A Beginner’s Guide to AI: Part Two

How might legal services be impacted by the development of AI? Robert Peat, D2LT Consultant, unpacks what you need to know about AI in legal services for the second part of our Beginner’s Guide to AI.

Guide to Legal AI

In Part One of our series on artificial intelligence (“AI”), Rachel Scanlon, our Asia Pacific Lead, outlined that AI is the use of computer systems to perform tasks requiring human levels of intelligence, including visual perception, speech recognition, decision-making and translation of texts. The systems “learn” from previous experiences to improve future performance. As Rachel outlined, AI has had a positive and increasing impact on many parts of financial services. Nevertheless, the provision of legal services has not yet been impacted to the same extent as financial services and many other areas. As a conservative profession, the legal profession has been slow to welcome the use of Legal AI.

Nothing natural about NLP

Natural Language Processing (“NLP”) is an integral part of AI. Essentially, the study of NLP is aimed at computer systems being able to understand natural language, i.e., the written word. As a result, the system can take the natural language and proceed with the task it is being asked to complete. NLP seems marvellous in principle, and its potential uses for legal services are evident given the document and word heavy nature of most legal tasks. It would be fantastic for lawyers (junior lawyers especially!), if manual tasks such as conducting document due diligence could be completed by an AI computer system.

The most revolutionary recent use of NLP is Google’s creation of a machine learning technique for NLP using bidirectional encoder representations from transformers (BERT); essentially a model that processes words in relation to all the other words in a sentence, rather than one-by-one in order. Known as BERT to us mere mortals (named like the Muppets character!), BERT is trained on representations from unlabelled text by jointly conditioning on both left and right context. In simple terms, BERT learns from both the words to the right and the left giving it a deeper understanding of the linguistic meaning.

NLP usually requires a huge volume of text to train computer systems. Fortunately, BERT comes pre-trained with over 3,000 million words of unlabelled text. BERT, therefore, could be used to reduce the burdensome tasks of lawyers by analysing huge volumes of text accurately and efficiently. Unfortunately, the pre-trained BERT Google has developed performs poorly when attempting NLP tasks in the legal sphere. This is due to the specific type of language used in legal services. For example, it is not uncommon to find Latin phrases in legal cases and complex syntax in contracts. Moreover, each sub-division of the legal industry has a hyper-bespoke lexicon of phrases.

Accordingly, firms have recognised that there is an ability to build on BERT and create a version for niche areas – for example D2LT’s work in respect of OTC derivatives master agreement documentation. There is precedent to this approach.[1] Using BERT as a foundation to then introduce further training of the NLP model on legal opinions and master agreements offers a superb opportunity to create a specialised legal BERT for financial services law. I’m sure many junior lawyers eagerly await the day when BERT will be working the late nights and not them…

A force for good

More often than not, the complexity caused by changes in law or regulation is not the change itself, it is understanding the scale of the impact. A recent example of this is IBOR, where the principal legal change was the cessation of IBOR publication; meaning it could no longer be used as contractual reference rate. The complexity arises due to the vast volume of documentation which may reference the defunct rate and needs to be transitioned to a new rate. Previously this would result in an expensive document review project which could take many months to complete. Legal AI has offerings to change this. Vendors such as Eigen, Kira Systems, LIKEZERO and Luminance have been able to provide institutions with cost savings and increased accuracy through their products, which can identify pre-trained sections of contracts and legal documents. The systems do not remove the need for human oversight, but they certainly do allow for an expedited process and reduce the number of experts required. They also serve to highlight the enabling function of Legal AI at this very moment.

Equally, by reducing expert resources required, these enabling tools mean current legal service providers (law firms) can be more readily challenged by alternative legal service providers (“ALSPs”). By opening the legal services market to ALSPs, it is likely that we will see an accelerated uptake and use of Legal AI due to the propensity of ALSPs to use more innovative methods.

Predictive Text

There are now several projects underway in the legal capital markets industry that aim to ensure a standardised use of language for contractual elections and amendments. Two model examples are the ISDA and ISLA Clause Library projects, which are looking at addressing their respective industry master agreements. Building on this, there is also the development of platforms such as AXDRAFT, Sirion Labs, SmartDX and ISDA Create which seek to provide intelligent document assembly functionality and move contract negotiation online, developing a more universal and standardised approach to capital markets documentation.

Currently, these tools and projects require a high degree of manual involvement. Hence, this area offers another fantastic opportunity to utilise developments in NLP such as Legal BERT. Once a comprehensive Legal BERT has been developed it is probable that an AI system could be given specific factual parameters (i.e., the counterparty’s jurisdiction of incorporation and its regulatory classification) and produce a master agreement draft. This AI system of the near future could accurately utilise the most recent legal opinions, and previous agreements with similar entities to produce accurate master agreements with minimal human intervention. In a future state, we can imagine that lawyers will be able to draft contracts and have language suggested to them which is based on the most informed legal advice and market standard in real time.


There are many opportunities for the legal services sector to embrace AI. As we have suggested, there is already important work ongoing in this field and the role firms like D2LT can play is clear. The development of legal AI will stem from the innovative work being undertaken by D2LT and others. It is an exciting time (even if your name isn’t Bert!).

[1] ‘LEGAL-BERT: The Muppets straight out of Law School’ Ilias Chalkidis, Manos Fergadiotis, Prodromos Malakasiotis, Nikolas Aletras, Ion Androutsopoulous. (https://arxiv.org/abs/2010.02559 )

Share and Enjoy !