Our Take
Knowledge management is finally useful—not because AI got smarter, but because firms can now choose what context the AI actually sees.
Why it matters
As Claude, Copilot, and Harvey become table stakes, law firms that simply dump document stores into them will lose to firms that curate. This is the infrastructure play that matters in 2025.
Do this week
Knowledge officer: audit your current AI integrations this week to see whether they're connected to curated metadata or raw document dumps, so you can estimate the control and quality gap you're running.
Lexsoft Connects Firm Knowledge to Three Major AI Tools
Lexsoft unveiled T3, a knowledge platform that sits between a law firm's document stores and generative AI tools, using Model Context Protocol (MCP) connectors to link with Microsoft Copilot, Claude, and Harvey. The founder and CEO Carlos García-Egocheaga demonstrated the system in a demo for Legal IT Insider.
The core claim: rather than feeding unfiltered document stores to AI, T3 structures and enriches firm knowledge with metadata, ensuring AI tools work with validated, curated content. Lawyers query knowledge conversationally inside the AI tools they already use—Copilot, Claude, or Harvey—without needing to know T3 is running underneath.
The demo showed a workflow where lawyers retrieve documents, analyse them with Harvey, and critique outputs with Claude in a single orchestrated sequence. Access to underlying documents in systems like iManage remains available, but firms retain control over what the AI actually sees: full documents or metadata only.
Control Over Context Is Now the Real Differentiator
The timing matters. Three major generative AI tools (Copilot, Claude, Harvey) are now embedded in legal workflows. All of them work best with high-quality context. The gap between firms that dump raw documents into these tools and firms that feed curated, structured knowledge is growing wider.
García-Egocheaga's diamond mining analogy captures the work: raw data is not useful until refined. A firm with 500,000 case documents benefits only if the AI can distinguish between relevant, validated precedent and noise. Metadata-enriched knowledge does that filtering. Unfiltered document stores do not.
This also addresses a real operational risk. Firms want to control whether an AI tool sees full case files or only summaries and tags. T3 makes that choice explicit and enforceable at the connector level, not buried in prompt engineering.
Three Steps to Assess Your Current Setup
First, map what your firm is currently feeding to Claude, Copilot, or Harvey. Is it raw document dumps, or is it structured, metadata-enriched content pulled from a knowledge base?
Second, identify the gaps. Do you have a way to control what each AI tool can access? Can you enforce data governance rules at the point of integration, or are you relying on post-hoc review of AI outputs?
Third, talk to your knowledge management team about whether your current document stores are tagged and structured enough to be useful at inference time. If they are not, that's the real project—not the AI tool choice.