Platform Overview · 01 / 05

Legal intelligence the firm actually owns

Rent the model.
Own the ontology.

AtlasAI deploys agents that map your firm's data into a firm-defined ontology. Inside your own Azure tenant. Curated by AI Agents. Read by any AI tool connected via our MCP.

01 / Agentic Curation · sources → ontology

Your firm's knowledge lives in twelve places. Atlas brings it to one.

iMiManageSyncing
SPSharePointSyncing
ODOneDriveSyncing
NDNetDocsSyncing
@Outlook · EmailSyncing
FSNetwork sharesSyncing

Atlas Context Graph Per-tenant nMatter nParty nClause nObligation nClient nDefinedTerm

02 / Build on the Graph · ontology → apps via REST · SDK · MCP

One graph. Every team. Every tool connected to our MCP.

Atlas Context Graph Per-tenant nMatter nClause nParty nClient nDefinedTerm

CAConflicts AssistantRunning
MIMatter Intake BotRunning
PEPE DiligenceRunning
CMCompliance MonitorBuilding
W·Word redline pluginRunning
DBPartner dashboardBuilding

In this overview → 02 Ontology · 03 Agentic Curation · 04 Enterprise Search · 05 Build with Agents

Platform Overview · 02 / 05 · The Ontology

02 / The Ontology

Not a search index. A firm-defined ontology.

Atlas turns the firm's institutional memory into a structured graph of entities, relationships, and policy. Editable by the KM team. Versioned. Queryable by any AI tool the firm picks.

An ontology is the firm's editorial schema for what gets extracted, related, and queried. Atlas ships with a base ontology for legal work and an in-product editor for the firm's KM team to extend it.

Every ingestion agent is bound to the ontology. That binding is what keeps the corpus coherent across thousands of matters and decades of memory. The agent does not invent fields. It reads the firm's source systems and maps what it finds onto the schema the firm has defined.

The ontology is also where governance lives. Policy hooks for privilege, ethical wall scope, retention, and confidentiality are first-class on entities and relationships. The query runtime enforces them at every read.

What's in the box

Entities for clients, matters, parties, judges, courts, statutes, clauses, deal terms, exhibits, witnesses, regulators.
Typed relationships like represented_by, opposed_by, cited_in, supersedes, governed_by.
Attributes for jurisdiction, matter status, fee arrangement, governing law, effective date.
Policy flags for privilege, ethical wall scope, retention class, confidentiality tier.
Practice-area taxonomies for M&A, litigation, regulatory, ship as starting points.

E / Entity types

Firm-extensible types.

Base set is editable. New types are added through the ontology editor without engineering involvement. Each type carries a definition the agents read.

R / Relationships

First-class and traversable.

Typed relationships connect entities. Traversal is native, not an afterthought of vector similarity. Queries can hop from client to matter to opposing counsel to outcome.

A / Attributes

Typed and validated.

Firm-defined fields with type constraints. Validated at ingestion. Dates parse as dates, jurisdictions resolve to a controlled list, amounts carry units.

P / Policy hooks

Privilege, walls, retention.

Per-entity flags govern visibility at query time. Privileged work product, ethical-wall-scoped matters, and retention class all enforced at the graph layer.

V / Versioning

Schema versions tracked.

Ontology edits ship without re-ingesting the corpus. Historical extractions remain bound to the schema version they were created under. Roll forward, roll back.

T / Taxonomies

Practice-area starters.

Reusable extensions for M&A, litigation, regulatory, employment, tax, and IP. Firms adopt as starting points and modify to match their own editorial standards.

Why ontology beats vector search

Vector similarity guesses. An ontology answers.

1

Hop, do not guess. "Which clauses did we negotiate for this client?" resolves through typed relationships, not nearest-neighbor noise.

2

Provenance, not paraphrase. Every fact in the graph is bound to a source document and offset. Verifiable, not summarized.

3

Policy at the field level. Privilege, ethical walls, ACLs, all evaluated per-read against the entity's policy hooks.

