Thomson Reuters Shifts Legal AI Work to Its Model

Thomson Reuters says it is moving token-heavy legal tasks, including bulk document review and tabular analysis, from frontier models to its own Thomson model in CoCounsel Legal. CEO Stephen Hasker says the move has increased speed and accuracy and may lower costs; the company plans to use that cost advantage over the next year or so.

Hasker says the Thomson model can handle these tasks more cheaply when its capacity utilization approaches 100%. That could let the company keep all-you-can-eat pricing for CoCounsel Legal in many cases, an offer he says is compelling to private partnerships. Thomson Reuters also offers enterprise-wide pricing or per-token usage pricing, and Hasker says those arrangements have maintained an attractive margin profile. The company has not yet explored higher prices for power users, though Hasker says it is exploring pricing based on value. He called the product's development an early stage in a five-to-ten-year journey.

Thomson Reuters is pitching CoCounsel as a legal AI assistant for work beyond legal research and know-how. Hasker describes that market as a “white space” and estimates a five-to-ten-year expansion in total addressable market. The rebuilt CoCounsel Legal has native access to Westlaw and Practical Law, and Thomson Reuters says 2,700 practice experts trained and built the agent. Hasker said users in a comparison exercise ranked CoCounsel's answer to an in-depth legal question above answers from five or six other applications, including frontier models and startup products.

In September, its first full month after general availability, CoCounsel NextGen surpassed the previous version's best month on every engagement and retention measure Thomson Reuters tracks. It carried the majority of users in the CoCounsel Legal AI bundle, and its active user base rose more than 60% from August. NextGen users completed 2.4x the work compared with the previous version's peak, were active 44% more days per month, and had higher retention. Product NPS scores were more than twice the previous version's best month on record.

Hasker says CoCounsel shows the steps it takes to answer a question—10, 20 or 30 steps—and identifies the authoritative citations or source materials used at each step, explaining why it moved from one step to another. He says customers see this transparency as a potential training and learning tool for young attorneys, amid concerns among law firm leaders about how firms would train new lawyers if less work were done through traditional apprenticeship. The article closes by saying customers will judge the company's case against rising AI costs in the coming months.

Thomson Reuters says it is moving token-heavy legal tasks, including bulk document review and tabular analysis, from frontier models to its own Thomson model in CoCounsel Legal. CEO Stephen Hasker says the move has increased speed and accuracy and may lower costs; the company plans to use that cost advantage over the next year or so.

Hasker says the Thomson model can handle these tasks more cheaply when its capacity utilization approaches 100%. That could let the company keep all-you-can-eat pricing for CoCounsel Legal in many cases, an offer he says is compelling to private partnerships. Thomson Reuters also offers enterprise-wide pricing or per-token usage pricing, and Hasker says those arrangements have maintained an attractive margin profile. The company has not yet explored higher prices for power users, though Hasker says it is exploring pricing based on value. He called the product's development an early stage in a five-to-ten-year journey.

Thomson Reuters is pitching CoCounsel as a legal AI assistant for work beyond legal research and know-how. Hasker describes that market as a “white space” and estimates a five-to-ten-year expansion in total addressable market. The rebuilt CoCounsel Legal has native access to Westlaw and Practical Law, and Thomson Reuters says 2,700 practice experts trained and built the agent. Hasker said users in a comparison exercise ranked CoCounsel's answer to an in-depth legal question above answers from five or six other applications, including frontier models and startup products.

In September, its first full month after general availability, CoCounsel NextGen surpassed the previous version's best month on every engagement and retention measure Thomson Reuters tracks. It carried the majority of users in the CoCounsel Legal AI bundle, and its active user base rose more than 60% from August. NextGen users completed 2.4x the work compared with the previous version's peak, were active 44% more days per month, and had higher retention. Product NPS scores were more than twice the previous version's best month on record.

Hasker says CoCounsel shows the steps it takes to answer a question—10, 20 or 30 steps—and identifies the authoritative citations or source materials used at each step, explaining why it moved from one step to another. He says customers see this transparency as a potential training and learning tool for young attorneys, amid concerns among law firm leaders about how firms would train new lawyers if less work were done through traditional apprenticeship. The article closes by saying customers will judge the company's case against rising AI costs in the coming months.