With law firms around the world turning to artificial intelligence solutions to provide a competitive edge in the digital age, a 2016 Deloitte study—entitled Developing legal talent: Stepping into the future law firm—predicted that 2020 will mark a “tipping point” for the legal profession as AI and other disruptive forces accelerate industry change. Looking at the U.K. market, the study noted that automation had already had a significant impact on industry employment, eliminating thousands of secretarial jobs while creating new high-tech positions. More dramatic change should be expected as automation shifts to associate-level work. By 2025, Deloitte predicts “a profound transformation of the profession due to the quickening pace of technological developments, shifts in workforce demographics and the need to offer clients more value for money.”
When it comes to the practice of law, it is unclear whether innovation drives efficiency or whether the need for increased efficiency drives innovation. Either way, as we contemplate the future, it is important to remember that the legal world has survived tech-induced seismic shifts in the past. It is also important to note that the industry’s previous adoptions of new technologies increased the efficiency of individual practitioners and law firms alike.
Given the sector’s previous experiences with new technologies, you could even argue that the eventual widespread introduction of AI solutions is likely to be more seamless than the shift from typewriter to word processor (which enabled the instantaneous modification of written text in lieu of redacting, retyping, or scrapping it altogether) or the shift from written correspondence to fax and then email (which enabled real-time communications and exchange of information), or the shift from library research to electronic databases and online search engines (which rendered trips to the courthouse library redundant). That said, we are in the early days of legal AI, and regardless of how law firms integrate the technology (most are licensing third-party solutions, while some develop proprietary systems), the promised disruption has limitations.
This article examines the current state of legal AI, highlighting current applications and future potential along with common concerns.
In general parlance, the expression “artificial intelligence” is without a precise definition and refers to a broad range of technological concepts. In its most basic form, AI is a machine’s ability to perform tasks that normally require human intelligence. Computers can achieve “artificial intelligence” through various technologies that aim to mimic or surpass aspects of human cognition: computer vision, natural-language processing, and machine learning (discussed below), paired with big data and modern processing power. In this regard, AI is not a single technology but rather a combination of technologies that can achieve varying degrees of what we might consider “intelligent behaviour” (e.g., from text recognition to more complex predictive analytics).
Machine learning is a subset technology of AI and is likely the most common application of AI as it pertains to tools currently available for the legal profession. Machine learning refers to the ability of a computer system to “learn” on its own by analyzing large datasets and identifying patterns that eventually allow it to draw conclusions and make probability-based predictions applicable to similar data.
While we expect to see boundless applications of AI to the practice of law in the long term, current applications generally aim to facilitate contract review, legal research, and the prediction of legal outcomes.
Due diligence is an essential phase of every M&A transaction, and yet it is arguably the costliest and most tedious phase, subject to tight deadlines. Corporate lawyers spend countless hours sifting through a data room (nowadays, generally a virtual data room) replete with contracts of the target corporation. The exploitation of AI to facilitate the document-review process could have a material impact on the costs, amount of personnel, and time to be devoted to the due-diligence phase, and in turn, would likely decrease overall transaction fees and timeframes.
AI software in this realm (from solution providers such as Kira Systems, Luminance, eBrevia, Leverton, iManage RAVN, and DISCO) is already at the stage where it is seeing widespread adoption by major law firms. For instance, Toronto-based Kira Systems’s flagship software, Kira, has been widely adopted by large Canadian firms. Kira is capable of performing automated contract analysis, en masse, and relatively quickly. The general process is as follows: (i) the lawyer uploads the trove of documents to be reviewed to a secure cloud; (ii) the AI system converts the documents into a machine-readable format (using optical character recognition); (iii) the lawyer selects a preset interpretation model best suited to the type and subject matter of the documents to be reviewed; (iv) the lawyer selects a number of target attributes to be identified (e.g., change of control provision); (v) the AI system automatically identifies and tags all the fields and clauses that it believes match the target attributes; (vi) the lawyer reviews each of the tags to confirm accuracy and to remove false positives; and (vii) the AI system formats the results into an easy-to-read table format that the lawyer can then import into or append to the due-diligence report.
