In Depth: What’s Standing in the Way of China’s Medical AI Agents?

1. Dr. Mao Hongjing, deputy head and chief physician at Hangzhou Seventh People’s Hospital, has dedicated nearly three decades to treating insomnia, observing patients mistreated with inappropriate sleep medications without addressing root causes like depression or sleep apnea. China faces a massive insomnia epidemic, but few receive standardized care due to specialist shortages [para. 1][para. 2][para. 3].

2. In August 2024, Mao partnered with Ant Group to create an AI agent on Alipay for sleep consultations, medical history intake, and personalized plans, launched two months later. Over 1 million users have consulted it, handling intake and cognitive behavioral therapy (CBT), freeing Mao for diagnoses [para. 4][para. 5].

3. This exemplifies AI agents' potential to ease China’s strained healthcare, concentrated in eastern provinces, by providing specialist-like guidance via smartphones, though obstacles persist [para. 6][para. 7][para. 8].

4. AI agents promise much but lag in accuracy; Mao’s agent started at 20% of his proficiency, now at 78%. Alzheimer’s screening agent Ruisi hits 50-60% accuracy, targeting over 90% with 4,000 cases validated by PET scans, per developer Liao Zhengluan of Zhejiang Provincial People’s Hospital. No standardized evaluation exists [para. 9][para. 10][para. 11][para. 12].

5. Tsinghua’s Wong Tien Yin warned in a June 2025 JAMA article that AI lacks clinical oversight, as patients turn to tools like DeepSeek, bypassing safeguards. Agents are limited to triaging, education, and post-treatment [para. 13][para. 14][para. 15].

6. Data access is a core barrier; clinical data is locked in hospitals due to privacy laws and bans on sharing with tech firms [para. 16][para. 17][para. 18].

7. Hospital data is poor: only 30% high-quality, per Fudan’s Zhu Tongyu, with unstructured, error-ridden files fractured across systems. Human tweaks, like inflated dosages for insurance, corrupt datasets [para. 19][para. 20].

8. A 2025 National Data Administration pilot uses blockchain “trusted data spaces” for in-hospital AI training without data export [para. 21].

9. Commercialization falters: hospital sales face slow procurement amid budget strains, with many below break-even due to insurance controls and competition [para. 22][para. 23][para. 24][para. 25].

10. Consumer tools have low retention; Ant Group spent hundreds of millions yuan monthly for Afu’s 27 million users. IP and profit splits complicate partnerships [para. 26].

11. Legally, only licensed physicians bear liability, confining AI to assistant roles. Zhang Wenhong opposes core diagnostic integration, fearing skill erosion; Wang Xiaochuan prioritizes patient health [para. 27][para. 28][para. 29][para. 30].

12. National Health Commission November guidelines promote AI for grassroots aid and pre-consults but mandate human oversight. Contrast: Utah legalized AI prescription renewals in January 2026 [para. 31][para. 32].

(Word count: 498)