How AI is being used to digitise property records across England and Wales

The problem

Across England and Wales, individual councils hold millions of local land charge records. A local land charge is an obligation or restriction on a property that often affects how it can be used, such as conditional planning permission, listed-building or conservation-area status, or a tree preservation order. A search of these records is a normal part of buying a property, and it shows whether the property is subject to a registered charge.

Before this information can be added to HM Land Registry's central Local Land Charges (LLC) Register, every record must be checked for accuracy and converted into a consistent digital format. The originals come in mixed formats: paper, electronic files, and scanned images.

Done by hand, this is a slow and labour-intensive process. A single council's migration can run to months of work for a large team. The greatest cost comes when a mistake slips in during data entry and is only caught later, at the quality assurance stage, because the affected records must be redone. That rework is the most demanding and expensive step in the whole process.

The solution

To speed this up, HM Land Registry's Data Science team built an AI tool that finds relevant local land charge information in a council's historical records, extracts it, and structures it for the register, rather than doing so by hand.

The tool combines three parts. An upgraded Optical Character Recognition (OCR) engine, software that reads text from scanned pages and images, performs the initial capture. A Large Language Model (LLM) is a type of Artificial Intelligence (AI) that can interpret and work with large amounts of text, then process that text and structure it. Finally, a set of purpose-built models fine-tunes the result for accuracy in this particular setting.

The tool is built to avoid errors that are common in manual data entry, such as misreading a character, for example, confusing a letter with a similar-looking number. It also has controls that keep stray page elements, such as menus picked up from a scanned document, out of the finished output, and that stop the AI from inventing data that was not in the original.

Once the tool has processed a batch of records, a team member checks its output, and a separate quality assurance team verifies it independently.

The tool reached its current form over more than five years of steady development. The earliest versions had to be coded specially for each council and ran on individual workstations. The team gradually replaced these one-off builds with a single, adaptable system, hosted on HM Land Registry's secure cloud, that can be pointed at a new council's records without being rebuilt each time.

The impact

1. A three-month job done in four weeks
Newham was the first council to run its text-based records through the tool. The programme had expected the work to take 20 staff three months; with the tool, four people completed it in four weeks.

2. No quality failures so far
Every batch of Newham's records cleared its quality checks at the first attempt, and across all the councils handled to date, nothing has fallen short on data quality or completeness. Because rework after a quality failure is the most time-consuming and costly part of a migration, avoiding it altogether yields the greatest savings.

3. Extended to more councils
Since Newham, the tool has been run on two further councils, Dacorum Borough Council and Cotswold District Council. The team has lined up other councils whose records are of a similar kind and is working to extend the tool further, including for handwritten material and lower-quality scans.

4. Staff freed for more complex work
With the tool handling the bulk of the extraction, staff who were previously tied up with manual data capture have been able to move on to work that calls for human judgment and expertise.

Lessons learned

• Long-term, iterative development is what got the tool working. Its current form is the result of more than five years of refinement, with each version improving on a more limited predecessor that had to be rebuilt for every new dataset. Reaching a reliable, reusable system took sustained investment and a willingness to learn from each attempt.
• Accuracy depends as much on the controls around the AI as on the AI itself. The safeguards against misread characters, stray content, and invented data are what keep the output passing quality checks. Building those controls in and checking every output before it is accepted is what makes the tool valuable for the quality of the result, not just its speed. The tool does the volume, but a person checks every output, and a separate quality assurance team verifies it. That combination is what lets the work move quickly while still meeting the standards the register depends on.
• Building for reuse can make scaling easier. Moving from a custom solution for each council to a single adaptable system meant each new migration could build on the last rather than starting over. For any programme that has to repeat a similar process across many organisations, investing in reusability early reduces the cost and time of each additional rollout.

This case study was written with assistance from artificial intelligence.