On August 22, Spellbook released a new product it is calling Spellbook Associate, that has been touted in press releases as “the first legal AI Agent”.
The product, which the company says can plan, execute, check its work, and adapt to accomplish larger scope assignments like a colleague would, is focused on executing transactional tasks, similar to the company’s original AI Assistant tool. Spellbook Associate can use a single prompt to work through legal matters such as producing complete financing documents from a term sheet, reviewing hundreds of documents for risks and inconsistencies and revising employment packages, doing so significantly faster than a human associate.
The product sounds compelling, especially if, like Spellbook’s other solution, it is primarily focused on mid-law. If that’s the case, then it may well be the first legal agentic solution squarely targeting that demographic. Also, as Spellbook founder Scott Stevenson says, their Associate product is able to execute across a range of tasks, revising and cross-referencing documents and undertaking work across multiple applications (including Microsoft Word), adjusting and replanning as it works. Some of the other implementations of agentic AI in the industry are focused in on single tasks.
But, to be clear: this is not the first legal AI agent. Spellbook Associate is the most recent legal AI agent to launch, marking the continuation of a trend we’ve been covering for over a year, namely the deployment of agentic AI to facilitate legal tasks.
We have written about AI agents multiple times. See previous articles here and here. We highlighted legal AI agents as a hot area for the year in our forward look at legaltech in 2024, and in our recent 2024 mid-year check-in.
Though definitions of AI agents may vary, most good definitions share common characteristics. An AI agent is an LLM-based entity that can manipulate its environment in some way (e.g., a computer, web browser, other agents, etc.) via tools, in order to achieve a multi-step goal that it is set (e.g., book a flight, write a business plan, research some topic, etc.) with significant autonomy over a long-time horizon. In other words, with minimal human input, an agent breaks down a broad goal into small sub-tasks and decides when that goal is satisfied, after checking its own work. Agents are usually able to course-correct, developing a plan B if the first plan did not successfully achieve the goal it has been set (and then a plan c and plan d, until the goal is achieved). Depending on how a product is engineered, the agent can reveal its plan to the user before it starts working, or the planning stage can be hidden from view. Like any product, simplicity often belies underlying complexity, so that agentic products that perform well with a seamless UI may appear like simpler tools than they in fact are. Agents and agentic AI have been in use since before ChatGPT, and have become more powerful with the improvements in LLM technology over the past year.
As we have been reporting since fall 2023, AI agents have particular applicability in the legal industry because lawyers rarely undertake a task in isolation. Instead, their work occurs in flows of related tasks, designed to be undertaken step by step in order to achieve an overarching goal. Agents are a better fit for this type of work than LLM chatbots alone, which many companies working to empower lawyers through technology have realized.
Given the recent launch of Spellbook Associate, this seems as good a time as any to take a look at just some of the products leveraging agentic AI that have launched in legal since last year.
273 Ventures – Dan Katz and Mike Bommarito started building legal AI agents in 2023. Six months ago, they showcased their six months of work building the Kelvin Agent, announcing that they had invested heavily into expanding its capabilities, “building out nearly 100 tools and over 40 API methods for "headless" interaction.” The Kelvin Agent (which is proprietary) was in alpha as early as August 2023, but has since been built out to address multiple complex scenarios, for example to quickly answer questions about 100+ page syndicated loans or dozens of docs in a deal rooms, showing its step-by-step work along the way. See a demo video here.
Harvey – Many people may not realize that Harvey is more than a single LLM and more than a chatbot. In fact, most of Harvey’s AI systems are agentic in nature. They all create a multi-step plan based on user queries and related uploaded documents. For each step of this plan, they produce output that is checked against the plan for accuracy, and the plan is adjusted accordingly. Once the plan has been carried out, Harvey algorithms check the work produced against the query and reference each step taken and the sources so users can verify that the plan was executed correctly. Harvey Workflows, in particular, leverage agentic AI to orchestrate hundreds of highly specialized models and complete full workflows beyond the scope of a single LLM, similar to how lawyers work together on complex matters. For example, users can produce a company profile, which is a long document that describes a company based on all public information. This workflow is powered by an agent that creates a plan based on the company and conducts research for each section of the profile to find, verify, and compose the necessary information. Other workflows like due diligence, redlining, and drafting are also supported by agentic workflows. As per its website, Harvey is building workflows for “all professional services use cases”.
