Our Take
Thomson Reuters is betting that owning both the content layer (Westlaw, Practical Law, 2,700 attorney-editors) and now a legal-specific model gives it an edge competitors cannot replicate, but 'fiduciary-grade' is marketing language until independent audits verify the hallucination rate.
Why it matters
Law firms are drowning in AI tools but few have operationalized them at workflow scale. Thomson Reuters is claiming to solve the accuracy and transparency problem that has embarrassed major firms in public hallucination cases, which matters if true and damages the company if it isn't.
Do this week
General Counsel: Run a live brief or memo through both CoCounsel and a competitor (Harvey, Legora) in parallel on the same task before committing to annual volume, so you can measure hallucination rate and citation accuracy on your own docket.
Thomson Reuters Opens Early Access to Rebuilt CoCounsel
Thomson Reuters announced early access to the next generation of CoCounsel Legal, the most substantial revision since the company acquired Casetext in 2023. The product now integrates Westlaw and Practical Law content directly into an agentic AI system trained by the company's 2,700 attorney-editors. CEO Steve Hasker told LawSites the acceleration was driven by beta feedback from the past month, claiming this is "the most successful beta trial I've ever been a part of" in terms of utility, accuracy, and reliability.
The rebuilt product runs on Anthropic's Claude SDK but is no longer locked to it. Thomson Reuters has also completed internal testing of Thomson, its own large language model for legal work trained using a research lab acquired from Cambridge and Imperial College. Hasker said the Thomson model is "starting to outperform all of the latest models for specific legal tasks" but the company has not yet decided whether to deploy it in CoCounsel or run different components on different models.
The Real Differentiation Is Content Access, Not the Model
Hasker positioned CoCounsel as the only legal AI meeting "fiduciary-grade" standards, defined by three things: native access to authoritative legal content, in-house domain expertise to train the agent, and transparency that prevents hallucination.
The claim rests on observed hallucination failures at rival firms. Hasker cited three recent public cases in which well-regarded law firms using frontier models without native content access produced briefs with false citations or opinions. In at least two cases, opposing counsel using CoCounsel identified the errors and alerted the judge. Competitors Harvey and Legora, he noted, lack access to Westlaw and Practical Law and do not employ legal editors to tune their agents.
This is verifiable but narrow. Thomson Reuters has a genuine structural advantage: its own content and editorial workforce. Whether that translates to a measurably lower hallucination rate or higher task success on independent benchmarks is unstated. The company provided no numerical comparison (false citation rate, accuracy on precedent lookup, etc.) and no independent audit of either product. "Fiduciary-grade" remains a label, not a published standard.
Hasker also emphasized transparency as a differentiator: CoCounsel shows its reasoning steps and cites sources for each conclusion. He framed this as benefiting both confidence and legal education, since young attorneys can understand how the tool reached its answer. This is a design choice, not a performance claim, and is harder to assess without using the product.
Test on Your Own Work Before Rolling Out
Law firms and legal departments should resist the narrative that tool choice is settled. The Thomson Reuters positioning is credible on content assets but unvalidated on outcomes. Run the same complex task (a motion, an opinion, a contract review) through CoCounsel and at least one competitor on real docket work before expanding seat count or signing multi-year commitments. Measure false citations, missed precedent, and time-to-usable-output. Hasker acknowledged that the real bottleneck is change management, not tool capability, which means the best product is the one your team will actually integrate into existing workflows.