概述
This guide uses named examples documented by Progressive and Zurich, a simplified expected-loss illustration, and NAIC guidance on insurer responsibility for AI systems.
深入探讨
In a simplified example, expected loss is claim probability multiplied by claim severity; actual insurance pricing and underwriting depend on product, jurisdiction and additional factors. Progressive documents telematics signals such as time of day, hard braking or acceleration, and amount driven. Zurich and Nearmap describe aerial imagery that provides property and roof-condition or maintenance indicators. NAIC guidance says insurer decisions made or supported by AI must comply with applicable insurance laws and describes governance and risk-management expectations. A variable correlated with a protected trait can be a proxy risk worth testing, but correlation alone does not prove discrimination. SHAP attributes a prediction to model inputs; it does not establish causality or fairness.
现实世界的实施
Progressive’s Snapshot FAQ lists signals such as time of day, hard braking or acceleration, and amount driven.
Zurich and Nearmap describe an aerial-imagery property assessment example.
A simplified expected-loss calculation weights claim severity by claim probability.
The NAIC model bulletin discusses insurer governance, risk management and compliance responsibility for AI-supported decisions.
人工智能在保险承保中的未来
Evaluate the actual data, decision process and subgroup outcomes for each product. Prescription, credit, motor-vehicle and location data are jurisdiction- and product-specific and should not be treated as universal practice without direct evidence.
付诸实践
为什么这很重要
- 背景与规则. 行业背景决定了人工智能创意能否与现实接触。
- 质量控制. 领域约束会影响可接受的错误率和监督模型。
- 构建选择. 成功的部署使技术能力与一线工作流程保持一致。
留意
- 监管要求可能会使原本强大的原型失效。
- 历史数据可能会编码损害特定社区的偏见。
- 遗留系统可能会造成集成瓶颈和隐性成本。
后续步骤
- 让领域专家参与从问题框架到评估的整个过程。
- 在启动前设计审计跟踪和文档。
- 尽早验证合规性和安全义务。
- 分阶段推出，并具有明确的停止和回滚标准。
不断探索
常见问题
What is AI in Insurance Underwriting?
Insurance underwriting evaluates information about a policy or exposure. This guide uses named examples documented by Progressive and Zurich, a simplified expected-loss illustration, and NAIC guidance on insurer responsibility for AI systems.
In a simplified hypothetical risk model, how is expected loss for one exposure calculated?
Expected loss is the probability-weighted average claim amount in this simplified example. It is a basic modeling definition, not a claim about every insurer’s production pricing formula.
Why might a regulated insurer document model inputs and validation?
NAIC’s AI guidance discusses governance, risk management, documentation and oversight. This does not imply that one model family is universally preferred for regulated underwriting.
Under the NAIC model bulletin, who remains responsible for AI-supported insurance decisions complying with applicable law?
The NAIC model bulletin places responsibility on the insurer using an AI system to ensure decisions comply with applicable laws. This is guidance context, not a statement that one rule applies identically in every jurisdiction.
What fairness risk should be tested when a model uses variables correlated with protected traits?
NAIC’s AI guidance highlights governance and unfair discrimination risks. A correlated variable can be a concern to evaluate; ZIP code is not asserted here to be a universal proxy in every product.
In an underwriting model, what is a feature?
A feature is a model input. Feature attribution methods such as SHAP describe how inputs contribute to a model prediction, but do not prove causation.