Thomson Reuters Just Built Its Own AI. Your SaaS Vendor Is Watching.

Last week, Thomson Reuters did something that should make every business owner pay attention. They announced "Thomson," a proprietary large language model trained on 30 years of their own legal and financial data.

The price tag: $40 million. The payoff: an AI model that runs at a fraction of what ChatGPT or Claude costs for the same legal tasks. And Thomson Reuters owns every bit of it.

This Is a Business Strategy Story

Thomson Reuters sits on one of the richest datasets in professional services. Westlaw case law. Practical Law templates. Checkpoint tax guidance. Reuters financial reporting. Decades of structured, verified, domain-specific content that no general-purpose AI can match.

So instead of paying OpenAI to process their customers' legal documents, they built a model tuned for legal and financial work. It's already running inside CoCounsel, their AI legal assistant, handling high-volume document review and tabular data analysis.

I think the most interesting detail is that they've only used about 10% of their proprietary content so far. They can keep expanding what this model does without starting from scratch.

What This Means if You Pay for SaaS

If you run a business that pays for AI-powered software, something just shifted under your feet.

Your software vendors are watching Thomson Reuters closely right now. Every SaaS company sitting on years of customer data is asking the same question: should we build our own model instead of renting access to GPT or Claude?

For some of them, the answer is going to be yes. And when they make that move, three things happen.

First, your costs could drop. Custom models tuned for specific workflows are cheaper to run than general-purpose models. If your legal software vendor builds something that handles contract review at a tenth of the inference cost, some of those savings should flow to you.

Second, performance gets better. A model trained on tax code and lease agreements will outperform one that also knows how to write poetry and explain quantum physics. Specialization wins here.

Third, you get more locked in. Once your vendor controls the AI layer, switching costs go up. Your data trains their model. Their model powers your workflow. That accumulated intelligence doesn't follow you when you leave.

The Small Business Version

You probably aren't spending $40 million on a custom LLM. But the principle scales down pretty well.

I work with businesses every week that have years of proprietary data sitting in spreadsheets, CRMs, and shared drives. Customer interaction patterns. Pricing decisions. Service playbooks. That data has real value when paired with AI.

You don't need to train a model from scratch. You can fine-tune existing models on your domain knowledge. You can build retrieval systems that reference your own documents. The point is the same one Thomson Reuters just proved at $40 million scale: your data is the real advantage.

General-purpose AI is a commodity. Everyone can access the same foundation models. What separates your AI setup from a competitor's is the specific data you feed it.

What to Do About It

Start by cataloging what proprietary data you actually have. Client communications, project outcomes, internal processes, pricing history. Then look at where AI tools are already touching your workflow.

Everyone's using AI at this point. The question worth asking is whether you're building on your own data or renting someone else's. Thomson Reuters just showed that the biggest players are choosing to own theirs. You should think about doing the same, even if your version costs $400 instead of $40 million.