Thomson Reuters previewed the latest version of its CoCounsel AI agent, which was touted as a fully autonomous assistant that can understand complex tasks, do its own research, and deliver a work product that was described as "human-level." Presented as a tool for legal-related tasks, its use cases also encompass tax law and regulation, as well as other applications such as due diligence, regulatory compliance and contract review.
"This architecture really supports any legal task," said Joel Hron, chief technology officer for Thomson Reuters, during a webcast. "This agent is able to develop a plan and execute tools to go solve that plan. We're not building narrow point solutions with this version of CoCounsel. We're really building an AI platform, a legal AI platform, that can handle the full spectrum of legal work."
While previous versions of CoCounsel already provide insights, draft content and autonomously perform tasks (Thomson Reuters also demoed the current version handling sales and use tax workflows for a hypothetical e-commerce retailer), Hron said they did not expose the company's content in a way that would truly bring its advantages to the fore. The latest version, he said, represents a fundamental change in that it is not going to be just answering questions or generating first-pass drafts but operating in a way that will allow it to deliver "human-level work product."
To demonstrate, he opened the latest version's interface, which features a chat window where people can enter their requests in plain language. He raised an example where someone wanted to analyze the recent antitrust case against Google, particularly the different proposals between Google and the U.S. government. He uploaded both proposal documents into the platform, each over 100 pages long, then typed a prompt asking the AI to draft a five-page summary on the key differences between the legal remedies proposed by Google and the government, and added a paragraph on the legal standard the court will need to apply when deciding between them, backed by legal research. By doing this, he said, he's asking the AI to do a few things in turn: to look at the documents, to draft a summary of the differences, and then perform research on precedents that would support the decision.
CoCounsel first searched the documents for key points, then summarized its findings, then used a deep research tool on the Westlaw database for more substantive information, went back to the document to check against what it found in the database, then produced the five-page summary and even included, on its own, a separate document on antitrust remedy standards from the information in the Westlaw database. He noted that the AI is not following a script or plan or predefined workflow but instead making dozens of independent decisions at each step based on the guidance of Thomson Reuters' own knowledge sources that has been infused into the product throughout the development process. He noted that this is a process that would have taken a human hours, or even days, to do, but with CoCounsel, was completed in minutes.
"This deliverable at the end of the day isn't just a good starting point; it's a complete, structured, human-level work product with supporting research documentation at a level that you would expect an associate to be at," said Hron. "The lawyer isn't out there cleaning up the work. They're reviewing and they're validating the output and maybe following further breadcrumb trails of work to go do from here. So you can see exactly what CoCounsel did, why it made those choices. You can trace back through the research. This is not a black box. It is meant to be a collaborator, a human collaborator."
He stressed that this is not a large language model spitting out legal-sounding words, nor did Thomson Reuters just take a foundation model and then wrap domain-oriented workflow experiences around it. Further, when the AI is conducting a search, it is not performing a simple web search but actively querying the same legal databases that any professional lawyer would be using based on subject matter expertise built directly into the product's structure, giving it the ability to understand and apply authoritative legal content.
"When it structures a document, it's applying editorial standards that have been refined over decades by practical law. And when it cites authority, it knows exactly what is authoritative because of editorial enhancements that have been made to our content. And this integration isn't something you just build with a series of API calls," said Hron.
While there are AI solutions that can do these things individually, he said CoCounsel's orchestration layer, which coordinates activities between multiple AI models, does them together as a holistic system.
"We're not building narrow point solutions with this version of CoCounsel," he said. "We're really building an AI platform that can handle the full spectrum of legal work. The question in legal AI isn't whether AI can generate text that sounds legal; the question is really around trust, and it's whether AI can really think legally and decompose complex problems, conduct rigorous research, apply authoritative sources and deliver work product that meets professional standards and, more importantly, earns the trust of those professionals because of the accountability of the system."
The beta for the new version is expected to be available early in the second quarter.
