The Mismatch Problem - MCPs Are Here

Last May, Microsoft Build got me thinking about Model Context Protocols (MCP). If AI systems could begin connecting more easily with the tools, data sources, and workflows people already use, then we were no longer talking only about better chatbots. We were talking about AI becoming part of the infrastructure of legal work.

So I wrote that MCPs could help unlock the future of legal AI by letting systems move beyond isolated silos and interact with the places where the work actually happens.

That future is beginning to arrive.

Last week, Anthropic announced a major legal expansion for Claude built around more than twenty MCP connectors and twelve new legal plugins. Thomson Reuters also announced a new MCP integration connecting Claude and CoCounsel Legal. In practical terms, legal work is beginning to move more fluidly between a general-purpose AI environment and purpose-built legal tools. And it is not just Thomson Reuters. NetDocuments says Claude can now connect directly to a firm’s document management system to search, retrieve, and draft from the firm’s own documents and precedents.

This is a meaningful moment for legal AI, and I think people should explore it. Lawyers should experiment with it. And courts should study it.

But before everyone starts connecting everything, it’s important to make sure that the pieces match.

I am no technologist, and I do not pretend to be. My expertise is in the justice system and in workflows. So if the people who spend their lives thinking about cybersecurity, enterprise architecture, and data governance see this differently, they should jump in and correct me. I mean that sincerely.

But from where I sit, legal professionals should be careful about connecting tools that may not have been built with the same security assumptions. An enterprise-grade legal tool and a $20-a-month consumer AI account are not necessarily the same kind of environment. Both may be useful and powerful. But they may not carry the same contractual protections, administrative controls, or expectations for sensitive work. Anthropic’s own materials distinguish between consumer Claude plans and commercial offerings in how data is handled, and Claude Pro remains a consumer plan priced at $20 per month when billed monthly.

If a lawyer connects a high-trust legal tool to a consumer AI environment, the profession should not simply assume that the stronger protections travel unchanged across the connection. Someone who understands the architecture needs to explain what data moves where and whether the pieces being connected were built to hold the same kind of work.

MCPs may help unlock the next stage of legal AI. And I think that is exciting.

Just be careful what you connect.