When legal luminary Roscoe Pound wrote, “The law must be stable, but it must not stand still,” little did he know the seismic shifts awaiting the industry. Beyond the growing body of AI laws, how lawyers approach their day-to-day work is likely to look far different in years to come.

Generative AI and large language models (LLMs) have the potential to redefine efficiency and service in the legal industry. Every aspect of the work — research, discovery, document creation, review, and management processes — can be streamlined and improved using these models. And those who act first can expect to increase their competitive edge over those who take a wait-and-see approach.

However, we’ve seen several high-profile stories of GenAI gone wrong in legal and other settings. From inaccurate and fictitious case information (hallucinations) provided by general-purpose LLMs, to open-source LLMs inadvertently disclosing confidential information, these mishaps are giving pause to firms who are building generative AI solutions. The threat of lawsuits alleging intellectual property rights violations by LLM providers only adds to the understandable anxiety.

In an industry that constantly assesses risk — to both clients and firms — the question is: How do you take advantage of the massive opportunity that this technology brings while avoiding a gaffe that makes the front pages?

The Growing Opportunity for Legal Tech Providers and Law Firms

Generative AI is significantly advancing modern legal technology, and we see two primary groups using it to build new solutions. The first, unsurprisingly, is the legal tech market, which is estimated to exceed $45 billion by 2030. Legal tech firms can give solo lawyers and small firms the technological leverage they need to compete against their larger counterparts. The second user group is large legal firms. They have the resources needed to deploy generative AI solutions themselves.

So how are these groups using LLMs today?

Legal tech providers are using generative AI to improve the accuracy of existing document management, legal research, and contract analysis systems, and to enable net-new capabilities for large and small firms alike.

One Employer of Record and now legal tech startup, Borderless AI, built an employment law knowledge assistant with Cohere models. The tool, called Alberni, is the first to enable lawyers to rapidly get answers regarding local labor laws and regulations worldwide. Alberni can also quickly analyze existing employment contracts, and draft new contracts in multiple languages.

Large, global legal firms are quickly embracing LLMs and generative AI, with nearly half of top U.S. law firms and a similar number in the UK reporting that they use the technology. One significant LLM use case is streamlining time-consuming tasks, such as summarizing or reviewing lengthy documents. This frees up lawyers for more strategic work, so they can deliver greater value to their clients.

Lawyers at one large law firm, for instance, can more quickly outline their litigation strategies and legal arguments by using AI to summarize witness statements. Another firm uses AI to summarize insights from government hearings and create relevant alerts on new policies, regulations, and risks for their client. At a third firm, partners are deploying the technology to assist with “horizon scanning” — proactively identifying bills, regulatory changes, news, and other developments relevant to their clients. See table below for more top uses cases.

Example use cases for generative AI in the legal industry.

Ways to Ensure Accuracy and Privacy

As impressive as the technology can be, it’s best to think of LLM applications as tools that can support workers in a variety of ways across the org chart – from a legal intern learning the ropes to a more seasoned advisor. When it comes to choosing an LLM consider how the application will be used and what levels of performance and security will be necessary to meet your requirements to deliver accurate outputs and protect client privacy.

Interviewing LLM Providers

Once you have a use case in mind, you’ll want to interview LLM providers to align your requirements with the right technology and supporting infrastructure. Here are three important questions to ask:

1. What deployment methods do you support? How you deploy an LLM-based application will affect privacy and security controls. Most large global law firms and legal tech companies will likely need to use a combination of approaches to match different client requirements and use cases. These include using private clouds and managed cloud services.

A virtual private cloud provides a firm with its own space to deploy and manage a dedicated instance of the LLM. Neither the cloud provider nor LLM vendor can view firm queries or data, making it ideal when clients don’t allow their data to be processed through a cloud provider. With virtual private clouds, a firm's IT and AI experts control the systems, access controls, and data security features.
Managed cloud services require less technical support and oversight by the firm. It’s possible in cases where client data isn't used, such as when summarizing public documents, or where clients have less stringent data requirements.

Cohere, for instance, provides deployment options across all major cloud vendors that can support either of these approaches, or a combination, based on a company’s need.

2. What strategic and technical capabilities do you provide either in-house or through partners? The technology is rapidly evolving, and few firms — even legal tech providers — will have the talent to keep up. It is a situation where it will “take a village” to stay current on the latest technical advances and how those may impact privacy, accuracy, and IP risks. What resources can an LLM provider deliver to help you customize your legal solution and stay abreast of changes in this rapidly evolving landscape?

