AI Underwriting and Parametric Flood Cover
Arden Mutual is considering entering parametric flood cover for UK SMEs. Its machine-learning pricing model uses eight years of claims data, satellite imagery and a gradient-boosted algorithm; the validation team cannot fully explain the model's drivers, and the data window excludes any major inland flood. The product would cover a defined rainfall threshold, so basis risk is high: a recent event caused widespread damage that fell below the trigger. Expected gross premium is £40m and the target combined ratio is 92%, but the capital treatment under the internal model is uncertain because the model has no parametric trigger module. The regulator has published guidance on AI explainability and consumer outcomes. The board sees the product as strategic diversification and wants to launch in six months. The CRO is asked whether the ERM framework can absorb the risk, and what would make the launch acceptable.
1. Identify the principal emerging and strategic risks in the proposed launch and assess whether Arden's existing ERM framework is adequate to manage them.
Read the worked answer
Principal risks include: (i) model risk - limited explainability and data bias, with an eight-year window missing a major inland flood; (ii) basis risk - the trigger may not pay when SMEs suffer damage, creating conduct and reputational risk; (iii) climate non-stationarity - historical data may understate future rainfall intensity; (iv) regulatory risk - AI guidance and consumer duty expectations; (v) capital risk - the internal model has no trigger module, so capital may be understated or diversification benefit may not be recognised. The existing ERM framework is not adequate: it lacks an emerging risk process with horizon scanning, a model risk policy for AI, and a parametric trigger module. The risk register would capture these only after the fact. I would recommend a pre-launch risk assessment, independent model validation and a board-approved risk appetite for basis risk and model risk before any launch decision.
2. Recommend the conditions, limits and staged decision gates you would require before launch, and explain how the product should be embedded in the ORSA and capital planning.
Read the worked answer
Conditions: (i) stage the launch with a pilot of £5m premium and a hard aggregate limit; (ii) independent validation of the ML model and a documented explainability standard; (iii) create a parametric trigger module or apply a prudential margin in the internal model; (iv) set basis risk limits, for example maximum 10% of premium where the trigger is more than 20% away from actual loss; (v) contractual clarity and customer disclosure. Decision gates: after 12 months, review claims experience, basis risk outcomes and model performance; scale only if the combined ratio is at or below 95% and complaints are below threshold. Embed in the ORSA with a stress where a major flood falls below the trigger, and a climate non-stationarity scenario. Capital planning should include a capital add-on until the model is validated. Governance should be through a product risk committee with the CRO, chief actuary and compliance reporting to the board. This makes the launch a controlled strategic experiment rather than an unquantified leap.