Jonathon's AI Wiki

Working at the Frontier: How Thomson Reuters Builds AI for High-Stakes Professional Work

9 min read

Source: ai-research/claude-blog-thomson-reuters-fiduciary-grade-ai.md — Anthropic first-party “Working at the Frontier” customer case study, claude.com/blog/working-at-the-frontier-how-thomson-reuters-builds-ai-for-high--stakes-professional-work (note the literal double-hyphen in the URL slug; fetched 2026-07-16, exact publish date not captured in extraction). Byline source: Joel Hron, CTO of Thomson Reuters. Part of the same case-study rollout as the sibling posts on Hebbia and Cognition.

Thomson Reuters — a 175-plus-year-old content and technology company behind Westlaw, Practical Law, and CoCounsel Legal — evaluates every new model by whether its output can withstand the level of professional review a lawyer applies before relying on it. This article covers Thomson Reuters’ “Fiduciary-Grade AI™” framing, the four requirements CTO Joel Hron’s team holds every model to, the CoCounsel Legal rebuild onto the Claude Agent SDK, and a concrete ROI data point (a production incident’s root-cause-to-fix time dropping from three hours to four minutes).

Key Takeaways

  • “Fiduciary-Grade AI™” is Thomson Reuters’ branded framing for professional AI: grounded in authoritative content, shaped by deep domain expertise, and embedded directly into professional workflows, so outputs are transparent, verifiable, and defensible when the stakes are high. The 2,700+ domain experts who annotate and enhance Westlaw’s case law daily are named as load-bearing to why a lawyer can trust an answer — not the model alone.
  • CoCounsel Legal was rebuilt on the Claude Agent SDK. It used to run separate skills sequentially; it now plans, delegates, and orchestrates across hundreds of tools and content sources in real time, letting a professional define the outcome instead of dictating every step. Customer data remains protected and is not used to train third-party models.
  • Four things a model must do before Thomson Reuters trusts it: (1) check its own citations before presenting findings for human review, (2) hold steady across long chains of tool calls, (3) bring a human into the loop rather than one-shot an answer, (4) free up time for previously out-of-reach work like motion drafting.
  • A concrete ROI data point: an internal Claude-based error-remediation tool turned a production incident from three hours of root-cause analysis into a four-minute fix. Hron’s broader ROI philosophy is contrarian — optimize for the cultural/mindset shift first, and “the returns follow on their own” rather than over-indexing on a rate-of-return calculation up front.
  • Thomson Reuters was one of Anthropic’s earliest enterprise customers, and Hron says the deciding factor wasn’t a benchmark — it was Anthropic’s approach to building enterprise AI (transparency, safety, responsible AI development). The first joint proof point was deep research in legal.

“We’re a technology company focused on professions that demand accuracy and precision,” Hron says. Thomson Reuters brings three advantages to professional AI that general-purpose systems can’t easily replicate: authoritative content, deep domain expertise, and workflow integration. A lawyer can rely on a Westlaw answer not because of the model alone, Hron argues, but because of decades of curated case law, the 2,700+ domain experts who annotate and enhance that content every day, and the evaluation infrastructure Thomson Reuters builds on top of models like Claude. “That human professional is still the one who is accountable for the end work product.”

That accountability is why verification matters more here than fluency. Thomson Reuters rebuilt legal research around agents tuned for “not just search and not just retrieval, but citation validation and verification” — surfacing sources clearly so professionals can review, verify, and apply judgment with confidence. The payoff shows up in customer-reported time: research that “would take dozens of hours” now arrives “in a matter of minutes,” giving professionals a high-quality starting point to evaluate, refine, and act on. “Deep research has been a profound shift in how to think about legal research.”

Building an agent-first product

For Thomson Reuters, agents aren’t a smarter chatbot — they’re a new way to deliver existing products. Hron’s team set out to teach a single agent to use all the tools the company used to sell as standalone software, simultaneously. That reframed how models get evaluated: “Our big test for Claude is to really assess how good it is at making plans and using these tools effectively and correctly.”