Consider the LLM provider's internal expertise, as well as its partnerships, to help you successfully and safely move from a proof of concept to production. To support our customers through this journey, we recently added Accenture to our partner roster, which also includes McKinsey and Oracle, as well as many others. For instance, our partnerships with vector database providers, including Elasticsearch, Weaviate, Pinecone, Quadrant, Milvus, and Zilliz, can enable organizations to implement advanced RAG capabilities that improve the accuracy of information retrieval. Other technology partnerships, such as our integration with Langchain tools, simplify the process of deploying complex, LLM-powered business workflows.

3. What protections do you provide regarding intellectual property rights? Beyond privacy and accuracy, firms should also examine how LLM providers indemnify organizations in the event of intellectual property (IP) claims. While many LLM providers offer some level of indemnification, providers often take different approaches. Learn more about Cohere’s IP legal protections for its customers.

Guiding and Training Your LLM

Firms and legal tech providers will need to provide LLMs guidance and training to gain the level of accuracy they require in model outputs. You can use a number of techniques to provide LLMs this training, including:

1. Giving your LLM access to authoritative data sources in real time. One popular natural language processing technique, called retrieval-augmented generation (RAG), enables LLMs to access information from authoritative data sources beyond their initial training data in real time. This infusion of context-specific data ensures the models can provide more accurate and timely answers to questions while reducing hallucinations. These data sources can include public records — for example, legal databases and libraries, government records, and business databases — and internal systems, such as document and contract management systems. Borderless AI, for example, uses Cohere Command R with RAG to enable its knowledge assistant to search its repository of legal documents on immigration laws, work permits, tax regulations, and employment contracts, so it can effectively draft contacts, analyze legal agreements, and answer questions in seconds.

2. Preparing your LLM for better retrieval. One way to do this is simply by using a reranking model with existing legacy search systems. These models can reorder AI query-response pairs to deliver significantly more relevant responses. Cohere Rerank, for example, can improve the relevancy of documents retrieved by up to 50% in some cases.

Another way to improve search query understanding is by implementing semantic search capabilities in place of keyword search. For instance, a semantic search on “contract termination” would look beyond the exact words in the phrase and provide a broader range of relevant documents related to cases where individuals didn't fulfill contractual terms. This enables a lawyer to have a more comprehensive view of the issue.

By using Cohere Embed to turn unstructured text into numerical vectors (called embeddings), Borderless AI's tool, for example, can semantically assess the relevance of documents to deliver contextually relevant results to its users. Using Embed's multilingual capabilities, users can also search and receive results in over 100+ languages. Similarly, a large law firm could use Cohere Embed to make e-discovery faster and more efficient, surfacing information from the large corpus of documents it must retain for each case. Such capabilities are crucial for enabling global legal practices to share information and access relevant documents across any of their practices.

3. Educating your LLM on legal jargon. General-purpose LLMs that can answer questions on a broad range of topics often need customization (also known as fine-tuning) to gain in-depth industry language and domain-specific knowledge. This fine-tuning might include teaching it legal terminology, case law, statutes, legal reasoning, or internal firm knowledge specific to a given area of practice.

A legal firm, for instance, may train an LLM on internal training materials, frequently asked client questions, and internal case files to bring together insights from colleagues anywhere in the world.

One firm plans to train Cohere Command R on examples of contract clauses that might be more or less favorable to its clients. This will enable the firm to develop an AI solution that can flag those areas of a lengthy contract that its lawyers may consider problematic.

Remember, when creating training data sets, firms must identify relevant and representative data for a given task or function, train the model, and then evaluate the outputs.

While many firms may still be skeptical about LLMs and generative AI, shying away from the technology is not the answer. The productivity gains are enormous, and those who lead with AI are likely to gain an edge. Instead, to build confidence, organizations should start testing the models. Developing a deep understanding of any potential privacy, accuracy, and IP concerns (and solutions) for legal use cases requires hands-on experience. Take inspiration from the law firms and legal tech providers highlighted in this blog post, and get started now.

To get started,

Let's talk

Join our exclusive webinar on Tuesday, April 30, 2024, 2:00 PM ET / 11:00 AM PT, to discover how Borderless AI leveraged Cohere models to build a cutting-edge, RAG-based legal assistant. Secure your spot today.

Kosta Starostin