For most of the last century, insurance pricing rested on a familiar actuarial toolkit: classification tables, experience rating, and, from the 1990s onwards, generalised linear models (GLMs). That toolkit served the industry well because it was transparent, auditable and grounded in statistical theory. Yet the pricing landscape is shifting. Machine learning methods - gradient boosted trees, neural networks and ensemble techniques - are now routinely outperforming traditional models on predictive accuracy, and insurers that ignore them risk adverse selection against better-equipped competitors.
Why machine learning outperforms traditional rating models
GLMs assume a linear relationship between rating factors and the response on some transformed scale, and interactions must be specified by hand. Machine learning models discover non-linearities and high-order interactions automatically. In motor insurance, for example, the interaction between driver age, vehicle power and annual mileage is notoriously difficult to capture with manually specified terms, whereas a gradient boosting machine will find it without being told where to look.
The practical consequence is more granular risk segmentation. Insurers using machine learning in technical pricing can identify pockets of mispriced risk that a GLM smooths over. When one market participant prices more accurately than its peers, the classic winner's curse dynamic follows: the less sophisticated insurer wins exactly the business it has underpriced.
The governance challenge: black boxes in a regulated industry
Accuracy alone does not make a pricing model deployable. UK insurers operate under the FCA's fair value rules and, for technical provisions, under Solvency II-derived requirements that demand models be understood, validated and controlled. A pricing model that cannot be explained to a regulator, a board or a customer is a liability regardless of its lift curve.
This is where the actuarial profession adds distinctive value. Techniques such as SHAP values, partial dependence plots and surrogate models make it possible to interrogate complex models factor by factor. A well-governed machine learning pricing framework typically pairs a champion model with an interpretable challenger, monitors drift in both data and predictions, and documents the rationale for every material modelling choice - exactly the discipline actuaries have always applied, extended to a new model class.
Where AI is changing the actuarial workflow
Beyond the rating engine itself, machine learning is reshaping adjacent parts of the pricing process. Claims triage models route claims to the right handler at first notification. Text mining extracts rating-relevant information from broker submissions in commercial lines. Telematics and IoT data streams feed usage-based pricing in motor and increasingly in home insurance. Large language models are beginning to accelerate documentation, peer review and regulatory reporting - not by replacing actuarial judgement, but by removing the mechanical drafting work that surrounds it.
What UK actuaries should learn next
The actuaries who thrive over the next decade will be bilingual: fluent in the statutory and professional framework that governs insurance, and fluent in modern statistical learning. In practice that means solid Python or R, an understanding of gradient boosting and regularisation, familiarity with model interpretability tooling, and enough MLOps awareness to ask the right questions about deployment, monitoring and retraining.
Crucially, none of this displaces the actuarial control cycle. Data quality, exposure measurement, credibility, and the economics of the underwriting cycle matter as much as ever. Machine learning sharpens the tools; it does not change the job, which remains turning uncertainty into decisions that a business can stand behind.
Conclusion
AI and machine learning are not a threat to actuarial science - they are its next chapter. Insurers that combine modern predictive methods with actuarial governance will price more accurately, respond to market shifts faster, and satisfy regulators more convincingly than those that treat the two disciplines as rivals. The future belongs to teams that do both well.