Trend analysis is another key part of contract review. Figuring out what terms of negotiated transactions are currently trending in a particular “market” is a common task for corporate lawyers. The answer generally comes with experience, access to a large number of contracts negotiated within a particular timeframe (generally a year), and a great amount of effort to interpret the contracts and tabulate the results. From time to time, legal associations release deal point studies that provide market metrics of key negotiated legal issues specific to a jurisdiction (e.g., the United States, Canada, or Europe) and type of transaction (e.g., private target vs. public target). On a rare occasion, a large firm may dedicate resources to manually produce a smaller-scale, internal study. A by-product of the ability to analyze contracts consistently, en masse, and quickly is the ability to identify trends and variations and gather statistics on the usage of particular clauses. For example, an AI system would be able to identify the presence or absence of a specific type of provision (e.g., change of control; materiality scrape) and the value of select terms (e.g., purchase price; indemnity thresholds; governing law). Firms with access to large amounts of transactional information (i.e., contract data) would be able to more easily conduct internal studies to benefit their lawyers and clients. National and international firms would particularly benefit from such a use of the technology, as they would generally be privy to a relatively large set of transactions in a given period.
AI systems are also particularly effective with respect to high volumes of standard-form contracts with relatively low variability between contractual terms, making such systems appealing candidates for use by corporations to manage large volumes of customer and supplier contracts. Consider the stresses placed on an in-house legal department that, traditionally, may have to manually vet thousands of commercial contracts annually. Now consider the cost and people-power savings where an AI system could review the bulk of the contracts (particular those in standard form) and tabulate the results in seconds. The anticipated cost savings were so great that JPMorgan developed its own AI system, COIN (short for “Contract Intelligence”), capable of recognizing approximately 150 attribute clauses, in order to facilitate processing of the bank’s over 12,000 standardized credit contracts per year. Similarly, an analysis of payment terms can help an organization optimize contractual payments by identifying which contracts provide for early-payment discounts and which contain late-payment penalties. Legal AI may also benefit corporate compliance departments, where there is often a need to ensure that the company’s current contracts meet specific regulatory obligations and industry standards.
AI-powered legal research tools aim to improve upon the previous generation of research resources by facilitating user input, assessing the relevance and persuasiveness of results, and anticipating the user’s research path. AI-powered functionality has been embraced by traditional market participants (e.g., LexisNexis, Westlaw) and new entrants (e.g., ROSS Intelligence). Research providers have turned to naturallanguage-processing techniques to allow users to ask a research question using plain language, instead of having to resort to precise and often frustrating syntax (e.g., Boolean operators). A research search engine can interpret a broader meaning from search terms, resulting in relevant search results that might not have appeared if one had only searched for a particular “keyword.” There is also a focus on presenting the most relevant results without overwhelming the user with a deluge of applicable case law and secondary sources. Search results can pinpoint the most persuasive text from a case in the context of one’s argument. Additionally, AI systems are capable of analyzing a document to automatically update any cited sources and to identify missing, key resources. The latter can be particularly useful when analyzing an opponent’s argument for weaknesses.
“Predictive AI tools can identify whether a particular case is likely to win in a particular court, whether it would be better to settle before trial, and if it does go to trial, a recommended strategy to pursue.”
Owing to its ability to assimilate and analyze large amounts of data, AI can “learn” to identify patterns in the outcome of legal cases and provide key analytical information to optimize litigation strategy. Providers of litigation analytics services (e.g., Lex Machina, Westlaw Edge) gather and process public court and docket data. Based on the type of case, the venue, the judge, and the general legal issues involved, predictive AI tools can identify whether a particular case is likely to win in a particular court, whether it would be better to settle before trial, and if it does go to trial, a recommended strategy to pursue before the court and presiding judge, as well as the expected timeframe for the case. Such information can be very useful when managing client expectations.
