Working at the frontier: How Thomson Reuters builds AI for high-stakes professional work
02:00 · July 8, 2026 · Claude Blog
Summary
Thomson Reuters has integrated Anthropic’s Claude models, including the recent Claude Fable 5 release, into CoCounsel Legal, its platform for high-stakes legal work. The company pairs these frontier models with its own curated legal content, domain annotations from more than 2,700 experts, and workflow-specific evaluations to produce outputs that lawyers can stand behind. This combination, branded internally as Fiduciary-Grade AI, prioritizes citation validation and source transparency over raw fluency, ensuring every result remains verifiable before a professional signs off.
The architecture has moved from sequential, standalone tools to an agent-first design built on the Claude Agent SDK. A single agent now plans, delegates, and orchestrates across hundreds of internal tools and content sources in real time. Evaluation criteria have shifted accordingly: models are tested on their ability to maintain coherent tool use over long sequences, to surface and confirm citations before presenting findings, and to insert human review points rather than attempt one-shot answers on complex tasks. These capabilities have reduced research that once required dozens of hours to a matter of minutes while preserving professional accountability.
Engineering teams report both productivity gains and a broader change in working practice. Traditional metrics such as time-to-resolution and DORA indicators show clear improvement—one internal remediation tool cut root-cause analysis from three hours to four minutes—yet leadership emphasizes cultural effects first. Engineers spend less time writing individual lines of code and more time on systems thinking, judgment, and cross-functional coordination, becoming more T-shaped in their skill profiles. The same pattern is appearing in product, design, and finance roles.
Looking ahead, Thomson Reuters is extending these agents toward longer-horizon drafting tasks, such as motion preparation, that demand sustained context and precision previously out of reach. The focus remains on environments where near-correct results are insufficient and every output must withstand formal scrutiny.
Why it matters
This article provides valuable architectural insights for product teams building high-stakes enterprise AI agents. Its focus on verifiable outputs, data privacy, and human-in-the-loop design strongly aligns with the ethical and transparent AI standards prevalent in the Dutch and EU markets.
