Execution
An agent should produce a real artifact or governed action
Examples include drafting a partner briefing, creating a research memorandum, preparing a chronology, reviewing clauses or queuing authority research. The work should remain attached to the matter and be available for inspection after completion.
A visible job state—queued, running, completed, review required or failed—helps lawyers distinguish real execution from a conversational promise.
Boundaries
Autonomy should not erase professional judgment
Creating a task is different from changing a verified fact. Updating workflow metadata may be safe to execute directly, while altering evidence, privilege, parties or legal conclusions should require stronger controls.
Sanctum's delegated work model is designed to keep substantive factual changes staged or unverified until a lawyer reviews them.
Audit
Every agent action should be attributable
The firm should be able to reconstruct who delegated the work, which matter it ran against, which policies applied, which model route was used, what artifact was generated and whether a reviewer approved the result.
That record is especially important when agentic systems begin doing work asynchronously while lawyers continue elsewhere in the platform.
Frequently asked questions
What is an agentic legal AI?
It is a legal AI system that can plan or execute multi-step work rather than only answer a single prompt. Enterprise use requires matter scope, tool permissions, review and audit controls.
Should legal AI agents be allowed to update matters automatically?
Operational changes can often be safe when explicitly requested and audited. Changes to evidence, legal conclusions, privilege or other consequential matter state should usually be staged or require lawyer approval.
Can a legal AI agent keep working while a lawyer does something else?
Yes. A persistent job queue can let delegated work continue asynchronously, provided the job remains matter-scoped, observable and governed.
Continue comparing
Distinguish selected evidence from exhaustive review.
Policy and approval boundaries for agents.
Keep agents attached to the legal matter.
Separate legal verification from drafting.
Explore the delegated-work product surface.
Evaluate on your own matter