Challenge
Matching different versions of the same UK address can prove to be a persistent data challenge. A single property can be recorded in several different ways across customer applications, internal systems, and external datasets. Addresses could contain abbreviations, spelling mistakes, reordered words, missing postcodes, or different descriptions of flats and other units.
Traditional matching methods can fail when two records describe the same property in different ways. Matching addresses can be simplified with automated processes, but it currently still requires manual intervention throughout, and remains an ongoing and complex challenge for financial organisations like LBG, and numerous other sectors.
LBG therefore wanted to develop an AI-based address matching tool; one that could achieve a high degree of accuracy for the varied structure of British addresses, while recognising uncertainty and referring inconclusive cases for human review.
LBG also wanted matching decisions to be transparent and auditable. Rather than allowing an AI system to make an unexplained guess, it needed to show why an address had been accepted, rejected, or referred for further investigation.
Solution
Lloyds Banking Group developed PropertyLM-UK, an innovative address resolution system that combined AI with deterministic checks based on the defining elements of a property address.
The system integrated:
- OS Open UPRN, an open-source dataset, which provided the location of properties through the Unique Property Reference Number
- AddressBase Premium, which supplied authoritative, UPRN-linked addresses for testing and benchmarking
- Energy Performance Certificate (EPC) data, which provided the reference property information used by the matching system.
The Unique Property Reference Number (UPRN) provided a consistent identifier. Rather than relying on address text alone, the system aimed to connect different representations of an address to the UPRN for the same physical property.
This created a stable property identity that could be used to join information across datasets and processes, even when the written address varied. It also helped distinguish between addresses that appeared almost identical but referred to different properties.
Combining AI with clear rules
PropertyLM-UK used AI to understand address variations and identify likely matches. It searched a national index containing more than 18.2 million property records and assessed which candidates were most likely to represent the address submitted.
The system then applied strict checks to critical address components, including the postcode, primary property number, and a flat, apartment, or unit identifier.
If one of these identity elements clearly contradicted the query, the system prevented the candidate from being accepted, regardless of how similar the two addresses appeared to the AI.
This separation was central to the innovation. AI handled differences in language, formatting, and spelling, while deterministic rules controlled the final property identity decision.
Knowing when not to match
The tool did not force every query into a match. It returned one of four outcomes:
- confident match
- probable match
- ambiguous and requiring review
- no match
Where evidence was incomplete or several properties remained plausible, the case could be passed to a person for review. This 'human in the loop' approach provided a failsafe, reducing the risk of an automated system making an unsupported decision.
Every result also included a log of reasons. This showed, for example, whether the postcodes agreed or whether the unit numbers contradicted one another. The record created a transparent audit trail that could help reviewers understand and reconstruct an individual matching decision.
Designed for secure deployment
PropertyLM-UK is designed to operate within an organisation’s own infrastructure, using open AI models and without sending addresses to an external AI service.
This keeps commercially sensitive and potentially personally identifiable address data inside an organisation’s environment. The system could also run on standard CPU hardware without task-specific AI model training, lowering the technical barriers to reproducing and deploying the approach.
Plus, the 'human in the loop' function means AI isn't running independently in the background, and organisations can still intervene, when necessary.
Result
PropertyLM-UK demonstrates how AI and UPRNs could turn inconsistent address text into trusted property identification.
When evaluated against 10,000 UPRN-labelled test queries generated from OS AddressBase Premium, the system identified the correct UPRN as its first result in 94.6% of cases where the property was present in the reference index. Its top five results contained the correct UPRN in 97.1% of those cases.
When the system declared a match to be confident, it achieved 99.64% precision. This meant that its highest-confidence matches were correct in more than 99 out of 100 cases within the benchmark.
The system was also tested against addresses for properties that were absent from its reference index, including new builds or properties without a relevant record. It correctly avoided making a match in 92.6% of these cases, either returning no match or sending the address for review.
The tool processed a query in a median time of 352 milliseconds on standard CPU hardware.
Benefits
This AI-driven approach has given LBG a foundation of UPRN-linked property identities, one that can be used with confidence that information from different systems relates to the same physical property.
PropertyLM-UK can connect address records with Energy Performance Certificate information, helping to identify properties and customers that might benefit from energy-efficiency improvements, and support the development of greener financing propositions.
It also has potential applications across the rental and commercial property markets, from customer onboarding onwards. It could help connect properties to landlords, distinguish between multiple units within one building, and understand complex premises with multiple owners or occupiers.
By combining OS data, UPRNs and AI, Lloyds Banking Group has demonstrated how innovation could improve a foundational process, and potentially revolutionise address matching, to make it much more efficient while still highly accurate and secure.
"OS data gave us the accurate, authoritative foundation we needed to explore a new approach to address matching - with the extra advantage that we could use OS's opensource datasets. By combining UPRNs with AI and clear validation rules, we created a solution that could recognise different versions of the same address while protecting against incorrect matches. This gave us greater confidence in the property data that could support our customer processes and sustainability ambitions."