How a Modern Data Platform Is Redefining the Future of Insurance

From Internet to Finance: Evolution Logic

Ping An Property & Casualty, the second‑largest player in the property‑insurance lane, contrasts with internet firms that prioritize traffic; insurance focuses on stable operations and strict regulatory compliance. The legacy data platform suffered from three "missing" problems: lack of standards (over 40,000 fragmented metrics), lack of unity (isolated data marts across business groups), and lack of openness (single‑point tools such as OCR were not linked to underwriting or risk control).

Kunpeng Intelligent Platform: From Silo to Systematic

Rather than rebuilding from scratch, the team layered a unified data‑asset management system and metric‑governance framework atop the existing warehouse and data marts. AI capability was split into two layers: a public platform group provides development tools (next‑gen low‑code framework, sandbox security), while each business group builds specific Skills, Tools, Memory, and Context for personal, group, claims, and public‑resource lines, enabling "manage, connect, use" of data in a systematic way.

Three Pillars of Platform Capability

Agent Platform : combines a knowledge base with large‑model components, tools, and a memory module. Recent work added a Decision mode that lets the model autonomously plan and execute tasks.

Data Governance (OSI Semantic Layer) : adopts the Open Semantic Interchange (OSI) standard—defined once, used everywhere—built on YAML‑based declarative semantics and the MetricFlow engine. It acts as a universal semantic "socket" that lets disparate systems speak the same language, enabling natural‑language understanding by large models.

AI Democratization (OpenClaw) : the super‑intelligent agent offers autonomous planning and execution, but in insurance it hit regulatory red flags due to uncontrolled permissions, massive token consumption, and hallucination risks. The team defined clear usage boundaries: high‑frequency, repeatable, highly standardized tasks (data collection, content processing, cross‑system sync) are suitable, while core underwriting, claims, and compliance‑heavy scenarios require template filling and human oversight.

Boosting ChatBI Accuracy with OSI

The original ChatBI workflow (problem rewrite → metric recall → slot extraction → answer) yielded 70‑80% accuracy because metric definitions were inconsistent. By integrating OSI and switching the workflow to a ReAct loop plus OSI, metric‑question accuracy rose from 58.4% to 80%, problem‑rewrite accuracy reached 96%, and slot‑extraction accuracy improved to 85%.

End‑to‑End AI Empowerment Across the Insurance Value Chain

Marketing : Four AIGC capabilities (copy, image, video, digital‑human live) generate PE‑level material (≈700 k pieces) automatically, cutting marginal content‑production cost.

Underwriting & Issuance : The group‑insurance underwriting assistant and car‑insurance issuance assistant turn expert experience into decision‑flow chatbots, automating the full process from opportunity decoding to risk assessment, thereby freeing underwriters.

Claims Automation (Non‑Car) : Combining OCR, image recognition, and chatbots, the team automated receipt, liability determination, loss assessment, and settlement, reducing processing time by 55% and improving case‑closure speed by 30% (award‑winning "Gold Hair" prize).

Agricultural Insurance : Drone + satellite mapping enables remote loss assessment for planting insurance; cattle‑face recognition ensures unique livestock identification for breeding insurance; OCR plus video‑frame counting supports pig‑sty counting, all leveraging biometric‑vector models.

Business Decision : A time‑series forecasting model predicts car‑insurance premium growth at T‑30, T‑1, and T+10 days, integrating internal policy data with public industry data to support the "three‑first" (foreknowledge, foresight, first‑move) strategy.

Key Takeaway

In a heavily regulated financial environment, AI "autonomy" must yield to "determinism". The optimal technology choice balances business pain points, rule constraints, and regulatory red lines; when large models cannot directly serve customers, template filling and human fallback become the safest engineering practice.