AI Doctor 'Dawei' in Financial Report Drives JD Health's Value Reassessment

On August 13, JD Health announced its 2026 interim results: total revenue of 40.9 billion yuan, up 15.9% year-on-year; Non-IFRS operating profit of 3.5 billion yuan, up 40.3% year-on-year. During the earnings call, management repeatedly mentioned the AI doctor 'Dawei', regarding it as the core product for full-scenario AI implementation.

'Dawei' covers over 1,000 traditional Chinese and Western medicine conditions. During the 618 shopping festival, the number of users served grew nearly 4 times year-on-year, with a satisfaction rate above 98%. Behind these numbers is the leap of medical AI from simple Q&A to the deep end of actual services.

Taking 'an 8-year-old child suddenly has a high fever, 38 degrees' as an example, 'Dawei' does not give a direct answer but instead asks step-by-step about the duration of the fever, cough, runny nose, and other symptoms, using visual options to lower the barrier to description. After three or four rounds of dialogue, it gives a hint of 'acute upper respiratory infection' and attaches 14 references.

In treatment recommendations, 'Dawei' distinguishes between etiological and symptomatic treatment, such as using acetaminophen to relieve fever and headache. It also breaks down the recommendations into specific medications, the process of consultation and prescription, and matches the user's address with nearby pharmacy inventory and delivery time, with the fastest delivery within 24 minutes.

This 'manage to the end' design stems from an emphasis on the user's 'sense of resolution'. According to Liu Hui, product director of JD Health's medical AI, about 70% of users do not return after completing their first consultation on ordinary medical Q&A products because they lack a clear sense of resolution.

'Dawei's' reliability lies in knowing when to stop online processing. A user consulted about left leg swelling, and as the inquiry deepened, symptoms such as numbness, tingling, and fatigue appeared, along with a history of hypertension. 'Dawei' did not give a definitive diagnosis but suggested checking for lumbar nerve compression and lower limb deep vein thrombosis, recommending prompt visits to orthopedics, vascular surgery, or neurology.

Conversely, a user in a remote area consulted about a hard lump on a toe. 'Dawei' combined photos and inquiry to judge it as thickened skin, giving home care advice, with medical visits only needed if pain worsens or infection occurs. This differentiated decision-making reflects clinical logic and risk control.

'Dawei's' underlying capability comes from JD Health's accumulated tens of millions of high-quality doctor-patient dialogue data, which records how doctors gradually clarify vague symptoms—key to learning clinical thinking. The team also collaborates with full-time doctors and experts to organize professional logic for model training.

To ensure judgments stay within clinical bounds, 'Dawei' calls on medical guidelines and literature, and has established over 300 standard diagnosis and treatment pathways, covering more than 90% of high-frequency diseases online. Meanwhile, structured medical rules act as 'guardrails' to reduce the risk of large model hallucinations in diagnosis, risk identification, and medication, such as not recommending a drug when contraindications are triggered.

The entire service process has 109 quality inspection standards. Before model deployment, offline evaluation is required; after deployment, a quality inspection model continuously screens real consultations, with full-time doctors reviewing, forming an uninterrupted inspection mechanism.

'Dawei's' uniqueness lies in combining medical judgment with real-world conditions such as prescription review, drug inventory, and delivery time. The system not only knows what should be done medically but also whether the user's location can purchase the drug, which drugs require a prescription, and whether urgent delivery is needed.

For example, if three drugs are medically suitable but only two can be delivered quickly near the user, the system weighs safety, timeliness, and cost among the realistic options. This planning and scheduling capability relies on infrastructure such as pharmacy networks, doctor resources, home testing, and instant delivery, which is difficult to replicate in the short term.

Liu Hui stated that model capabilities can be iterated quickly, but the fulfillment capability in the physical world cannot be caught up in three months or half a year. This is the hardest part of 'Dawei' to replicate.