Insurers Adopt 'Affirmative AI' to Manage Model Risk
Insurers are reportedly tightening underwriting standards for their technology clients by using "affirmative AI" insurance products. These policies are designed to address specific risks associated with AI systems. Key areas of concern include algorithmic drift, model hallucinations, and data poisoning vulnerabilities.
- Affirmative AI coverage is an explicit endorsement in a Tech E&O or Cyber insurance policy that grants coverage for liabilities connected to AI use, such as algorithmic errors, copyright claims from generated content, and model hallucinations. Insurers use these endorsements to define the scope of covered AI risks and cap financial liability with targeted sublimits, turning the policy into a loss-control mechanism. - Agentic AI systems are being deployed to manage complex, multi-step insurance workflows without direct human instruction at each step. For example, Allianz launched an agentic AI solution in Australia to automate food spoilage claims, reducing processing time from days to hours by using seven specialized AI agents for tasks like coverage checks and fraud detection, with a human making the final payout decision. - Multi-agent systems (MAS) differ from single-agentic systems by using a decentralized network of specialized AI agents that collaborate to handle complex processes like underwriting or claims. In property claim underwriting, this architecture has achieved up to 92.9% accuracy by using coordinated, self-modular agents for different aspects of risk assessment, with a human-in-the-loop for final verification. - Data poisoning, the malicious alteration of a model's training data, is a key risk; studies have shown that poisoning just 1-3% of data can significantly impair an AI's predictive accuracy. This can lead to incorrect risk assessments and fraudulent claim approvals, representing a systemic risk for AI-driven insurance operations. - Integrating AI with legacy insurance systems is a major challenge, with 57% of insurance decision-makers citing it as a significant hurdle. Common architectural patterns to bridge this gap include using APIs to expose legacy functions, employing an "Overlay Model" where AI services run alongside old systems, or using a "Strangler Fig" pattern to incrementally replace legacy components with modern, AI-enabled microservices. - LLM orchestration frameworks like LangChain and LlamaIndex are essential for building enterprise-grade AI applications by managing the complex coordination between multiple LLMs, external data sources, and business systems. These frameworks handle model abstraction, conversation memory, and tool integration, allowing for the construction of sophisticated agent-based systems and Retrieval-Augmented Generation (RAG) pipelines. - After a multi-year contraction, global insurtech funding rose 19.5% year-over-year to $5.08 billion in 2025, with two-thirds of that capital flowing to AI-focused companies. The rebound was most pronounced in Property & Casualty insurtechs, which saw a 34.9% increase in funding to $3.49 billion.
