Thomson Reuters Launches Proprietary AI for Legal Industry

Thomson Reuters has unveiled its first proprietary large‑language AI, a move that could reshape how the company delivers legal, tax and professional content. The Thomson engine is built on the firm’s extensive research libraries and is billed as a “fiduciary‑grade” tool for practitioners who need verifiable answers.

Company seeks to replace third‑party providers

The new system is intended to reduce reliance on external foundation models such as those from OpenAI or Anthropic. By embedding its own intelligence, the firm hopes to avoid price hikes or policy shifts that could affect its suite of subscription services.

According to the filing, the development cost was roughly $40 million, a fraction of the multibillion‑dollar budgets typical for large AI projects. Training leveraged an open‑source base before being fine‑tuned with decades of curated material from Westlaw, Practical Law, Checkpoint and Reuters news archives.

Only a small slice of that proprietary corpus—under ten percent—has been used so far, suggesting the company may still have a large reservoir of data to enrich the engine later.

Related: Florida Guide to Establishing Paternity Step by Step

Benchmarks show competitive performance

The firm’s internal tests placed the AI ahead of several leading frontier systems on a legal‑specific benchmark, scoring 0.352 on the PrBench Legal Hard test. It also posted favorable results against Claude Opus 4.8, GPT‑5.5 and Gemini 3.1 Pro on assorted professional tasks.

Those numbers come with the usual caveats: most of the evaluation was internal, and real‑world usage can differ from controlled leaderboard conditions.

Legal scholars have begun independent assessments. Professor Jonathan Choi of Washington University compared the tool to ChatGPT and Claude on complex corporate‑tax queries and said the answers from the new engine were generally preferred because of their linked citations. Professor Samuel Dahan, affiliated with Queen’s University and Cornell’s Legal AI Lab, noted that citation quality was on par with top‑tier models.

CoCounsel will continue to operate as a multi‑model service, routing certain queries to the proprietary engine while still calling on external models for tasks where they have an edge. This hybrid approach reflects a pragmatic view that no single system can dominate every legal workflow.

From an economic standpoint, owning the intelligence layer gives the firm more control over pricing and product roadmap. Competitors that rent external models may face higher costs if providers decide to enter the legal market directly—a scenario already seen with Anthropic’s push into law‑firm tools and Google’s Gemini Enterprise expansion.

Related: Burgundy Reads Everything You Type

Industry observers have highlighted the strategic value of the firm’s content moat. While large AI labs possess massive compute resources, they lack the curated, cross‑referenced legal material that underpins the new engine’s answers.

The underlying thesis is that a specialized system, trained on high‑quality professional data, can outperform a generic model that merely ingests the same documents. Early testing appears to support that view, though broader adoption will be the true test.

Looking ahead, the firm may expand the engine’s reach beyond legal services into tax and accounting solutions, leveraging the same data‑driven methodology. If the approach scales, it could set a precedent for other information‑rich industries to develop their own AI capabilities rather than depend on external providers.

For now, the rollout remains limited to a specific feature within CoCounsel, and the company has opened its sandbox to additional academic partners for further independent evaluation.