Before the Click: Legal Automation Approval Decisions

What must a human approve before a legal-work automation runs?

Define the objective, information boundary, authorized actions and accountable decision owner before enabling execution. Existing authorization may cover routine steps; a new external commitment, sensitive disclosure or materially changed objective needs its own authority check. The decision tree below is a proposed operating method, not a universal legal approval rule.

A preflight decision tree

Start with the next concrete action, rather than the agent's entire job. Reading an approved public source, drafting a clause, sending that clause and accepting it are four different actions. Record the action in a sentence that names its object, destination and effect.

• Is the information and destination already authorized? If not, stop before reading or transmitting it.
• Is the action within the approved objective and limits? If yes, continue to the consequence check.
• Will it commit money, accept terms, waive a position, contact someone or change a live record? Identify who can authorize that consequence.
• Can the outcome be observed? If not, define a receipt or readback before execution.

Separate client choices from tool permissions

ABA Model Rule 1.2 assigns decisions about representation objectives and settlement to the client while recognizing lawyer consultation and some implied authority. A software permission does not answer those allocation questions. [S1]

For a founder, the practical record should identify the business decision maker and, where applicable, counsel's role. A tool's ability to access a signing system is only a technical capability; it does not resolve the underlying authority question.

• Decision owner: person entitled to choose the outcome.
• Operator: person or system carrying out the authorized step.
• Reviewer: person checking the artifact against instructions.
• Evidence owner: person preserving the approval and observed result.

Synthetic example: a vendor renewal

Assume a fictional company asks an agent to prepare a comparison of three public subscription plans. The approved scope permits public research and a local draft, with a fictional annual budget ceiling of $3,000. The agent then finds a renewal button in an existing account.

The preflight result is CONTINUE for the comparison and HOLD for renewal. The budget ceiling limits spending if spending is authorized; it does not itself authorize a purchase. The reviewer receives the exact plan, total price, renewal terms and proposed account action before deciding.

• Approved output: comparison memo version 2.
• Held action: click Renew on vendor account.
• Missing decision: authorized purchaser's approval of identified terms.
• Completion evidence if approved: vendor confirmation and account status, with sensitive details kept out of general logs.

Make the review narrow enough to perform

Use a short approval packet: the proposed artifact, its source material, consequential changes, unresolved questions and a description of what execution will do. A reviewer cannot meaningfully approve a future document whose recipient, price or operative wording can still change.

NIST's AI RMF is voluntary risk-management guidance rather than a grant of legal authority. It is a useful reference for designing the surrounding control process. [S2]

• Recheck approval if the recipient, data category, commitment or material terms change.
• Use clear states: prepared, approved, attempted, confirmed and failed.
• After a timeout, inspect the external result before repeating an action.

Scope and limitations

Educational workflow design. ABA sources describe model-rule principles, whose adoption and application vary by jurisdiction.

This page does not approve any real action, establish representation or determine who has authority in a particular organization. The vendor example is entirely synthetic.

Sources

1. ABA Model Rule 1.2: Scope and Allocation of Authority — Allocation of decisions between client and lawyer, including settlement and limited representation. Checked 2026-11-05
2. NIST AI Risk Management Framework — Voluntary AI risk-management framework. Checked 2026-11-05

Explore further

This approval-tree topic connects to the human review questions explored around Agentic Attorney. Visit https://agenticattorney.com/ to examine that central project and compare its current information with these workflow criteria. Explore Agentic Attorney.