4

Stable identity. Counterparties keep one identity across thousands of matters. Aliases resolve. Mergers resolve. The graph remembers.

Platform Overview · 03 / 05 · Agentic Curation

03 / Agentic Curation & the Librarian

Agents do the mapping. The librarian governs the graph.

Configurable ingestion agents read the firm's source systems continuously. The librarian workflow keeps the firm's knowledge managers in the loop on every extraction, every confidence score, every policy flag.

How an ingestion agent works

01

Bind

Attach the agent to one source connector. iManage workspaces, SharePoint sites, NetDocs cabinets, an Outlook mailbox, or a folder. Scope: single matter type, a practice group, or the whole firm.

02

Instruct

Describe the extraction job in plain English. Reference entity types from the ontology, not freeform tags. The agent's instructions become a versioned artifact.

03

Score

Every extraction returns a calibrated confidence score. Above the auto-commit threshold the fact enters the graph. Below it the fact routes to the librarian's review queue.

04

Run

Agents run continuously, not in nightly batches. New filings, new mail, new folders appear in the graph (or in the queue) within minutes of arriving in the source.

Throughput

Tested up to 60,000 documents per agent per matter, with 12 simultaneous extraction questions, in 90 minutes. Horizontally scales with tenant resources.

Confidence model

Calibrated per agent during a 24-hour shadow run on historical data. Defaults: auto-commit ≥ 0.85, review 0.55 to 0.84, discard < 0.55. Firm-tunable.

Hallucination control

Every extraction is bound to a source citation and document offset. Nothing without provenance enters the graph.

Re-run and reconciliation

Agents can be rerun against updated ontology or improved models. Re-extractions are diffed against existing graph state. Conflicts route to the librarian.

Source mutation

Read-only by default. AtlasAI never writes back to source systems unless a connector is explicitly granted write scope.

The librarian workflow

Knowledge managers govern the graph through one console.

1

Review queue. Per-agent queue of low-confidence extractions. One-click accept, edit, or reject. Decisions retrain the agent's calibration.

2

Ontology editor. In-product editor for entity types, relationships, attributes, and policy hooks. Versioned and previewable against the live corpus.

3

Agent tuning. Per-agent dashboards for precision, recall, queue depth, and override rate. KM tunes thresholds and instructions without writing code.

4

Audit trail. Every graph mutation (agent or human) is logged with actor, timestamp, source citation, and prior state. Roll back at entity or relationship level.

5

Editorial standards. Firm-authored style guides and curation policies live alongside the ontology. New reviewers onboard against the firm's actual standards.

6

Roles. reviewer, steward, ontology editor, auditor. Mapped to Entra ID or Okta groups. Ethical-wall membership respected throughout.

Platform Overview · 04 / 05 · Enterprise Search

04 / Enterprise Search

Contextual retrieval. Citations on every answer.

Natural-language queries run over the firm's curated graph, not raw documents. Every fact in every answer carries a back-link to its source. Ethical walls and ACLs enforced at retrieval, not bolted on after.

Atlas Search is a contextual query layer over a structured corpus. Attorneys ask questions in plain English. The query layer plans across entity lookups, document retrieval, and relationship traversal to assemble an answer.

Because the corpus is curated, retrieval is precise rather than probabilistic. The system does not paraphrase. It hops across typed relationships, pulls the relevant clauses, and returns an answer with citations down to the document offset.

The same retrieval surface backs the chat experience, the tabular review tool, and the document editor. Anything an attorney can ask in chat is also available to an agent through the application data layer.

First-party surface

Chat with the curated graph, citations on every fact.
Tabular review over thousands of documents in one workflow.
Document editor with Word-compatible export.
Assistants for practice-area workflows.
Research across the firm's matters and public sources.

Q / Query

Natural-language planning.

The query layer plans across entity lookups, document retrieval, and relationship traversal. Multi-hop questions resolve in one pass.

M / Matter-aware

Scope to the matter.

Queries scope to a specific client, matter, deal type, or counterparty. Cross-matter intelligence stays inside policy boundaries.

