Google’s new legal AI exposes a bigger battle over the enterprise stack
The New Stack Meredith Shubel
Google launched a legal AI built for contract review, drafting and discovery. It shows the real fight is moving from the model itself to the whole enterprise stack.
Based on reporting by The New Stack, Meredith Shubel — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Google Cloud has rolled out Gemini Enterprise for Legal, a new agentic AI package aimed at legal work. The pitch is simple enough: help with contract review, regulatory monitoring, document drafting, data discovery, DSAR responses, and the rest of the paperwork jungle that keeps law teams busy. Google also introduced Gemini Enterprise for Financial Services at the same time, and says these are the first in a wider set of specialized industry products built on its governed Gemini Enterprise platform.
The timing is awkward, in a good way for readers. Google’s legal launch landed about 24 hours after Thomson Reuters unveiled Thomson, its own AI model for legal, tax, and compliance work. Thomson Reuters says it spent $40 million developing it. On the surface, both companies are chasing the same prize: AI that’s actually useful inside professional workflows. Under the hood, they’re taking very different routes.
Google is not trying to make one giant legal model and call it a day. It is wrapping its existing Gemini models in agents, integrations, tools, and governance, with some model tuning possibly included. The company says Gemini Enterprise for Legal sits on its broader AI stack, from global infrastructure and custom silicon to foundation models and an AI-ready data platform. The legal product’s core pieces include specialized skills for agents, secure MCP integrations with platforms such as DocuSign, access to third-party agents and partners like Accenture and Deloitte, and a control plane for risk management, audit logs, and governance.
That setup is meant to do very specific work: track legislative updates, court dockets, and supervisory bodies so policy drafts can be updated; speed up contract review and negotiations; draft NDA documents; prepare court filings; redact legal documents; and build or refresh contracting playbooks. Google is betting that specialization can come from the layers around the model, not just the model itself.
Thomson Reuters is making the opposite argument. Its Thomson model was trained further on proprietary content and expert input drawn from its own material, including Westlaw, Practical Law, Checkpoint, and Reuters. The company says Thomson beat Gemini 3.1 Pro, Claude Opus 4.8, and GPT-5.5 in some benchmark evaluations. And yet Thomson Reuters is not betting everything on Thomson. Its CoCounsel Legal assistant runs on Anthropic’s Claude Agent SDK, and Thomson will first be used inside Tabular Analysis, the document-review tool for large volumes of documents. The company wants to route work to whichever model fits best.
That’s the real shift here. The next enterprise AI fight may not be about who has the biggest model. It may be about who owns the data, the integrations, the workflow, and the bits of software that make the model hard to replace.
My take — AI-written commentary, not fact-checked reporting
The neat little myth that “the best model wins” is already getting old, which is inconvenient for anyone selling pure-model fireworks. Enterprise buyers want the messy stuff: permissions, audit trails, integrations, and data they don’t have to beg for. That’s not glamorous, but it’s how serious software gets built, and it’s why the stack is where the moat now lives.
Read more about this at: The New Stack