AI & Intelligence · Lex AI & Agents
The intelligence layer of the firm — Lex follows the work everywhere, and only acts when you say so.
Included with EdgeLex · Available across web, mobile, and desktop
Good morning, Maya — you have 8 overdue tasks.
/task/deadline/emailType / in the message box to route a turn — filing, deadline, draft, email, billing, and more.
Ask Lex to research, review, draft, or explain… (type / for commands)
The problem
Cloud legal AI lives in a separate chatbot: it answers from a model you can't see, trained on data you can't audit, disconnected from where the work actually happens. For work that is malpractice-adjacent, that's not good enough.
What working with Lex is like
Asked. Touched. Organized. Everywhere. Delegated. Governed.
From one grounded answer to a standing workforce under firm policy — every screen below is the real product.
01 · Asked
Answers arrive with their receipts — or refuse to arrive.
Ask for the MSJ deadline and Lex shows the rule-to-fact calculation, cites CCP §437c, and refuses to state a final deadline it can't verify — giving the conditional outer boundary instead. The right-panel workbench holds the document with one-click workflows beside the conversation.
📄 Result — No single verified combined filing-and-service deadline can be stated, because no reserved MSJ hearing date is present and the filing date has not been independently established.
October 16, 2026 is only the adjusted, trial-date-derived conditional outer boundary.
Governing authority: CCP §437c(a)(2)–(3) · trial date Feb 5, 2027 from 1 verified Matter source · 81-day base notice counted backward · weekend adjustment applied.
Ask Lex to research, review, draft, or explain… (type / for commands)
02 · Touched
Highlight a passage, and it becomes work.
The workbench renders any DMS document beside the conversation as governed context, exact version pinned. Select any passage and a menu appears: summarize it, ask Lex about it, save it to Matter Intelligence, turn it into a fact note, chronology entry, depo question, issue, or draft — or check it against rule authority and dates.
SUPERIOR COURT OF CALIFORNIA
COUNTY OF ALAMEDA
TENTATIVE RULING
HEARING DATE: 8/14/2026 · JUDICIAL OFFICER: HON. A. OYELARAN
CASE NO.: 26CV004182 · CASE TITLE: Hale v. Northstar Logistics
Defendants’ motion to dismiss is DENIED.
A. The complaint pleads each element of breach with the specificity Code Civ. Proc. § 430.10(e) requires; the agreement, the performance rendered, and the July 7 cessation of work are alleged with dates and amounts.
B. The first amended complaint was filed without leave of court. See Code Civ. Proc. § 472(a). The court declines to strike the FAC to avoid further delay, and directs plaintiff to follow all applicable rules going forward.
C. Like attorneys, self-represented litigants must check every citation to ensure that the cited authority exists and supports the proposition for which it is cited.
Tentative Ruling — Hale v. Northstar… Esc
⚡ Ask & use
📌 Save to the Matter
🗂 Turn into work product
⚖ Check against authority
03 · Organized
Projects hold the whole case, honestly.
A Project binds instructions, files, memory policy, and living work products to one matter — a Discovery Matrix that proves it was built from the current record, a preflight that freezes scope and budget before a single token. Context engineering as a feature, not a prayer.
The full Projects story →📁 Civil — Hale v. Northstar
Project chats use the instructions, files, and saved AI work attached here.
Living project work products
📇 Matter Source Inventory CURRENT
Built Aug 14 · 419 sources
✓ deterministic build · build receipt recorded
✓ 419 source fingerprints verified
🧮 Discovery Matrix FRESH
Built Aug 14 · 19 sources
✓ deterministic build · build receipt recorded
✓ 19 current source fingerprints verified
🗺 Matter Evidence Map SCOPE NEEDED
Tests lawyer-defined propositions against the exact current Source Inventory.
Current inventory ready · lawyer proposition required
🕰 Case Chronology NOT BUILT
What happened at each stage: filings, responses, hearing outcomes, next expected hearings.
⚠ Issue Register NOT BUILT
Issues, evidence, adverse facts, missing proof, risk, and next action.
＋ Describe any work product GOVERNED
Create a governed living Project artifact from a sentence.
04 · Everywhere
One Lex, inside every surface.
The same assistant — with the same evidence rules and the same audit trail — works beside the document in EdgeDocs, answers @Lex mentions in EdgeMessage, sits second chair in EdgeMeet, and takes voice on mobile. Not a chatbot in a separate tab.