Benefits, Concerns, and Limitations
As mentioned above, over the long run and based on sufficient training data, AI systems are likely to reduce costs, produce relatively consistent and accurate results, and do so in considerably less time than ever before. They also alleviate a number of human resource limitations. AI systems do not suffer from the law of diminishing returns (e.g., fatigue), making them ideal for repetitive tasks across a high volume of documents. They are also able to review almost any number of contracts, eliminating the need to prioritize high-value contracts at the detriment of not being able to adequately service the company’s other contracts.
There are, of course, issues that need to be addressed. The following concerns are common to the general proliferation of AI, but they warrant particular discussion with respect to legal AI.
As with the adoption of any technology, implementation costs can be high. For firms that have opted to develop their own AI systems in-house, at least one has done so with the backing of a client as part of a cost-saving initiative that would benefit both parties over the long term. Major firms that have enlisted third-party contract review services have likely done so by negotiated contract, on a tiered basis, often dependent on the volume of documents uploaded/processed or the number of contractual provisions tagged. While the costs are not trivial, there are still considerable cost savings to be had when one considers that, on average, fewer lawyer resources will have to be deployed and for shorter periods. Small and mid-sized firms may find the costs associated with the implementation and training of an AI-based contract review system prohibitive. Further, cost savings will generally increase as transaction sizes scale up, which is less applicable to smaller firms. Nowadays, legal research services are generally offered on a per-user or per-firm flat fee basis. For practices that regularly rely upon case law research and secondary resources, these services are generally considered cost-effective.
Many have voiced a concern that AI technologies will supplant human capital. While this may be true with respect to those mundane and repetitive tasks for which we rely upon human labour (e.g., basic due diligence; contract review), it is less so for more sophisticated tasks for which we rely upon human judgment (e.g., persuasion and strategy). In addition, legal AI requires professional resources for training, supervision, and the revision of AI work product. In this respect, legal AI should be seen as a powerful tool to facilitate the practice of law, allowing legal professionals to focus on more stimulating activities.
For legal AI to be effective, it must engender trust in the reliability of its results. It is well recognized that biased data can lead to biased AI. The legal profession should adopt a systematic approach to address bias early on, particularly when training AI systems. As much as practicable, the algorithmic decision-making process requires transparency, in order to allow professionals to evaluate and challenge instances of bias. The need for such transparency will become ever apparent as legal AI sees increased adoption in the justice system (e.g., predictive risk assessments in bail determination; sentiment analysis in forum and jury selection).
As for expectations, legal AI technologies suffer from some notable limitations. It is hoped—and somewhat expected—that machine-learning AI systems will deliver more accurate results over time with the benefit of experience. But there are reasons why this is not necessarily the case.
First, as is common to machine-learning AI systems, the system’s accuracy is generally dependent on it being trained on massive amounts of data and subsequent human verification of the results. Current contract review software may come pre-trained on certain types of contracts (e.g., corporate; real estate; financial services) and may allow for firms to train the software based on their own contract archives. For accuracy, such training is generally supervised by subject-matter experts. While it may seem advantageous to train the software on high volumes of material to increase clause-recognition accuracy, firms are often reluctant to dedicate lawyer resources to do so. Further, the marginal increase in accuracy from increased training appears to level out relatively quickly (based on our discussions with programmers of AI-based contract review software).
Second, legal AI is hampered by the segregated nature of legal data and the possessive nature of its owners. For many reasons, including a firm’s confidentiality obligations, work product is rarely shared outside of the firm or its clients, and is mostly shared to the limited extent required to execute client mandates. The proprietary nature of legal data imposes a direct limitation on the ability of an AI product to learn from the aggregate data upon which it was trained. Instead of being able to fully exploit the pool of contracts across all of the AI provider’s clients, an AI system is generally implemented on a firm-by-firm basis. Where a technology operates under a cloud-based software-as-a-service model, any firm data uploaded to train the AI system is generally stored in a sequestered AI cloud, provided solely to benefit the firm and its clients, and no others. Even if data is to be anonymized and shared, the types of commonly used contracts, the laws governing their interpretation, standard practices, and even document language can vary by jurisdiction. Consequently, any self-learning functionality performed by AI systems would most likely have to be contained to a specific jurisdiction.
