Overview
As artificial intelligence penetrates insurance operations, a number of insurers have begun classifying use cases into tiers — defining where AI can be used freely, where dedicated approval is required and where it must be avoided altogether. The regulatory floor was set in June, when guidance on the safe development and application of AI in banking and insurance stated that users bear responsibility, and classified generative AI uses directly tied to customer interests and the formation of insurance contracts — notably underwriting and claims — as high-risk, requiring approval by the institution’s risk management committee.
According to reports, one property insurer divides AI uses into three categories: low and medium-risk scenarios suitable for routine use, high-risk scenarios under strict control, and areas to avoid. Low-risk uses center on internal retrieval, policy Q&A, document processing and office productivity. Fund transactions, underwriting, claims, risk management and public-facing generative services fall under a triple mechanism of special approval, human review and fallback to manual processes.
Background & Interpretation
1. Regulatory and industry context
The industry’s stance is shifting from encouraging exploration to pairing it with safety. The regulator’s guidance makes institutions accountable as both service providers and technology users, preventing them from attributing failures to model vendors. By naming underwriting and claims as high-risk, the framework links technical governance directly to consumer protection.
2. Core drivers and mechanisms
Tiering reflects an asymmetry between risk and benefit. Errors in internal productivity tools typically stay inside the workflow and are cheap to correct. A model-generated underwriting or claims decision, by contrast, turns immediately into a wrongful rejection, mispriced claim or unfair rate, with legal and conduct consequences. Institutions therefore use whether a conclusion takes effect for the customer as the dividing line, keeping final decisions human and maintaining a path back to manual processing. Privacy data is treated as an absolute boundary: sensitive information does not enter model training, and human control is retained over decisions, liability and customer interactions.
3. Market structure and value-chain impact
Alongside tighter controls, efficiency tools are spreading. A fintech showcase in Jiangsu described an insurer using AI in small personal-injury traffic claims: photos are recognized automatically, injuries assessed with assistance, a settlement recommendation generated within 30 seconds and payment made as fast as one minute after online agreement. Agricultural platforms combine satellite imagery, drones and weather data for pre-disaster warnings and post-event loss assessment. In both cases AI handles information processing and decision support while institutional processes and human staff retain final responsibility. This is pushing insurance-technology vendors from selling models toward delivering explainable, traceable, reversible workflows.
Implications & Outlook
Dedicated approval will lengthen time to production for some projects but reduce compliance rework. Three variables are worth tracking: whether risk committees converge on common approval standards, how model explainability is audited, and whether the frequency of fallback to manual processes becomes a measurable operational metric. Simultaneous evidence of faster claims and fewer complaints would accelerate adoption.
Takeaways for Industry Participants
Insurers should turn scenario inventories, data boundaries and approval responsibilities into enforceable rules rather than generic confidentiality warnings, scaling low-risk uses quickly while keeping humans in the loop for underwriting and claims. Vendors should deliver audit logs, explainable outputs and rollback interfaces as standard capabilities. Policyholders, meanwhile, should be aware of review channels so disputes can reach a human decision-maker.
This column compiles industry information and shares technical perspectives; it does not constitute investment advice.