, and performance varies widely. No standardized evaluation framework exists to measure how well they work.
The numbers bear that out. When Ant Group first launched Mao’s sleep agent, it achieved only about 20% of his clinical proficiency. It has since improved significantly, with diagnostic alignment now reportedly reaching 78%, Caixin has learned. It’s a meaningful jump, but still a considerable distance from the standard a patient might reasonably expect from a specialist with three decades of experience.
It’s a similar story for an agent designed to screen for Alzheimer’s disease. Called Ruisi, the agent accuracy is currently in the 50% to 60% range. Its developers are aiming to push that figure above 90% by collecting 4,000 community cases and validating results against biological gold standards like PET scan biomarkers, according to Liao Zhengluan, Ruisi’s developer and the chief of clinical psychology at Zhejiang Provincial People’s Hospital.
The limitations have drawn serious academic scrutiny. Wong Tien Yin, dean of Tsinghua University’s school of medicine, published a warning in the Journal of the American Medical Association in June 2025, cautioning that while AI has given underserved populations unprecedented access to health advice, these interactions often lack the clinical oversight necessary for safe implementation.
Patients are increasingly turning to broad AI tools like DeepSeek as cheap alternatives to traditional medical consultations, inadvertently bypassing essential clinical safeguards, Wong wrote.
For now, the functional scope of most agents remains largely confined to pre-consultation triaging, health education and post-treatment management — useful, but far from the transformative role their proponents envision.
Behind locked doors
Even if accuracy problems could be solved overnight, developers would still face a deeper, more intractable problem: they can’t get their hands on the data needed to solve them. The clinical data required to train better agents sits inside hospital networks and getting it out is extraordinarily difficult.
The reasons are both legal and practical. Patient privacy concerns mean that there are heavy restrictions on accessing sensitive medical data, with breaches punished severely. Risk-averse hospital administrations have responded with blanket bans on sharing data with external commercial partners. Tech companies possess the capability to process and structure the data but are legally barred from accessing it. The result is a standoff.
The raw data that does exist inside hospital networks is itself deeply problematic. Zhu Tongyu, vice dean of Shanghai Medical College at Fudan University, estimates that genuinely high-quality original hospital data accounts for about 30% of the total available pool. The rest exists as unstructured data, riddled with missing elements and coding errors. Critical information is fractured across disconnected administrative and clinical computer systems.
Human interference corrupts data sets further — a doctor might officially prescribe a higher daily dosage of medicine than necessary simply to provide a patient with enough pills under strict insurance limitations, Zhu said. That decision renders the data point clinically useless for training an AI model.
A potential bridge is slowly forming. In 2025, China’s National Data Administration launched a trusted data space pilot initiative, functioning like a secure digital highway built on blockchain technology. These spaces allow raw data to remain within hospital networks while permitting authorized AI enterprises to deploy and train models directly on internal servers, extracting only refined parameters and algorithmic upgrades, leaving the patient data untouched. As a workaround, it shows promise, but it’s still in its early stages.
A tough sell
Even agents that clear the accuracy and data hurdles face a third obstacle: turning a working product into a sustainable business. Commercialization has proven stubbornly elusive, and the path to profitability looks different — and difficult — from every angle.
For companies selling directly to hospitals, the challenge is one of institutional inertia. Traditional procurement, bidding and network-listing procedures are agonizingly slow, and hospitals operating under budgetary strain are not natural early adopters.
A significant portion of domestic hospitals are already below the break-even point, squeezed by cost-control measures applied to the national medical insurance fund, competition from elite tier-three hospitals, and reduced government subsidies.
Consumer-facing products face a different set of problems. Basic health consultation tools suffer from low barriers to entry, poor user stickiness and low use. Ant Group reportedly burned through hundreds of millions of yuan in a single month to build its current base of 27 million users for its Afu health assistant — a sobering illustration of what it costs to acquire and retain users in this space. Intellectual property arrangements add another layer of complexity. Investors frequently scrutinize whether a jointly developed product operates under the banner of the enterprise or the hospital, and how profits are split, said Yan Jingjing, a founding partner at Probe Capital. Some corporate-led projects funnel all financial returns back to the enterprise, while paying doctors a one-time upfront fee — meaning the company shoulders all the commercial risk if the product fails to sell.
Legal limits
Underlying all of these challenges is a question that is difficult to resolve: when an AI agent gets it wrong, who is legally responsible?
Under China’s existing legal framework, the answer is unambiguous — and limiting. Only a licensed human physician can assume liability for a medical decision. Because an AI agent lacks legal personhood, it cannot hold a medical license and therefore cannot bear responsibility for clinical failure. This reality restricts AI’s role as a subordinate assistant in any serious medical situation, regardless of how good the technology gets.
The debate within the medical community reflects the weight of that constraint. Zhang Wenhong, director of the National Center for Infectious Diseases at Fudan University-affiliated Huashan Hospital, has explicitly opposed integrating AI into core hospital diagnostics, warning that doctors who rely on AI from their internship onward will fail to develop the rigorous clinical thinking skills needed to verify whether an AI’s diagnosis is correct. Wang Xiaochuan, founder of Baichuan Intelligent Technology, has taken the other side of the argument, saying that the professional development of doctors must not come at the cost of patient health.
So far, regulators have tried to thread the needle. The National Health Commission issued guidelines in November encouraging the use of AI to assist grassroots practitioners and enable smart pre-consultations, while maintaining that human doctors must serve as the ultimate gatekeepers.
The contrast with developments elsewhere is instructive. In January 2026, Utah became the first U.S. state to legally empower an AI system to authorize prescription renewals for chronic disease patients — a small but significant crack in the wall of human-only medical authority.
Contact editor Lu Zhenhua (zhenhualu@caixin.com)
京公网安备 11010502034662号 
