Even after just six months, the use of ‘legal grade’ Artificial Intelligence (AI) is starting to make highly manual legal documentation research much easier at UK law firm, Clyde & Co.
As a result, Damian Rourke, a partner at the London-based international law firm who deals with fraud claims, says:
What we're looking to do is remove as many non-value-added tasks within our space that we can, so our lawyers can then do lawyering instead.
The way Rourke and his team are freeing up the team is natural language processing in combination with OCR (Optical Character Recognition).
Like other Generative AI applications, the system works off a proprietary Large Language Model. However, unlike general purpose LLMs like OpenAI’s ChatGPT, Clyde & Co is using one built for legal-specific applications - Luminance.
Luminance claims to have trained the model on over 150 million verified legal documents.
That dataset is the fuel for what the supplier calls its Legal Pre-Trained Transformer (LPT) technology, which understands and ‘learns’ what its users want to know from a corpus of legal forms.
In Clyde & Co’s case, that’s sifting through thousands of pages of what it specializes in - fraud claims.
We do more insurance litigation than anyone else in the universe, and about five years ago clients started to approach us for digital solutions in the casualty claims space.
This was something of a challenge for the team, as it always concentrated more on the purely legal element of what it delivered, he says, and a lot less on the tech solution.
However, prompted by the request, the team created an Excel-based application extended by Microsoft VBA to show the potential of detecting fraud on our casualty claims. The aim being to help insurers calculate potential costs.
The firm’s potential customers liked the idea but disliked the lack of version control over the calculations, as well as the need to jump from one application to another.
That resulted, says Rourke, in a decision to create a more centrally controlled solution with functionality held on one platform that could be automatically populated with the right data and formulas.
To do so meant use of OCR to quickly ingest the required data from claims at speed.
The problem the team soon discovered was that most legal OCR suppliers were geared to aiding lawyers with less relevant workflows to Clyde & Co’s needs, such as disclosure.
Rourke’s colleague Ben Parsons - Head of Digital at the firm - says:
None of them had really approached the casualty market before and were all more geared towards the big ticket, high-value exercises that you might do on a huge piece of litigation; there wasn't really anything out there that had the flexibility to go over different types of documents and extract the concepts that we would want to extract to be able to automate processes.
However, the Luminance solution is much more flexible in learning concepts to quickly understand what specific information staff need to extract from documents, says Rourke.
It can also spot useful patterns without being told: no-one had specifically spelled out the concept of ‘time off work as a consequence of the accident,’ yet the system was very soon able to do this after only two days of training, and with a very high level of consistency.
For example, 19 out of the 20 key fields that the team needed to be captured were recognized in uploaded claim forms after just 20 hours of use.
The team can also show the technology just one example of a new concept (such as ‘injury prognosis’) and teach its meaning via a simple point and click tagging system.
Many data points
In UK law, a casualty case means that when presented with a claim for injury, specialists evaluate the claim to decide what level of compensation is or isn’t due.
To get there, what type of injury it was, what part of the body it affects, how long it takes to recover from and other medical data needs to be assessed.
That results in a large number of data points that need to be fed into any final assessment of what a claim is deemed to be worth.
But the medical reports that a specialist needs to read to try to get that information are really diverse in terms of their structure and content, says Rourke.
You end up with thousands of different medical experts all writing their own different reports in their own different style, and in their own different format.
Most non-AI PCR products are simply not able, he claims, to see that - even though ‘Dr. Smith’s’ reports use a different template than those of ‘Dr. Jones,’ the AI tool can look at both sorts of reports (and more) - but still be able to pull what Rourke calls “quite nuanced concepts and tags” from different reports.
That’s really useful, as we will always want the date of an accident, type of injury, the prognosis and so on, from a medical report no matter how it’s laid out. [We want to] extract it and automate the claims evaluation process as much as we can to save time.”
Human in the loop
So: is it working?
The answer is a definite yes, say the pair. Something that would take a person about half an hour to do, assuming a medium complexity personal injury claim, is now down to five minutes at Clyde & Co.
The company stresses that while it’s a machine pulling out the data and making the initial valuation, a human specialist still checks its findings.
But, says Rourke, his colleagues really appreciate the time saved by the system pulling out data from what can be thousands of pages of notes.
This enables us to gather data at a much more granular level - to capture it and then analyze it on a much more macro level - and look for trends and patterns in terms of how injuries have been put forward and presented. And that is massively valuable to our clients in terms of the insights we're able to deliver back to them.
And as the UK sees about 350,000 personal injury claims, states the company, the potential savings in terms of people hours and time could be very significant as well as accurate and consistent.
After having proved its usefulness in the claims arena, the next use of AI and OCR at the firm include applying the system to achieve similar levels of timesaving in other Clyde & Co legal processes, like case creation.
This also involves extracting data from often thousands of pages - and again, this is now being accurately set up in not hours but more like 22 minutes, claims Parsons.
That’s very promising, adds Rourke, as most parts of the firm’s processes are extracting information from documents and then populating its systems to initiate processes.
We are now working with our clients on what they send us and what we send back and how we can streamline those processes with our clients. That will reduce our client's costs by us reading and sending the data they want back to them, rather than sending documents or emails.
Ultimately, I've not met a lawyer in my years of working here or elsewhere that really enjoys completing this step.
Lawyers want to lawyer, and this is a fantastic way of letting them do only that.