Richer Signals, Better Pricing: Explainable AI and the Case for Understanding

And that’s where the real challenge begins. 

Accuracy alone isn’t enough. A model is only as useful as your ability to explain it. And more importantly, understand what to do with it. That’s where explainable AI (XAI) steps in. Done well, it gives pricing professionals the ability to challenge, calibrate and communicate their models with confidence. 

But XAI isn’t just a tool for governance or validation. It’s a lens. A way of seeing things differently. And when paired with rich data and the right domain expertise, it can reveal strategic opportunities hiding in plain sight. 

More Than Marginal Gains 

At Consumer Intelligence, we use a range of XAI tools like SHAP, HSTATS and partial or 2-way dependence plots to interrogate the behaviour of our proprietary pricing engine, Apollo. These tools help us understand not just which features drive predictions, but how those features interact and whether the model is responding to genuine patterns or just noise. 

But these tools don’t give us answers on their own. They’re most powerful when used alongside pricing expertise and domain context especially when supported by rich feature data that helps explain why a signal exists, not just where it exists. 

This is where our postcode enrichment layer, Atlas, comes into play. 

Making Sense of Risk with Atlas 

Atlas is our geospatial data engine,  built to describe the environment around each UK postcode using over 200 engineered features. These include public datasets from the Office for National Statistics, Department for Transport and Met Office, alongside proprietary engineered measures. 

These features span areas such as transport patterns, environmental stress, road network accessibility, and contextual indicators of traffic collisions. While some variables,  like commuting modes or local economic conditions,  derive from Census sources, others capture more external, structural conditions that influence how and where risk emerges. 

Importantly, Atlas doesn’t attempt to infer causality directly. But when used in combination with feature outputs from machine learning models, it becomes a powerful lens to explore and refine hypotheses about what might be driving certain pricing behaviours or performance patterns. 

For example, deprivation indices—summarised from various underlying measures,  are a familiar component in pricing. But when you can isolate and test specific subcomponents like long-term unemployment, educational attainment, or transport availability, you can better understand the likely causes of elevated risk in particular areas. And that gives pricing teams clearer options for refinement, segmentation or messaging,  not just rating. 

Similarly, Atlas includes airport proximity features. Collision data from the Department for Transport shows that the area surrounding major airports can be significantly riskier than the national average. Independent analysis by Angelica Solutions showed that injury-causing collisions near Heathrow were over twice as common per capita than elsewhere. While this kind of spatial correlation is interesting in itself, it becomes far more powerful when explored in the context of modelled uplift. It opens up discussions around potential causes,  like driver fatigue, unfamiliar vehicles, or increased congestion and how to manage or mitigate them. 

This kind of thinking isn’t about explaining the model for the sake of it. It’s about bringing together model output, real-world context, and pricing expertise to understand what’s really going on and what can be done about it. 

Why This Matters 

Correlation is the backbone of much of insurance pricing. But when we can begin to understand the cause, we can do more. Not just build better pricing, but help shape safer behaviours, fairer outcomes and more informed conversations across the business. 

Explainable AI tools help pricing teams do more than spot uplift. They help them make sense of it. They turn opaque outputs into understandable logic. And when used with geospatial enrichments like Atlas and tested against rating factors like age, NCD, mileage or occupation, they reveal relationships that can reshape how risk is viewed,  not just within pricing, but across underwriting, marketing and beyond. 

And that’s the real opportunity here. It’s not just about defending a model. It’s about informing the organisation. Helping every stakeholder, from analyst to underwriter to executive, understand what matters and why. So we can price with confidence, adapt with agility, and move from reactive modelling to proactive strategy. 

Because richer signals aren’t the end goal. It’s what we do with them that counts.