Even prior to workflows, Harvey’s assistant product has used agentic algorithms since Harvey first gained early access to GPT-4 in August 2022. Harvey’s assistant algorithm creates a multi-step plan based on a user query and documents, and leverages multiple custom language models and other techniques to complete complex user queries, verify their correctness, and adjust its plan dynamically.
Harvey’s research team is deeply experienced in agent systems. Gabe Pereyra, Harvey’s cofounder, was a research scientist at Google DeepMind—the research lab famous for popularizing reinforcement learning and agents through work like AlphaGo—and at Meta AI, where he also worked on LLM agents. Niko Grupen, Harvey’s head of applied research, has a PhD from Cornell in multi-agent systems and was also a research scientist at Google DeepMind.
BRYTER – in July 2024, BRYTER announced the launch of its AI Agents product. Using proprietary, pre-trained AI, BRYTER’s AI Agents help law firms and legal departments with the manual, tedious elements of their work: from speeding up the review of contracts using the Review Agent, to getting draft email replies to recurring requests from commercial teams directly in MS Outlook or Gmail through the Email Agent. BRYTER agents work by planning work through breaking down a bigger goal into smaller tasks that are undertaken step by step. Agents are able to revise and reference documents or passages of documents and work across several applications, like MS Teams, Outlook, Excel. The BRTYER homepage lists six discrete agents built especially for legal, as well as the capability for users to build their own legal-specific agents. Also, this just in: a new agent, built specifically for higher volume contract review, is about to launch (stay tuned!).
Dioptra – Built with the innovation team at Wilson Sonsini, Y Combinator start-up Dioptra is an AI Agent that reviews contracts with the accuracy of lawyers, be it for negotiation, due diligence or compliance use cases. Former Chief Innovation Officer at Wilson Sonsini David Wang shared that "For an AI Agent in the legal space, quality is the most important factor to gain true efficiency. Dioptra was consistently 92%+ accurate on first-party reviews, third-party reviews, and issue detections we tested." With backgrounds in AI research and academia and more than 10 years each of hands-on experience in AI, the founding team says they are committed to bringing the most reliable AI Agent to the legal industry. Dioptra launched in May 2024.
Leya – Swedish AI company Leya has been running agentic flows since September last year. Leya’s agent can select from a range of specialized tools, depending on the use case. The team says that as most agentic workflows are not yet mature enough to run truly end-to-end tasks autonomously, Leya keeps the lawyer in the loop to make sure each solution is optimized for accuracy and validity in the results.
Hebbia – Originally targeting the financial sector, Hebbia turned its focus towards the legal industry with the hire of Ryan Samii as Head of Legal, leveraging its AI agents to to process, organize and analyze complex data precisely as a user instructs. Hebbia provides organizations with an LLM-linked agentic copilot called Matrix, which allows knowledge workers to ask complex questions associated with internal documents — from PDFs, spreadsheets and Word documents to audio transcripts — with an infinite context window. Once a user provides the query and the associated documents/files, Matrix takes the prompt and breaks it down into smaller actions that the LLM sitting under the hood can execute. This enables it to analyze all the information contained in the documents at once and extract exactly what’s needed in a structured form. Hebbia says the platform enables the model to reason over any amount (millions to billions of docs) and modality of data while also providing relevant citations to help users trace every action and understand how exactly the platform got to the final answer.
Centari: Centari, a legal deal intelligence platform, has developed AI agents that identify and interpret information from complex deal documents. Just as lawyers often apply multi-step reasoning to derive an answer from cross-references and multiple documents, Centari’s agents are trained to perform a similar workflow to achieve a high degree of data extraction accuracy.
Lawdify – Not a company I was familiar with until recently, Lawdify (based in Singapore) is building agents that run specialized tasks for lawyers, including a Legal Due Diligence AI Agent product and an eDiscovery AI Agent product. Lawdify’s website boasts that its agents have the ability for contextual understanding, reasoned analysis, autonomous execution, and connect with external data sources.
This list is by no means exhaustive. In many of the sophisticated products using advanced AI for legal work, there is a likelihood that agentic AI plays a part under the hood. Whether that means those products constitute “fully fledged legal AI agents” is a question up for debate.
