Despite being at its core a knowledge industry, the legal profession has been remarkably slow to adopt information technology outside of online databases such as LexisNexis and e-discovery software. The conservative nature of the profession, the intensive training that focuses on developing individual skills and knowledge and the lack of incentive for efficiency that is built into the hourly billing model all contribute to that resistance. Over the last few years, however, numerous artificial intelligence (AI) solutions have been developed for legal use, and the profession has begun to embrace, or perhaps be embraced by, those tools.
A number of forces have converged to catalyze this market. First, the software products meet specific, well defined needs and therefore have been well accepted. Second, they provide efficiencies that reduce costs, which have been under considerable downward pressure. Those same efficiencies help lawyers, particularly first-year associates, get some relief from tedious, repetitive tasks. Finally, law firm clients who are using software products to increase productivity in their own work have begun to question why their lawyers are not doing the same thing.
Mitratech estimates that the U.S. market for legal software is $3 billion, split evenly between law firms and corporate law departments, and estimates that the potential market is about $16 billion. The report does not break out AI products as a separate category and those products constitute only a small percent of the legal market at this point. However, legal AI products have significant potential and are likely to grow more dynamically than established technologies such as document management and matter management.
Many legal AI products are based on technologies that are well established in knowledge management, including text analytics and business process automation. What pushes them into the AI category is the degree to which they incorporate intelligence into their functionality, including the ability to learn and to process natural language inquiries. Because they are focused on issues specific to the legal profession, they are working with a finite knowledge domain, which allows them to be more precise in their performance.
Typically the companies that developed the products are relatively new firms or startups. Often they were established to meet needs observed by the founders in their previous law practices. The products do not attempt to replace lawyers but rather to augment their skills and to automate that portion of the task best supported by computers.
Extracting contract information
Kira Systems is a Canadian firm that was co-founded by Noah Waisberg, a lawyer who had been reviewing contracts and supervising other lawyers who were doing the same task. “People were looking for the same things over and over,” says Waisberg, CEO of Kira Systems. “The work was tedious and it was easy to make mistakes.” He could see potential for a product that searched contracts for the required information and presented it in a systematic way.
Alongside co-founder Alexander Hudek, a computer scientist and CTO of Kira Systems, Waisberg developed a software product that extracts data from sets of contracts (and other documents), saving anywhere from 20 percent to 90 percent of the time required for manual review. “We had one client who needed to check for 35 data points from several thousand agreements,” Waisberg explains. “Using human review, each person could have reviewed about 10 to 20 documents a day. With Kira, they got through 100 per day per reviewer.” In addition to being used for contract analysis, Kirais also deployed on due diligence and lease abstraction projects.
Out of the box, Kira can automatically find more than 100 data points, and users can teach Kira to identify custom information. To teach the system to find new information, the user simply highlights relevant phrases in a training set, typically ranging from 15 to 30 examples and up. “The software can then identify similar phrases from other documents, even if the same wording is not used,” Waisberg says, “because it is looking for meaning rather than keywords.” The software provides a matrix summarizing how many hits were found for each category for each document, and the user can drill down in each cell of the matrix to see the details and can also export results.
Deloitte has been using Kira for almost two years, first for internal use to analyze client documents and then by client teams for client support. Deloitte also uses Kira for its audit practice. When Deloitte first licensed Kira, it came with about 30 categories already available, and Deloitte trained it for approximately 1,000 more. Deloitte has customized the platform as Argus for its audit practice and D-ICE for its consulting practice, using its own models in combination with Kira’s platform to address a growing range of corporate and client needs. The company credits Kira Systems with helping it to win about $30 million in new business.
The founders of RAVN Systems, a British company, have backgrounds in search, engineering and data science. The company provides two major technologies, an artificial intelligence platform to analyze and summarize information and a search application. Its Applied Cognitive Engine (ACE) focuses on the legal industry, including corporate law, and is also applied to finance and real estate.
RAVN’s cognitive computing capabilities for the legal industry include a Refine application that categorizes information in large data sets; Extract for Due Diligence, which automatically extracts information in preparation for a financial transaction such as a merger or acquisition, and Extract for Contract Analysis. In addition to organizing and distilling information, the Extract function asks a series of iterative questions to aid in the interpretation of document sets.
The Extract capability can determine relevancy for such factors as whether information is subject to attorney-client privilege. “It can take a very long time to do that manually when people have to search through millions of documents,” says Peter Wallqvist, chief strategy officer of RAVN. “After about 1 percent or less of the total document set has been used as a training set, the system can classify information and test its own model. It will fairly quickly become as good or better than a person and definitely much faster.”
The technology has gotten to the state where it can make deductions and judgments on text that in the past might just have been searched. “What people are after is not a list of documents but a specific piece of knowledge in a document,” says Wallqvist. “We focused our R&D on surfacing that information.” For example, ACE can filter on rent amount in an unstructured document, and rent then becomes a category for metadata into which rent amounts can be placed.
British telecommunications giant BT announced that it has been working with RAVN for the past year using its technology to review contracts. The contracts are checked for accuracy and to make sure no changes have been subsequently made, and are compared to invoicing data to make sure the two are consistent. BT estimates that RAVN is saving it tens of millions of pounds a year.
An unusual feature of the ACE platform is that it incorporates a graph database, which shows how data objects are linked. Normally a separate application, the graph database in ACE is meshed with the search function. “Exploiting links in a graph is essentially what makes Google so accurate,” says Wallqvist, “because that approach makes connections a relevant search feature, whereas search has traditionally focused on term frequency.”