How do you build AI lawyers can trust? Inside Thomson Reuters' AI strategy!

Hey everyone, it's Shii! Today's story is a bit different: an interview about how Thomson Reuters builds AI for work where mistakes really aren't an option.

Claude Blog claude.com

What was announced?

The Claude Blog published an interview about how Thomson Reuters, a company with over 175 years of history, builds AI for high-stakes professional work in law, tax, accounting, and compliance. Joel Hron, CTO of Thomson Reuters, walks through how the company thinks about AI development.

Why it matters

Alongside legal research tools like Westlaw and Practical Law, Thomson Reuters offers CoCounsel Legal, a professional-grade legal AI platform built to boost lawyer productivity. As Hron puts it, "We're a technology company focused on professions that demand accuracy and precision." The company's bar for choosing a model is whether its output can survive the kind of professional review lawyers apply before relying on it.

Thomson Reuters calls this approach Fiduciary-Grade AI™: "AI grounded in authoritative content, shaped by deep domain expertise, and embedded directly into professional workflows, so outputs are transparent, verifiable, and defensible." The company rebuilt legal research around agents dedicated to citation validation, and customers report that research that "would take dozens of hours" now arrives "in a matter of minutes."

What changes

Rather than build a smarter chatbot, Thomson Reuters rebuilt its products around agents that can reach hundreds of internal tools at once. CoCounsel Legal was rebuilt on the Claude Agent SDK, letting it plan, delegate, and orchestrate across tools and content sources in real time. Hron notes that Thomson Reuters was one of Anthropic's earliest enterprise customers, saying "the number one thing that spoke to us was Anthropic's approach to building enterprise AI" - citing transparency, safety, and responsible development.

Hron's team distilled what knowledge work demands from a model into four capabilities:

• Citation validation: models need to check their own citations rather than just retrieving sources, validating findings before human review
• Extended tool use: staying focused across long chains of tool calls and many systems without losing context
• Human-in-the-loop design: bringing humans into the process of building a work product, instead of just returning a one-shot answer
• Expanding what's possible: tackling work once considered too demanding. With Claude Fable 5, advanced drafting for complex legal work, including motions and filings that professionals "would otherwise spend days or weeks perfecting," is now within reach

Dive Deep

Hron's take on AI's ROI is refreshingly unconventional. "If you try to optimize too much for the rate of return calculation, you miss the forest for the trees," he says, encouraging teams to experience the cultural and mindset shift before chasing cost-per-task optimization.

That said, the company still tracks concrete numbers. Alongside standard engineering measures like DORA metrics and time to production, an internal error-remediation tool built on Claude cut production issue analysis from three hours down to four minutes. "The ability to get back to health within minutes versus hours is a material difference," Hron says.

He also emphasizes that AI is changing the nature of the work itself. "The act of writing lines of code is no longer the job," he says of his engineers, adding that systems thinking, judgment, and taste now matter most. AI is helping people become more "T-shaped," able to work across product, design, and finance.

Hron uses these tools himself: Claude Code helps him quickly get up to speed on codebases he hasn't touched in months, while Claude Cowork lets him borrow a CFO's or a strategist's perspective to pressure-test ideas. For work that ultimately has to hold up in court, he sees this frontier as one of the next critical directions for AI.

Wrap-up

• Thomson Reuters builds AI for high-stakes professional work in law, tax, accounting, and more
• CoCounsel Legal was rebuilt on the Claude Agent SDK as an agent that reaches across many tools
• The model requirements Hron's team cares about: citation validation, extended tool use, human-in-the-loop design, and expanding what's possible
• An internal error-remediation tool cut issue analysis time from three hours to four minutes
• The company is looking to Claude Fable 5 for longer, more demanding professional work

If you're curious how AI gets built for a field where mistakes truly aren't an option, this one's worth a proper read!