AI
Google and Thomson Reuters chart competing paths to specialized legal AI
As enterprises demand AI tailored to specific industries, rival approaches emerge: one built on integrated cloud infrastructure, the other on proprietary training data.
The New Stack
Google Cloud rolled out Gemini Enterprise for Legal this week alongside a financial-services sibling, marking the tech giant's first wave of purpose-built industry AI agents. The moves underscore a shift in enterprise AI competition: success now depends less on who builds the most powerful foundation model and more on who can best assemble specialization across the entire stack.
The timing drew attention because Thomson Reuters unveiled its own legal AI just hours earlier. Yet despite surface similarities—both target legal workflows like contract review and regulatory tracking—the two companies took fundamentally different routes to solve the same problem.
Two paths to specialization#
Thomson Reuters' approach: The company invested $40 million to develop and train its own proprietary model called Thomson. Rather than simply taking an existing general-purpose AI and fine-tuning it, Thomson Reuters fed the model decades of proprietary legal content from Westlaw, Practical Law, and Checkpoint, along with structured feedback from legal experts. The result outperformed Gemini 3.1 Pro, Claude Opus 4.8, and GPT-5.5 on several legal benchmarks.
Google's approach: Gemini Enterprise for Legal layers specialization above the foundation-model level. The product combines four core elements: specialized agent skills designed for legal tasks, secure integrations with industry tools like DocuSign, a network of third-party legal tech partners such as Accenture and Deloitte, and built-in governance controls. Google hints its solution may also include model tweaks, but the emphasis falls on orchestration, not model training from scratch.
Why owning a model doesn't mean using it exclusively#
A telling detail emerged in Thomson Reuters' announcement: despite investing heavily in a proprietary legal model, the company isn't abandoning other frontier AI systems. Thomson Reuters' legal assistant, CoCounsel Legal, runs on Anthropic's Claude Agent SDK rather than on Thomson alone.
The strategy amounts to pragmatic selection: when Thomson outperforms rivals on a specific task, it gets priority. When another model suits the job better, the system routes accordingly. This multi-model flexibility suggests that even companies building proprietary AI recognize that no single model wins every scenario.
