AXA

AXA's underwriting teams operated in an environment where accurate, comprehensive guidance review was not optional — it was a regulatory and commercial necessity. Underwriters navigating Property & Casualty risk decisions relied on dense reference documents covering policy wordings, actuarial guidance, and compliance requirements. Manually searching these materials averaged 10 minutes per query, a friction cost that compounded across thousands of daily decisions spanning underwriting, contact centres, and claims processing. Beyond speed, the deeper risk was completeness: a missed guidance document could mean mispriced risk, a disputed claim, or a regulatory breach. The status quo created both operational inefficiency and latent compliance exposure at global scale.

AXA deployed Retrieval Augmented Generation (RAG) to give underwriters a structured, AI-assisted interface for extracting and summarising relevant sections of large reference documents in real time. Rather than replacing actuarial judgment, the system augmented it — surfacing pertinent guidance and reducing the manual search burden while keeping the human underwriter accountable for final decisions. The RAG deployment was preceded by more than four years of testing, trialing, and iterative refinement before group-wide rollout. In parallel, AXA developed over 60 agentic AI use cases across underwriting, contact centres, and claims processing, all built on a common AI platform engineered around open market standards with governance, security, and regulatory compliance embedded from inception rather than retrofitted.

The RAG pilot delivered a 70% reduction in per-query research time, compressing the average from 10 minutes to under three minutes. Critically, speed gains did not come at the cost of confidence: 86% of users rated the tool 8 out of 10 or higher, and reported confidence that they had considered all relevant guidance — a meaningful outcome in a regulated environment where missed information carries real liability. Qualitative adoption signals reinforced the quantitative results, with AXA's Tech, Data and AI teams treating the rollout as a platform milestone rather than a standalone tool. The broader agentic AI programme now counts 60+ use cases in testing or partial deployment across core insurance operations.

• Accuracy and completeness matter more than speed alone in regulated insurance workflows — AXA's 86% confidence rating validated the tool as much as the 70% time reduction did.
• Four years of testing before full deployment signals that enterprise AI in insurance demands structured QA cycles, not accelerated release timelines.
• A common, governed AI platform is a prerequisite for scaling across business lines; bolt-on governance creates fragility as use cases multiply.
• Augmented intelligence framing — positioning AI as a complement to underwriter judgment rather than a replacement — drove adoption and maintained regulatory defensibility.
• End-user confidence metrics should be tracked alongside efficiency KPIs when validating AI tools in compliance-sensitive environments.