Triage legal AI requests before work begins

How should internal legal AI requests be triaged before work begins?

Classify what the requester wants, what information is needed, where it may go, whether the output affects others, and who can approve the result. Route the request to a bounded workflow with an accountable owner. Urgency should change prioritization, not silently expand authority or data access.

Collect five facts that determine the route

A request such as 'review this agreement' hides several decisions. Does the requester want a summary, a negotiation position, or an approved signature? Does the file contain sensitive information? Is a tool approved for that information? These questions should be resolved before processing begins. NIST's AI RMF is a voluntary risk-management reference, not authorization for a use. [S1]

• Desired output and intended reliance.
• Data category, source, and authorized destination.
• External effect: internal draft, shared advice, negotiated term, signed document, or filing.
• Decision owner and reviewer.
• Timing and the source of any claimed deadline.

A proposed three-route decision model

Route by the combination of consequence and permission rather than by document length. A one-line instruction to send a termination notice can be more consequential than summarizing a hundred public pages.

• Route A: public-information research or synthetic experimentation; define source and output limits.
• Route B: internal work on authorized private material; confirm approved environment, owner, and review.
• Route C: legal advice, external communication, commitment, or exceptional data use; require the applicable professional and organizational approvals.
• Unresolved data permission or authority: hold that dependent step and continue only separable authorized work.

Synthetic intake: 'approve the vendor today'

A fictional procurement request arrives with a draft vendor contract and a same-day target. Triage splits the work into a commercial comparison, a security review, and legal interpretation. The agent can prepare the comparison from authorized material. It cannot treat the phrase 'approve today' as proof that the requester may waive security conditions or bind the company.

• Requester clarifies the business need and supplies the current contract version.
• Security owner assesses the proposed data flow.
• Legal owner reviews identified legal issues.
• Authorized business signatory makes the final commitment under the company's actual delegation.

Leave a route receipt

Record the route, allowed inputs, allowed actions, reviewer, stop conditions, and output status. ABA Formal Opinion 512 discusses confidentiality and competence in lawyers' generative-AI use; it does not endorse a particular triage scheme. [S2]

• Example receipt: 'Internal comparison authorized; no vendor contact; contract version D3; legal review pending.'
• A changed input, recipient, tool, or objective triggers re-triage.
• Retain the reason for exceptions and their approving owner.
• Measure whether requests arrive at the right owner, not merely how quickly the intake queue empties.

Scope and limitations

Proposed internal operating model. Risk categories are illustrative and are not a statutory classification or a declaration that any particular use is permitted.

No organization-specific policy, delegation of authority, data classification, or legal matter was reviewed.

Sources

1. NIST — AI Risk Management Framework — Voluntary AI risk-management framework; not a certification or legal safe harbor. Checked 2026-11-05
2. American Bar Association — Formal Opinion 512 (2024) — ABA interpretation of Model Rules for generative AI, including competence, confidentiality, supervision, communication, and fees; local law must be checked. Checked 2026-11-05

Explore further

This triage model provides concrete questions for workflow assessment with Agentic Attorney at https://agenticattorney.com/ . It does not authorize any tool or offer legal approval. Explore Agentic Attorney.