Transforming Contract Review with Machine Learning
Duration: 9 months
Team: 8
Technology: Machine Learning
Introduction
Our client, a large US-based law firm, sought ways to speed up its contract review processes. With a vast volume of documents to analyze, the firm was looking for a solution that would allow extraction of critical contract provisions, enabling lawyers to focus on higher-value tasks while reducing the manual workload.
Engagement Challenges
While there were many off the shelf solutions available for various pre-signature activities like contract drafting and signature workflows, there was no simple way to extract information from existing signed contracts. Traditional text search based technologies proved to be ineffective, since the contract language had significant variations in the words and phrases used.
Engagement Approach
Our solution involved leveraging Natural Language Processing (NLP), a machine learning method to train a model using the firm’s archive of existing contracts. As a first step, these contracts were manually annotated by human subject matter experts.
The next step was to train the model against the annotated data. The contract text is cleaned and tokenized into smaller words/phrases. This allows identification of key terms, clauses, names, dates, and other such elements, and assigning probabilities. This would then be output to the human experts for review. This review and feedback cycle was critical for the further model training and improvement.
We also built a custom user interface, allowing easy upload of documents and reviewing the output. The solution was hosted on the firm’s existing IT infrastructure, for data security and compliance.
Engagement Outcomes
The implementation of this machine learning-powered contract review system provided significant and measurable benefits to the client. There was more than 50% reduction in the time spent by the legal team in reviewing contracts, saving thousands of person hours. There were also fewer errors post implementation, compared to prior text search based processes.