C / Citation

Every field cited.

Every fact carries a back-link to its source document and the exact offset it was extracted from. Verifiable without leaving the surface.

W / Walls

Ethical walls enforced.

The query runtime obeys iManage and SharePoint ACLs and firm-defined ethical walls. No cross-wall leakage at retrieval.

T / Tabular

Tabular review at scale.

Run structured extraction questions across thousands of documents in a single workflow. Output is a structured table, exportable to a memo.

D / Drafting

Document editor.

Authoring surface for memos and drafts assembled from retrieved facts. Word-compatible export. Native Word plugin in private beta.

95%

Lower token spend vs. raw-doc AI

60K

Documents per tabular review, 90 min

100%

Of retrievals carry source citations

0

Cross-wall leakage at retrieval

"What work have we done for this client, and which clauses did we negotiate?" Atlas answers from the firm's own graph, matter, counterparty, and clause-level, in one hop. No vector-similarity guessing.

Platform Overview · 05 / 05 · Build with Agents

05 / Build with Agents · Governance · Compliance

Claude Code, Cursor, your own agents. Governed by your policy.

The MCP server, REST APIs, and SDKs let any compliant tool consume the firm's curated graph. Identity, authorization, and audit flow through every call.

Atlas exposes the firm's curated graph through one open surface. A documented REST API, a GraphQL endpoint, webhooks, TypeScript and Python SDKs, and an MCP server that any MCP-compatible model or coding agent can call.

That means a partner can ask Claude Code to scaffold a conflicts assistant, point it at the Atlas MCP server, and have a working tool against the firm's own graph within a working day. Cursor and Copilot do the same. So do the firm's own engineers.

Governance is not bolted on. Every MCP tool call carries the calling user's identity. The graph runtime enforces the firm's ACLs and ethical walls per call. Every call is logged. Every model invocation is policy-routed.

The tool surface

MCP server for Claude, Cursor, Copilot, any MCP-aware agent.
REST API with an OpenAPI 3.1 spec.
GraphQL endpoint generated from the firm's ontology.
Webhooks for graph mutations, new entities, librarian decisions.
SDKs in TypeScript, Python, .NET.

# MCP client manifest, AtlasAI server entry
{
  "name": "atlas-graph",
  "transport": "https",
  "endpoint": "https://<tenant>.atlas.firm.com/mcp",
  "auth": "oauth_pkce",
  "tools": [
    "graph.lookup_entity",
    "graph.traverse",
    "graph.retrieve_documents",
    "graph.tabular_extract",
    "graph.citation"
  ]
}

Authentication

OAuth 2.1 with PKCE. Per-user tokens carry the user's identity into the graph runtime. ACL and ethical-wall enforcement applies at every tool call.

Authorization

Source ACLs honored at retrieval. Ethical walls enforced at the graph layer. Policy hooks evaluated per read.

Model routing

Firm policy controls which models can be called for which purposes. Routing decisions are audited. Default deny on any model outside the firm's allow list.

Audit

Every MCP call logged with calling user, tool, arguments, returned facts, citations. Flows to firm SIEM in CEF / JSON.

Token economics

MCP tools return retrieval results, not entire documents. Model context windows stay small. ~95% lower token spend versus dump-the-DMS-into-context patterns.

Deployment · Security · Compliance

Inside your Azure tenant. Audited to your SIEM.

1

Three operating models. Firm-operated Azure, AtlasAI-managed, or hybrid.

2

Any Azure region. US, EU, UK, CA, AU, JP, GovCloud supported.

3

Identity. Entra ID and Okta. Per-user identity flows through every call.

4

Encryption. Customer-managed keys in Azure Key Vault. TLS 1.3.

5

Certifications. SOC 2 Type II, ISO 27001. HIPAA-eligible.

6

Egress. No data leaves the firm's tenant. Telemetry opt-in.

See it in your environment. 30 minutes with the founder.

Request a demo →

The AtlasAI Platform: rent the model, own the ontology.

Rent the model.Own the ontology.