CoCounsel Legal is the concrete example: it used to run separate skills one after another; rebuilt on the Claude Agent SDK, it now plans, delegates, and orchestrates across tools and content sources in real time, so a professional defines the outcome instead of dictating every step. Customer data stays protected and is never used to train third-party models. Hron traces the partnership back to Thomson Reuters being one of Anthropic’s earliest enterprise customers — the deciding factor wasn’t a benchmark score but “Anthropic’s approach to building enterprise AI,” citing transparency, safety, and responsible AI development. The first proof point was deep research in legal, built jointly as both teams noticed how Anthropic’s own engineers used the tools the way Thomson Reuters was already shipping them.

What knowledge work demands of a model

Hron’s team has settled on four requirements a model has to clear before Thomson Reuters trusts it with professional work:

  1. Check its own citations. The CoCounsel Legal system validates what it cites before presenting findings to a human for final review — rather than retrieve a source and move on.
  2. Hold steady across long chains of tool calls. Longer tasks demand better context management and dependable tool use over an extended run, so an agent finishes real work instead of stalling halfway.
  3. Bring a person into the work, not just the answer. For the hardest jobs, Hron wants a model that will “bring the human into the loop of developing a work product rather than just relying on the agent to one-shot an answer.”
  4. Free up time for work the team didn’t have bandwidth for before. Thomson Reuters is developing advanced drafting for complex legal work — including motion drafting, filings professionals would otherwise “spend days or weeks perfecting” — work that “always required far too much context and precision” for earlier models. With Claude Fable 5, Hron says it’s now within reach.

The ROI of AI

Hron takes a contrarian view on AI ROI: “If you try to optimize too much for the rate of return calculation, you miss the forest for the trees.” He wants teams to feel the cultural and mindset shift before tuning for cost per task — once that shift happens, “the returns follow on their own.” He still tracks conventional engineering measures (DORA — DevOps Research and Assessment — and idea-to-production time), and points to a concrete result: an internal Claude-based error-remediation tool turned a production issue from three hours of root-cause analysis into a four-minute fix. “The ability to get back to health within minutes versus hours is a material difference.”

The deeper change, per Hron, is to the work itself: “The act of writing lines of code is no longer the job” for his engineers — systems thinking, judgment, and taste matter most now. He sees the same pattern spreading beyond engineering, with AI making people “more T-shaped,” able to reach across product, design, and finance rather than staying in one lane.

What’s next

Hron’s team wants to keep pushing longer-horizon work, better context management, and tool calling reliable enough to trust across a full chain of agent tasks. He’s also a personal user of the stack: Claude Code has let him “be far more technical again,” coming up to speed on a codebase he hasn’t touched in months within minutes rather than a day, and he turns to Claude Cowork to take on the perspective of a CFO or strategy officer and pressure-test ideas. His framing for the frontier still ahead: professional AI has to work in environments where “being almost right is not good enough” — work that ultimately has to hold up in court.

Try It

  • Name your own “Fiduciary-Grade” (or equivalent) bar explicitly if you operate in a regulated or high-stakes domain — Thomson Reuters’ three-part framing (authoritative content + domain expertise + workflow integration) is a reusable checklist for any team building AI a professional has to be willing to put their name on.
  • Evaluate models on planning-and-tool-use quality, not just raw output quality, when building an agent that orchestrates many existing tools/products into one interface — this is explicitly Thomson Reuters’ “big test” for CoCounsel Legal.
  • Resist over-optimizing the ROI calculation before the cultural shift happens. Hron’s advice: let teams feel the mindset change first; the cost-per-task numbers improve once people actually change how they work.

Open Questions

  • The exact publish date of this post was not captured in the Tavily extraction — it sits in the same July 2026 “Working at the Frontier” case-study series as the Hebbia and Cognition posts, but no explicit dateline was visible on the fetched page (unlike the Cognition post, which carried a “July 10, 2026” dateline).
  • No specific figures are given for CoCounsel Legal’s user base, accuracy benchmarks, or the “advanced drafting capabilities” still in development for motion drafting — these are described as directional, not measured.
  • The internal error-remediation tool (3 hours to 4 minutes) is not named or described beyond the single anecdote — unclear if it’s a packaged product or an internal-only build.

Graph View

Backlinks