AI model flags 1.2 million ‘undefended’ buildings at flood risk in England

A new AI-driven flood modelling project has identified up to 1.2 million buildings in England that could be at risk of flooding despite sitting outside current flood defence and planning frameworks.

The Intelligent Flood Readiness Model, developed by Snowflake using geospatial data from Ordnance Survey (OS), combines detailed building-level data with government datasets and statutory Flood Risk Management Plans (FRMPs) to produce what the partners describe as a new “structural intelligence” layer for policymakers.

The findings suggest that a significant proportion of at-risk buildings are located in areas of high social deprivation, raising concerns about communities’ ability to recover from flood events.

Vulnerability beyond defences

By layering flood risk, mapping and deprivation data, the model estimates that as many as 68 percent of the identified buildings are highly vulnerable – not only exposed to flooding, but also lacking the social and economic resilience needed for recovery.

Much of this risk appears rooted in the age of England’s housing stock. Around 84 percent of the buildings flagged by the model were constructed before 2001, when planning rules began to more consistently account for flood risk. A further breakdown shows 15 percent date from before 1919 and 23 percent from between 1919 and 1959.

This reflects how both the built and natural environment have shifted over time, leaving legacy properties exposed to new or evolving flood risks.

Surface water risk dominates

The model also challenges common assumptions about flooding in England. It suggests that 85 percent of at-risk, undefended buildings are vulnerable primarily to surface water flooding, rather than river or coastal inundation.

This has implications for urban policy, as high-density housing – rather than riverside or coastal properties – may account for a larger share of at-risk households.

Geographically, areas along the east coast of Yorkshire and the Humber show the highest concentrations of vulnerable properties. However, the risk is widely dispersed: 37 percent of neighbourhoods contain at least one such building.

From static plans to dynamic models

The model integrates six distinct data streams, including OS building datasets, deprivation indices, Environment Agency flood risk data, and AI analysis of more than 3,000 pages of FRMP documents.

This approach aims to address a long-standing challenge for government: flood planning cycles are typically updated every six years and rely on relatively high-level data, while real-world conditions change far more rapidly.

By contrast, the model offers near real-time insights at neighbourhood—and potentially street—level, supporting more targeted interventions.

Tim Chilton, managing geospatial consultant at Ordnance Survey, said the model could help local authorities prioritise investment and better understand where protections are weakest.

Policy implications

The findings point to several emerging pressures for central and local government. Around 64 percent of affected buildings are residential, suggesting much of the recovery burden could fall on households. Meanwhile, 15 percent are commercial and 10 percent industrial, with a further five percent classified as infrastructure -assets that are often critical for recovery but may sit outside existing planning assumptions.

The project partners argue that protecting every at-risk property using current methods is unlikely to be feasible, reinforcing the need for more granular, data-driven approaches.

Among the recommendations are:

• shifting from area-based planning to building-level risk assessment
• targeting clusters of vulnerability across administrative boundaries
• increasing investment in surface water infrastructure such as drainage
• introducing “vertical” risk assessments for high-rise buildings
• factoring social deprivation into resilience planning

Towards digital twins

The model reflects a broader shift in digital government towards predictive, data-led decision-making. By combining AI with frequently updated geospatial data, it raises the prospect of “digital twins” that simulate how flooding could impact specific locations before events occur.

Fawad Qureshi, global field CTO at Snowflake, said the project demonstrates how integrating fragmented datasets can improve policy insight.

Rather than replacing human decision-making, the partners position the model as a tool to “ask better questions”, stress-test plans, and identify gaps in existing flood strategies.