What Happened

Ping An Technology’s head of health insurance intelligence, Li Xiang, is scheduled to speak at AICon Global Artificial Intelligence Development and Application Conference in Shenzhen on August 21–22. His session, titled “Innovation and Practice of Agentic AI in Inclusive Health Insurance,” will be part of the forum on high-value commercial AI Agent scenarios.

The presentation will focus on three insurance scenarios: health insurance product innovation, intelligent underwriting, and intelligent claims handling. It will cover the use of multimodal AI risk prediction, an AI underwriting agent, and an AI claims agent in inclusive health insurance.

Why Health Insurance Is a Difficult AI Scenario

Health insurance is a demanding field for AI because the data and rules are both complex. Medical bills, handwritten medical records, examination reports, and test results come from different sources and formats. Sending all of this material directly into a large model for end-to-end parsing can increase token consumption, latency, and hallucination risk.

The rule system is also difficult to maintain. Health insurance policies contain many clauses, and underwriting and claims review can involve hundreds of detailed checkpoints. Traditional hard-coded rule engines require engineering changes and regression testing whenever clauses or products are updated. Claims handling is also a long workflow, from case reporting and data entry to review, calculation, and final decision output.

Ping An Technology’s proposed approach is to move from a “static rules plus single large model” setup to an Agentic Skills multi-agent architecture. In this model, different agents take on different responsibilities, such as task routing, specialized review, and final decision aggregation.

Underwriting and Claims Use Cases

For underwriting, Ping An’s AI underwriting agent is designed around one-click upload, one-click underwriting, and one-click consultation. After a user uploads medical reports, the system identifies and parses the information, connects medical signals with insurance responsibilities and underwriting rules, predicts disease risks, and provides underwriting options and professional answers.

The architecture includes three main components:

• Disease prediction models, which process images, test results, or structured factors and output disease risk probabilities;
• Medical large models, which understand unstructured medical records and reports and convert them into computable risk factors;
• Multi-agent collaboration, where a scenario scheduling agent assigns tasks, specialized review agents make graded judgments, and an aggregation agent produces the final conclusion.

For claims handling, Ping An’s AI claims agent targets complex materials, lower manual review efficiency, and high risk-control costs. A policy parsing agent extracts more than one hundred core liability review points from complex policy responsibilities and converts policy content into structured data. A scheduling agent coordinates the case, distributes subtasks in parallel, and aligns context. Specialized agents then process detailed review tasks, while a summary agent resolves conflicts and merges decisions.

Engineering Lessons and Controls

The disclosed outline emphasizes engineering reliability rather than simply applying a large model. One challenge in multi-agent collaboration is context loss: when agents pass work between one another without a strict format, conclusions may conflict. Ping An’s solution is to introduce a constrained state machine and standardized schema contracts, defining each agent’s inputs, outputs, and process boundaries.

Long medical records can also cause token expansion and slow responses. The team uses RAG plus key-information pre-extraction. RAG, or retrieval-augmented generation, means retrieving relevant source material first and then letting the model reason over those selected passages. This allows review agents to retrieve medical course records and test indicators as needed, keeping average processing time at the seconds level.

Hallucination control is another key requirement. Ping An uses an “AI evidence tracing plus deterministic rule fallback” mechanism. Model-generated conclusions must be linked to the original text location. If confidence is insufficient, the system falls back to traditional deterministic rule engines or human review. According to the disclosed outline, after these optimizations, AI review accuracy remains above 96%, and the review automation rate exceeds 70%.

Industry Outlook

This case reflects a broader shift in AI adoption: competition is moving beyond model capability alone toward reliable agents, engineering systems, and business integration. Health insurance requires accuracy, traceability, compliance, and user trust, so an AI system must be auditable and able to fall back to rules or humans when needed.

Ping An has already launched several innovative medical insurance products, including chronic disease and family-oriented million-yuan medical insurance offerings. The likely direction for inclusive health insurance is not full automation at any cost, but specialized models, engineered multi-agent workflows, and controlled human-rule fallback. If implemented carefully, Agentic AI could make underwriting clearer, claims faster, and insurance services more accessible while keeping risk management in place.