The Disruptors Just Got Disrupted
Three years ago, everyone in professional services was bracing for the same story.
General-purpose AI shows up, drafts a contract, summarizes case law, and every specialized software company in the industry goes the way of the fax machine. Why pay for Westlaw when ChatGPT can do the job for free?
Thomson Reuters has an answer. And it's not the answer Silicon Valley expected.
We got into this on Episode 503 of The Accounting Podcast, and Hector Garcia, CPA, joined us for the conversation.
They built their own model
Thomson Reuters launched Thomson, a proprietary large language model (LLM) trained on 175 years of proprietary content from Westlaw, Practical Law, Checkpoint, Reuters, plus input from thousands of subject matter experts.
The project traces back to their 2024 acquisition of Safe Sign Technologies, a UK-based startup that developed legal-specific LLMs.
Thomson Reuters took an open-weight model and fine-tuned it on the library they've spent a century and a half building. It’s the stuff that's always been behind a paywall because it's actually valuable.
Think analysis from top tax and legal experts, not text scraped off the open internet.
The benchmarks are striking.
On instruction following, Thomson Reuters' model reportedly scored 0.914, ahead of Claude (0.861), Gemini (0.848), and ChatGPT (0.885).
On deep research using Westlaw and Practical Law, it hit 0.83 on factuality (meaning claims trace back to a correctly cited source), versus 0.65 and 0.68 for two leading frontier models.
These are self-reported numbers, so take the specifics with a grain of salt until someone independently verifies them. But the direction of the story is hard to argue with.
The "general AI wins everything" thesis assumes data doesn't matter
It does.
A model trained on the open web is a generalist. It knows a little about everything, meaning it's more likely to hallucinate the second you ask it something specific and technical.
As my cohost David Leary said on the podcast, he wants "dumb AI," a model that only knows accounting and tax. Narrow its knowledge, and it's less likely to make things up.
Thomson Reuters didn't stop at the data.
They brought in partner-level practitioners who spent months building rubrics for research tasks, tested outputs against them, and had the model improve from the feedback. Around 1,500 attorney-editors reviewed and rated outputs.
A startup with a slick interface and an OpenAI API key can't replicate that level of intelligence.
You can't rent 175 years of proprietary analysis and a standing bench of experts to grade your homework.
And Thomson Reuters says it's used less than 10% of its own content to train this model so far.
Imagine how the metrics move once they train on the rest.
Data ownership is the whole game
Everyone spent the last few years worrying AI would flatten specialized software’s advantage. Turns out the advantage is the data underneath the interface, and whether you own it.
Thomson Reuters owns theirs. So do other big legal and tax research providers who've spent decades building it. The frontier models, for all their compute and talent, train on what's freely available. And freely available data isn’t the hard part.
The model's entering production now inside CoCounsel, starting with high-volume structured document review, with broader legal and tax integration planned from there.
If you're evaluating AI tools for your firm, this changes the question you should be asking.
Does the tool know your domain, or is it guessing based on whatever it happened to read online?
Where did the training data come from?
Who's reviewing the outputs?
A specialist grounded in real authoritative content beats a generalist every time, especially when the answer has to be right, not just plausible.
