The Curated Knowledge Graph: Why Firm Data Needs to Stay Inside the Tenant

The Data Layer Problem in Legal AI

For the past two years, AmLaw firms have been caught between two bad choices: either adopt AI tools that send sensitive work product to third-party LLMs, or build everything in-house at enormous cost. The industry has slowly realized there's a third path, keep the firm's data inside the tenant, shape it with AI, and build everything on top of that curated foundation.

The shift is real. Firms are no longer asking "How do we add AI chat to our existing systems?" They're asking "How do we rebuild our data layer so it's shaped by AI from the moment a document lands?" That's the move from passive, searchable storage to active, continuously-curated knowledge.

What We Built: Agents as Curators, Not Indexers

Singularity v6 introduces agentic curation. When a new document lands in iManage or SharePoint, Atlas agents don't just tokenize it and throw it into a vector store. They read it, classify it, extract structured metadata (matter, document type, parties, key clauses, privilege status), link it to related documents and prior work, and refine the firm's knowledge graph continuously.

The graph itself is stored in the firm's tenant, alongside direct integrations to iManage, SharePoint, OneDrive, NetDocuments, and email. Firms point Atlas at specific matter folders or SharePoint sites via the source browser, and the curation agents wake up. As new documents land, agents immediately classify and index them, so the graph stays sharp without manual intervention.

Access controls flow through the same layer. If you can't see a document in iManage, you can't retrieve it from the knowledge graph in Atlas. The firm's security model is the graph's security model.

How This Changes the Workflow

Before Singularity v6, the firm's data was inert. A document landed in iManage. It sat there until someone searched for it or asked an LLM about it. AI wasn't shaping how the firm's own work was organized.

Now: a document lands in an iManage matter folder. Within minutes, Atlas agents have classified it, extracted its key terms, linked it to similar documents, and added it to the firm's curated knowledge graph. When a partner runs a document review in an Atlas Workspace, or when they ask Claude a question via the MCP integration, the AI is reasoning over a graph that's been shaped by AI on the firm's own precedents, not a vanilla LLM trained on public data.

For an M&A team working on a 200-document deal: the agents have already classified each document by type (SPA, exhibit, LOI, disclosure schedule), extracted key financial terms and indemnity provisions, and linked them to the firm's precedent SPAs. The review process starts with a hot, curated graph, not a cold index.

Inside Singularity: Four Surfaces

The curated knowledge graph is the foundation. On top of it live four product surfaces:

Workspaces run agentic document review at scale. Upload hundreds of documents, agents classify and route them, execute substantive review (privilege, conflicts, clauses, indemnification, diligence findings), and draft new client documents from the findings, all grounded in the curated graph.

Lists are reusable agent workflows. A practice team builds a Deal Review list once; agents execute it on every new matter, extracting terms, comparing to precedent, generating redlines, and producing diligence memos.

Agents do two jobs: inside Workspaces they review documents in parallel and draft new client work; outside Workspaces they curate the knowledge graph itself.

MCP + API let Claude and any MCP-aware tool read the firm's matters, documents, clauses, and Lists directly, with the same access controls as the in-app experience.

What's Next

Start with the source browser: pick a few iManage matter folders or SharePoint sites. Watch the agents classify and curate the incoming documents. Then move a matter into an Atlas Workspace and run your first agentic review. The graph gets smarter as you use it.

Singularity v6 is the shift from bolted-on AI features to a curated data layer that's shaped by AI on the firm's own work. That's the platform.