How JD Insurance Uses AI Agents to Automate the Entire Insurance Supply Chain

AI Agent Landing Methodology and Technical Points

The core challenge is selecting appropriate insurance scenarios and defining clear objectives. Each niche service (e.g., extended warranty, freight insurance, pet insurance) can be built as a fully automated AI‑driven supply chain.

Economic Benefit Formula

R = (Ch - Ca) × D × A × S

Where:

R – Economic benefit of the deployed agent

Ch – Unit human labor cost

Ca – Unit agent operating cost

D – Directness of the conversion chain (0~1)

A – Agent knowledge coverage (0~1), calculated as A = M / TI S – Scale

M – Amount of information input

TI – Feedback cycle (time units)

I – Baseline knowledge difficulty (higher value = harder)

The formula shows that higher directness, broader knowledge coverage, and larger scale increase the economic return.

Why Insurance Supply Chain Fits AI Agents

The product is a virtual probability‑based item, making full‑life‑cycle automation feasible.

The production process is highly procedural and rule‑driven, easy for agents to learn.

Agents improve the most expensive, repetitive, error‑prone steps, directly impacting revenue.

Designing AI‑Driven Insurance Products

A domain‑specific large model is created by fine‑tuning a small‑size model on insurance data. Knowledge‑augmented retrieval (RAG) and continuous self‑play reinforcement learning (RL) further improve performance.

Key Technical Highlights

Domain large model : Fine‑tuned on insurance data to capture industry nuances.

Knowledge base : Structured, searchable knowledge objects for accurate, controllable decisions.

Tool integration : External tools (memory, knowledge, utilities, actions) extend agent capabilities.

Scheduling strategy : Coordination layer that plans, reflects, and optimizes multi‑agent workflows.

Continuous evolution : Self‑play RL loop (state → call → new state) with reward signals drives autonomous improvement.

Agent Planning Strategies

Prompt‑based planning : Encode the entire workflow in prompts; suitable for low‑latency, high‑reliability scenarios.

Search‑enhanced hierarchical planning : Use retrieval to reduce context pressure and plan with external knowledge.

RL‑driven autonomous orchestration : Agents learn from execution‑reward loops to handle long‑term, evolving environments.

AI‑Driven Pricing Agent

Combines multi‑agent coordination, machine‑learning risk estimation, and operations‑research optimization.

Accurate risk estimation : Model error < 2% using massive data and ML.

Fast business adjustment : Real‑time monitoring and proactive optimization.

High quoting efficiency : From request to quote in under one minute.

Key components:

Master Agent : Handles intent recognition and task scheduling.

Underwriting Agent : Generates risk‑controlled quotes.

Actuarial Agent : Produces explainable rate tables using ML.

Operations Agent : Monitors performance and suggests adjustments.

AI‑Driven Claims (Fulfillment) Agent

Automates claim verification with >94% accuracy, >95% coverage, and a per‑claim cost of ¥0.02, reducing fraud and operational costs.

AI‑Driven Risk Control (Full‑Chain)

Risk control is split into underwriting, claims, and recovery stages, each using a combination of rules, lightweight models, and large models.

Underwriting : Pre‑compute risk tags and apply them asynchronously.

Claims : Real‑time rule + small model + large model triage.

Recovery : Cross‑temporal data mining and unsupervised clustering to uncover hidden fraud networks.

Technical Foundations for Model Development

Domain large model : Small‑size model fine‑tuned on insurance data to address knowledge gaps, style mismatches, and compliance differences.

RAG knowledge base : Retrieves structured knowledge objects to supplement model predictions.

Training pipeline : Continuous pre‑training → supervised fine‑tuning → alignment optimization.

Data efficiency : Seed data expanded with methods such as WizardLM, MAGPIE, GraphGen, Condor, Self‑Instruct, Self‑QA, Self‑KG.

Evaluation : Measures generic ability, insurance‑specific ability, and business‑level performance across multiple dimensions.

Knowledge Base and Table Processing

Insurance contracts contain many tables. A simple serialization approach converts tables to markdown/HTML, allowing large models to understand irregular structures.

<table>
| 保障类型 | 场景 | 区域 | 报销比例 | 免赔额 |
|---|---|---|---|---|
| 住院医疗 | 一般住院 | 国内 | 80% | 1000元 |
| ... |
</table>

Structural Chunking for Long Documents

Documents are split at the "条" or "款" level, preserving metadata such as chapter path, product name, and effective date. Late chunking retains global semantics.

Reinforcement Learning Loop

During execution, each decision point records State , Call , and New State as a tuple (component, input, output, reward). This sequence forms the training data for RL, enabling agents to improve through self‑play.

Agent Evolution Architecture (Eva)

Four modules:

Expert Agents : Domain experts implemented via large models.

Tool Modules : Memory, knowledge, and action tools augment capabilities.

Scheduling Strategy : Central planner that coordinates sub‑agents, performs reflection, and optimizes plans.

Continuous Evolution : Self‑play RL refines models based on reward‑augmented execution traces.

Future Directions

Strengthen core risk‑control capabilities and expand coverage.

Deploy a master risk‑control agent that orchestrates sub‑agents in a perception‑decision‑execution‑evaluation‑optimization loop.

Continue evolving AI‑native pricing, decision, and end‑to‑end models.

Scale AI digital employees to over 200 agents, reducing marginal costs and boosting efficiency.