When 'Mostly Right' Isn't Good Enough: How Thomson Reuters Builds AI You Can Defend

The starting point

In a courtroom, a plausible answer that turns out to be invented is worse than no answer at all. Lawyers have been sanctioned for filing briefs with fabricated case citations produced by general-purpose chatbots, and the count of documented hallucinations in legal filings passed 1,600 cases globally by mid-2026, up from around 700 at the start of the year. For a profession where a single bad citation ends careers, the honest verdict on consumer AI is that fluent and confident are not the same as correct.

Thomson Reuters sits directly in the blast radius of that problem. It is a 175-year-old content and technology company serving lawyers, tax advisors, accountants, and compliance teams, with products like Westlaw and Practical Law that professionals have trusted for decades to be right. The company had a choice most incumbents face: bolt a chatbot onto its existing search and hope, or rebuild for a standard no general model was designed to meet.

Joel Hron, the company's chief technology officer, framed the test plainly. He assesses any model on "whether its work can withstand the level of professional review lawyers apply before relying on it." That is a much higher bar than sounding right, and it shaped everything that followed.

What they built

A named standard, not a disclaimer

Thomson Reuters gave its approach a name and a definition: Fiduciary-Grade AI, which it describes as "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 phrasing is worth reading twice. Defensibility is the target, meaning a professional has to be able to stand behind the output in a courtroom, an audit, or a regulatory proceeding. Accuracy that sounds convincing but cannot be traced back to a source does nothing for you there.

That reframes the whole engineering problem. In high-stakes work, the useful question is not "what is the answer" but "can I show my work and defend it." Hron says the company builds so customers can "trust, verify, and defend" what the system produces. Everything below is downstream of that sentence.

Verification, not search

The sharpest design decision was to rebuild legal research around citation validation and verification rather than retrieval. Old-model legal search returned documents and left the lawyer to check them. The new system is tuned to check its own citations before it presents a finding, so the professional reviews and confirms rather than hunts for errors. In a domain where the failure mode is a confident wrong reference, moving verification to the front of the process is the entire game.

An agent architecture that earns the trust

Thomson Reuters rebuilt CoCounsel Legal on Anthropic's Claude Agent SDK, orchestrating hundreds of internal tools at once instead of running skills one after another. Hron set four requirements for any model at the core of that system: it has to check its own citations before presenting findings, hold context across long chains of tool calls, bring humans into the development of the work rather than only handing over a finished answer, and open up tasks that were impractical before, such as drafting a legal motion.

On top of that sits Deep Research, launched in August 2025, which lets a lawyer hand over a full research question and get back a reasoned, sourced answer grounded in Westlaw and Practical Law content. The company's own framing is that professionals move "beyond prompting and start delegating," with human oversight kept in the loop throughout. The lawyer stays accountable for the final work product, which is the point, not a limitation.

Domain experts treated as infrastructure

The part most businesses underestimate is the human layer. More than 2,600 subject-matter experts shape the company's AI systems, annotating content and defining what good looks like. Its evaluation standard, CoCoBench, was built with over 100 legal experts contributing more than 15,000 hours, testing real legal tasks against attorney-authored gold standards rather than generic benchmark scores. To capture the visual structure of legal documents, the company partnered with Invisible Technologies on pixel-level annotation, because plain text extraction loses the relational meaning that a table or a signature block carries.

None of this is a temporary bridge until the models improve. It is a permanent input. The experts are part of the architecture, and Thomson Reuters treats them that way.

The results

What makes this case interesting

Verification is the product, not a feature. Most AI programmes bolt a review step onto the end and call it governance. Thomson Reuters moved verification to the front and made it the design goal, rebuilding research around citation checking rather than search. For any business where an output has consequences, this is the shift worth copying: build the system to prove itself, not to sound convincing.

Domain experts are permanent, not a phase. Over 2,600 experts and 15,000 hours of evaluation work are not a training cost the company expects to retire once the models get good. They are how the system knows what "defensible" means in a specific field. If your plan treats subject-matter expertise as a one-off data-labelling exercise, you are building for demos, not for professional use.

Optimising the ROI calculation too early is a trap. Hron is direct on this: "If you try to optimize too much for the rate of return calculation, you miss the forest for the trees." His argument is that the cultural shift to working differently comes before the cost optimisation, and leaders who invert that order stall. Coming from the CTO of a public company, that is a notable thing to say about business cases.

The partner choice was about behaviour, not benchmarks. Thomson Reuters selected Anthropic on its approach to building enterprise AI, citing transparency, safety, and responsible development as the deciding factors rather than raw model scores. In a regulated setting, how a vendor behaves under scrutiny turns out to weigh more than where it sits on a leaderboard this quarter.

The challenges

This is not a story of AI removing human effort. The lawyer remains accountable for the final work product, so the system refocuses professional review rather than removing it. The reviewer stops hunting for errors and starts confirming edge cases, which is a better use of expensive time, but it is still a human cost that never goes to zero.

The value is also hard to pin to a number. Hron's own "forest for the trees" comment is a tell: a company this advanced still finds the return calculation slippery enough to warn against over-indexing on it. Budget for the same ambiguity in your own business case, and do not promise a board a clean payback figure you will struggle to defend later.

And the expert layer is expensive. More than 2,600 specialists and a bespoke annotation partnership are not line items a mid-size business will match. Copy the principle behind the scale rather than the headcount: the quality of a high-stakes AI system is capped by the quality of the human judgement fed into it, and that judgement has to be resourced, not assumed.

Lessons for your programme

Ground the system in content you own and can stand behind. Thomson Reuters built on authoritative, curated content, not the open web, because a defensible answer needs a defensible source. Before you point a model at a business problem, decide what corpus it is allowed to reason over and who owns its accuracy. Section 10: RAG and Knowledge Systems covers how to prepare and govern that content.

Make verification the design goal, not the last step. The lesson from legal AI is that a review bolted on at the end catches too little, too late. Design so the system surfaces its sources and checks its own claims up front, and your governance stops being paperwork and starts being a control. Section 08: Risk and Governance sets out the classification and control work this needs.

Resource domain expertise as a standing input. The experts are part of the machine, not a phase that ends. Whatever your field, decide whose judgement defines a good output and build their time into the operating model permanently. Section 14: Operating Model for AI covers where that capability sits and how it is funded.

Do not let the ROI calculation lead the decision. Hron's warning applies well beyond legal software. Lead with the behaviour change you are trying to create, then track value honestly rather than promising a precise return you cannot defend. Section 13: Business Case Production shows how to build benefit tracking without overselling the numbers.

Sources

• Working at the frontier: How Thomson Reuters builds AI for high-stakes professional work (Anthropic, 2026)
• Thomson Reuters Shows Trusted AI Starts With Data (Forbes, 2026)
• One Million Professionals Turn to CoCounsel as Thomson Reuters Scales AI for Regulated Industries (Thomson Reuters, 2026)
• Thomson Reuters Launches CoCounsel Legal: Transforming Legal Work with Agentic AI and Deep Research (PR Newswire, 2025)
• Thomson Reuters Reveals New Agentic AI in CoCounsel Legal (Thomson Reuters, 2025)
• Building Trust in Agentic AI (Thomson Reuters Institute, 2026)
• Accurate and reliable AI: Five key ingredients (Thomson Reuters, 2026)