EdgeDocs
Lex Document Expert
Briefs the document, proofreads a selection, drafts the next paragraph — edits always stop for review.
EdgeMessage
@Lex in the thread
“@Lex when is our opposition due?” answered in-thread with receipts. DMs can never route to the AI.
EdgeMeet
Second chair, live
Flags a contradiction against the transcript, cited to page:line — attorneys only, accept or dismiss.
EdgeMobile
Voice, governed
“Lex, what’s tomorrow on Hale?” — same evidence rules, same audit trail, in your pocket.
05 · Delegated
Tell it once, and it keeps working.
A sentence becomes a standing delegation with a locked work contract, budgets, and a kill switch — run from the Agent Control Center, every run a real task on the firm's ledger.
The full Delegations story →🤖 Agent Control Center
Standing Delegations, committed work, runtime health, and governed delivery.
Matter Watch — Hale v. Northstar ✎
💬 Discuss with Lex📄 Process a document…⏸ Disable✎ New version🗑 RetireHealthyEnabled
Summarize each newly linked Matter document in exactly 5 bullets.
Definition v3 · 24 linked Tasks · Updated 8/13/2026, 10:24 PM
ACTIVE TASKS
4
IN REVIEW
20
COMPLETED
0
DELIVERY FAILURES
0
06 · Governed
Everything above runs under firm policy.
Which models Lex may use, what each turn may touch, and what every token cost — decided by the firm in the database, enforced before any model call, attributed to the matter.
The full AI Governance story →AI Model Governance
Control which AI models the Lex Runtime can use, set defaults, and monitor usage.
🛡 Runtime Governance Active authority
The Lex Runtime checks firm policy in the database before any model call. Browser clients cannot override this policy.
The difference
Grounded, or it doesn’t ship.
Every claim in a Lex answer is typed to its evidence — a document span, an email span, a verified citation, a rule authority. A claim with no supporting span is blocked and shown as blocked. Other tools cite; EdgeLex enforces.
How Lex researches →Typed evidence
Document spans, email spans, verified citations, rule authorities — per claim.
Approval gates
Lex proposes; you approve. Nothing consequential happens on its own.
On the record
Every answer, tool call, and action lands in the matter's audit trail.
Capabilities
What it does
Grounded conversational AI
Multi-turn chat that answers from your matters, documents, emails, and court rules, with streaming responses and cited sources. Answers that can't be grounded are blocked rather than guessed.
Follows the work, not a chat tab
Lex participates where the work lives: as a teammate you can @mention in Messages, a second chair in depositions, a document expert inside EdgeDocs and Word, a triage layer in EdgeMail, and a voice assistant on mobile.
Matter-aware
Scope a conversation to a matter and Lex pulls that matter's documents, tasks, deadlines, notes, and communications into context.
Standing delegations
Beyond conversation, Lex takes standing delegations: plain-language instructions that wake on the firm's events and run inside explicit, whitelisted actions — see Lex Delegations for the full engine.
Approval-gated actions
High-impact actions — send email, save a document, complete a court-mandated task — require explicit human approval. The AI proposes; you approve.
Cross-domain workflows
One request can span the platform: an incoming court order becomes identified events, calculated deadlines, proposed calendar entries and tasks, and a team notification — with the evidence and action history attached to the matter.
Model choice & routing
The firm picks the models — frontier, local open-source, or its own model fine-tuned in the Lex Training Center. An internal router scores models per task and refuses to let underpowered models perform destructive actions.
Governed memory
Lex can remember a user's preferences and facts under a scope-safe contract. Memory is context, never used as legal authority.
In your control
Three-layer model governance (platform / firm / user) with default-deny, full token-and-cost attribution per user, matter, and model, and tool allowlists per agent.
Related
Works with
Lex Delegations
Standing delegations you create in plain language. Lex watches the matter, wakes when your firm's events fire, does the work inside your approval gates — and every run lands on the task ledger like any other member of the team.
Lex Research
Lex researches the way an associate does — plans the question, works your matters and documents, verifies caselaw against a national citation ledger inside your deployment, pulls opinions on demand, and returns a cited memo where every claim carries its evidence or gets blocked.
AI Governance
A governance layer built into the platform: classify the task before the model runs, bind only allowed tools, require typed evidence, gate actions on approval, and audit every decision.
EdgeMessage
One universal inbox across every EdgeLex surface: direct messages, team threads, document conversations, and Lex as an @mentionable teammate.
See EdgeLex on your own terms.
We'll walk through self-hosting, model control, and your firm's workflows.