Thomson Reuters CTO Joel Hron: legal industry is on fire for AI, RAG and domain experts key to hallucination control
Jul 1, 2025 with Joel Hron
Key Points
- Thomson Reuters treats its 4,500 domain experts as a structural moat, using their real-world evaluations to give model providers feedback that yields more specialized versions of frontier models in return.
- Legal tech adoption has accelerated over two years as capital floods the sector, while tax and audit lag behind due to weak quantitative reasoning, though recent agentic advances are opening those workflows to automation.
- Thomson Reuters buys frontier model capability from Anthropic, OpenAI, and Google rather than pre-train its own, leveraging a proprietary 1.5 trillion token corpus as its primary negotiating asset.
Summary
Thomson Reuters is far more than a news business. Joel Hron, CTO, says the Reuters brand obscures what is actually a top-two or top-three global software provider across legal, tax, compliance, audit, and risk — with news representing a relatively small share of revenue.
Legal Is the Lead Vertical
Of all its markets, legal has been the standout AI adopter for the past two to two-and-a-half years. Hron describes customer demand as intense, with capital flooding into legal tech and AI tools visibly reshaping how lawyers structure their work and underlying business models. Tax and audit are accelerating but started from a lower base, held back initially by models' weaker quantitative reasoning. The rise of agentic frameworks and improved tool-call capabilities over the last six to nine months has changed that calculus, opening the door to meaningful workflow automation in those sectors.
Build vs. Buy Strategy
Thomson Reuters is primarily a buyer of frontier model capability, with active partnerships across Anthropic, OpenAI, and Google. The company has completed fine-tuning engagements with several of those providers and is tracking reinforcement fine-tuning closely as a next step. It is not pre-training GPT-4 or Claude-scale models.
That said, roughly 18 months ago Thomson Reuters acquired a small company staffed by former DeepMind researchers to build out internal deep learning training expertise. The strategic asset underlying all of this is approximately 1.5 trillion tokens of proprietary content spanning legal, tax, and news — one of the largest domain-specific corpora of any enterprise software company.
Domain Experts as a Competitive Moat
With reinforcement learning increasingly driving model improvement, Hron positions Thomson Reuters' 4,500 domain experts across legal and tax as a core differentiator. Those specialists build real-world evaluation rubrics — assessing outputs not on right-or-wrong binary terms but on whether a response is genuinely useful or potentially harmful to a practicing lawyer. Foundation model providers rely on that feedback to improve their training data, and Thomson Reuters receives more domain-optimized models in return.
Public benchmarks, in Hron's view, fail to capture how professionals actually work. The proprietary eval infrastructure Thomson Reuters has built is part of what it brings to the table in commercial negotiations with model providers — alongside the content corpus itself.
Hallucination Remains the Core Engineering Challenge
Hron is direct that hallucinations will persist regardless of model improvements, given the probabilistic nature of how large language models operate. Thomson Reuters invests equally on two tracks: algorithmic guardrails to reduce hallucination rates, and UX design to ensure lawyers can identify what needs verification and complete that validation quickly. The goal is to preserve net time savings even when human review remains necessary. In high-stakes domains like legal and tax, the product experience around error detection is treated as just as important as the underlying model accuracy.