Claude for Legal is a significant moment in legal AI - not because it ends the legal AI race, but because it clarifies the stack.
Claude is becoming the work surface. MCP is making connectors table stakes. Open-source projects like MikeOSS show that the legal AI front end can be cloned, open-sourced, and run with a user's own model key in a weekend.
That changes the question.
The question is no longer about connectivity
The question used to be: can I connect AI to iManage? To Box? To SharePoint? To Westlaw or DocuSign?
Increasingly, the answer is yes. MCP connectors are proliferating. The integration problem is getting solved. What was a moat is becoming table stakes.
The harder question - the one that doesn't get solved by a connector - is: which result should the AI trust? Which version of the contract is official? Which clause is approved? Which memo supersedes the older one? Which answer actually reflects the firm's current position?
Across silos, where the same document exists in fourteen places and none of them are labeled "authoritative."
This is the knowledge authority problem
The knowledge authority problem is not a connector problem. It's an infrastructure problem. It requires:
- Federated retrieval that surfaces candidates from all relevant sources simultaneously
- Canonical version identification - finding the approved version, not just any version
- Source authority ranking - understanding which source should be trusted more than another for a given type of question
- Organizational ratification - mechanisms for the firm to formally designate which answer is authoritative
None of these are solved by connecting Claude to iManage. They're solved by the knowledge layer underneath the AI interface.
What the stack looks like
The legal AI stack is clarifying into two distinct layers:
The interface layer - Claude, Copilot, MikeOSS, Harvey, or any AI work surface - handles the user interaction, the generation, the conversation workflow. This layer is commoditizing rapidly.
The knowledge authority layer - federated retrieval, canonical version finding, authority ranking, organizational ratification - handles the question of what the AI should actually believe and act on. This layer is not commoditizing. It requires enterprise-specific configuration, governance, and ongoing management.
This is where SWIRL 5 is headed. The interface is becoming a commodity. The knowledge authority layer is not. And in legal, where the wrong answer isn't just inconvenient but potentially sanctionable, that distinction is everything.