AI & Intelligence · AI Governance
Your firm — not an AI vendor — controls the models, the data access, the tools, and the approvals.
Built into EdgeLex
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 problem
Most legal AI governance is a policy PDF and a prompt asking the model to be careful. When the model is wrong, you find out after the answer — with no record of what it was allowed to see or do.
How model policy actually works
Allowed. Enforced. Keyed. Metered. Attributed.
Follow one firm’s model policy through the system — every screen below is the real product.
01 · Allowed
Every model is a deliberate decision.
Provider catalogs arrive with every model disabled. Each one is allowed or denied by the firm — with its capabilities, context window, and price on the row. Refreshing a catalog never changes firm policy.
Official provider pricing · View source
New catalog models are added disabled. Refreshing prices never changes firm model policy.
…59 models in this catalog · 11 enabled by firm policy
02 · Enforced
The effective policy is one honest list.
25 allowed. 214 denied — each with its reason, per realm. This is what the runtime will actually run, and the list includes your own fine-tuned firm model alongside frontier and local options.
✓ Allowed Models (25)
GPT-4.1(openai)
GPT-4o Mini(openai)
Claude Haiku 4.5(anthropic)
Claude Sonnet 5(anthropic)
Grok 4.3(xai)
Llama 3.1 8B(local)
DeepSeek R1 14B(local)
Qwen3 14B(local)
Harborview Firm Model v2(local · fine-tuned)yours
…16 more
⊘ Denied Models (214)
gpt-4o — Model denied for realm harborview
o3-deep-research — Model denied for realm harborview
mistralai/mixtral-8x22b — Model denied for realm harborview
nvidia/cosmos-reason2-8b — Model denied for realm harborview
claude-3-5-haiku — Model not activated for this realm
moonshotai/kimi-k2-instruct — Model denied for realm harborview
…208 more, each with its reason
03 · Keyed
Your keys, your billing relationship.
Bring your own provider credentials, realm-scoped, validated on demand, revocable in one click. Frontier AI under the firm's own accounts — never a black-box pass-through.
Tenant provider keys · 2 active
Store realm-scoped provider credentials for tenant BYOK routing.
validated 5/23/2026active · valid
…k4Tz · a41c9e02d7b8ValidateDelete
5/23/2026disabled
…xxxx · 08f2d1c39aa4ValidateDelete
validated 6/12/2026active · valid
…mQ2w · 5b7e88f01d23ValidateDelete
04 · Metered
Every token lands on a ledger you can check.
Requests, tokens, cost, and monthly projection — plus provider-side reconciliation that fetches the organization's usage straight from the source, so the firm's ledger can be audited against the provider's.
TOTAL REQUESTS
10,786
INPUT TOKENS
305,944,215
OUTPUT TOKENS
7,698,068
EST. COST
$975.41
PROJECTED MONTHLY
$254.45
Cost by provider · all recorded usage
anthropic2,524 req · 51,054,819 tok · $595.38 · 61%
openai5,402 req · 148,640,487 tok · $275.44 · 28%
zai1,688 req · 64,350,721 tok · $94.82 · 10%
local (firm GPU)950 req · 4,762,581 tok · $0.00 · 0%
05 · Attributed
Cost has a client, a matter, and a feature.
Spend breaks down by credential source, by feature, by user, and by client and matter — flowing through the same expense pipeline as every other cost in the practice. AI time becomes billable truth, not overhead mystery.
Cost by credential source
Platform Keys87.2%
Tenant BYOK12.8%
Cost by feature
legal_graph58.1%
lex_chat25.9%
lex_tool15.8%
lex_summarization0.2%
Cost by client / matter
Hale v. Northstar Logistics — breach of contract119 req$9.97
Harborview — bylaw revision110 req$5.90
In re Delgado — Chapter 1173 req$3.66
Every run bills to a matter or is absorbed knowingly — the same expense pipeline the rest of the firm uses.
The difference
Policy lives in the database, not the browser.
The runtime checks firm policy before any model call. A client application cannot override it, a clever prompt cannot route around it, and a new model in a vendor catalog arrives disabled until the firm says otherwise. Governance isn’t a settings page — it’s the authority layer everything runs under.
The architecture underneath →Runtime authority
Checked in the database before every model call — clients cannot override it.
Default-deny
214 models denied with reasons; new catalog arrivals start disabled.
Any model, governed
Frontier under your keys, local on your GPUs, your own fine-tune — one policy.
Capabilities
What it does
Default-deny model governance
Three layers of policy — platform, firm, user — decide which models may run at all. No model runs unless your firm allows it — frontier, local open-source, or the firm's own fine-tuned model, all under the same policy.
The task is classified before the model runs
Each request is classified into a governed task type first, producing a contract for the turn: what evidence is required, which tools are allowed, whether the model may answer at all.
Capability binding
Only the tools a task actually needs are bound for that turn. The model can't call what it was never given.
Typed evidence, not vibes
Every claim's basis is recorded — a rules-engine packet, a document span, an email span, a billing ledger, a court order's own words, an approval artifact — and verified against a typed evidence ledger. Ungrounded legal claims are blocked.
Approval-gated actions
Actions that change the world — send, save, update, delete — require an approval artifact. Proposed and completed actions are never conflated.
Model routing with teeth
A router scores models per task and structurally refuses to let underpowered models perform destructive actions.
Real-time cost accounting, down to the matter
Every request is metered at the runtime with registry rate cards: cost by provider, by model, by user — and by client matter, so AI spend maps to the file and can be billed, absorbed, or examined deliberately. Local models run at zero marginal token cost.
Audit without leakage
Every governance decision is logged in one event taxonomy — and prompts, documents, and privileged content are redacted from the logs themselves.
In your control
Governance is architecture, not policy prose: the deny happens before the model runs, the evidence check happens before the answer renders, and the audit trail is written either way.
Related
Works with
Lex AI & Agents
Lex isn't one application among many. It's the horizontal layer connecting matters, documents, email, messages, meetings, deadlines, and billing — grounded in your data, gated by your approvals.
Cascade
Deadlines computed from the actual rules, chained to the facts that triggered them, and recalculated with full lineage when the case moves.
EdgeCite
Citation governance: extract, validate, and verify citations against your case library before they ever reach a filing.
See EdgeLex on your own terms.
We'll walk through self-hosting, model control, and your firm's workflows.