Third, by itself, an AI-based system is not a silver bullet for the contract-review process. In practice, a significant amount of human oversight is still required to prepare the materials for processing, vet and interpret results, and clean up the resulting reports. For instance, AI systems often rely upon a professional to verify the applicability of flagged provisions. While AI systems are most efficient at identifying relevant language in large deposits of documents with low variability between them, due-diligence document deposits often include non-standardized contracts. In such scenarios, AI contract review systems are subject to non-trivial rates of false positives, where a keyword taken out of context results in the errant flagging of an unrelated provision. Based on discussions with users of AI-based contract review software, accuracy also appears to be higher in reviews targeting simpler contractual clauses that employ standard keywords (e.g., termination), as opposed to more complex clauses that may take on myriad forms (e.g., most favoured nation provision).
Fourth, there is a real risk that once AI-assisted technologies become ubiquitous, their results may be factored into the legal strategy of all participants. For instance, if commonly accessible litigation analytics tools determine that it would be better to settle an ongoing case, opposing counsel may enter the settlement negotiation with this knowledge in hand, and may use such knowledge to extract a more favourable settlement. Similarly, the analytical and strategic information may also be available to decision makers (e.g., judges), which raises issues as to the extent to which courts should be able to use such information (exploration of such issues is outside the scope of this article, but we encourage others to delve deeper into the ethical and procedural issues that may ensue).
Finally, the law is not static. It changes over time with the evolution of court doctrine and new legislation. Similarly, market standards for negotiated transactional terms vary from one year to the next. Consequently, basing AI-driven insights on outdated law and trends may yield inaccurate results and recommendations. Methods of tempering the impact of outdated law would be for AI systems to lend more weight to more recent decisions and trends, to have subject-matter experts regularly monitor the AI algorithms and results for accuracy, and to eventually have the AI system update itself based on new jurisprudence and market standards. (The ability to identify bad law already exists, to a limited extent, in legal research features that identify case law that has been implicitly undermined due to its reliance on overruled or otherwise invalidated cases.)
So, Where Do We Go From Here?
Legal technology has featured image-recognition capabilities (e.g., OCR) for years, and the more recent AI-labelled technologies have increased the functionality and value of such techniques. But current applications are focused on finding and filtering information in an efficient manner.
Once the base AI technology has proven itself, there will be improvements to the user interface (UI) and user experience (UX). For instance, we expect to see the implementation of speech recognition in combination with enhanced natural-languageprocessing algorithms, whether in the form of an input method for existing software or voice assistants. Such assistants or chat bots could be useful in walking lawyers and paralegals through legal procedures and search results.
Taking this one step further, by combining image recognition, speech recognition, natural-language-processing techniques, robotics, and a variety of legal applications, the realization of a robotic legal assistant becomes a distinct possibility.
The vast majority of practitioners have yet to use the latest generation of legal AI tools. And as law firms contemplate how to integrate legal AI into their existing IT infrastructure, it will take time for providers to deploy comprehensive, cost-effective solutions that fully utilize the potential of this technology—and for the legal profession to accept them. But once the tipping point is reached, clients will very likely come to expect the associated benefits that such technologies bring, precipitating widespread adoption of legal AI.
In the meantime, as the Deloitte study advised, law firms “must have a clear strategy for dealing with these changes now if they want to remain competitive and ensure they attract the best talent to support their business.”
Authors’ Note: We would like to thank Candice Hévin for research assistance.
 Due to the limited scope of this article, we have not addressed automated document creation (i.e., technology-assisted drafting), which is currently less dependent on AI technologies and more based on questionnaire input and populating fields in precedents.
 ROSS Intelligence was founded in Toronto, but its functionality currently only extends to U.S. legislation and case law.