What AI may and may not do during legal intake

The safest legal-intake design begins with a boundary: the system can manage a conversation and a workflow, but the firm remains responsible for legal judgment, supervision and client decisions.

Appropriate operational tasks

A controlled system can greet a caller, identify language, give an approved disclosure, capture contact details, ask factual questions, follow a deterministic branch, schedule, attempt a transfer, create a task and write an organized record. It can repeat exact firm-approved office information when the source is current.

Decisions that stay with the firm

The system should not determine conflict, fault, liability, damages, case value, deadlines, strategy or whether representation exists. It should not accept a matter, reject a matter, negotiate a fee or sign a legal agreement for the firm. A caller’s request for those decisions should trigger the approved boundary response and, when appropriate, a human path.

Why a prompt is not enough

A prompt can tell a model to be careful. A controlled workflow enforces what state the conversation is in, which fields are allowed, which tool may run and what happens when the tool fails. For legal intake, the workflow should own the decision boundary and the model should own natural phrasing inside that boundary.

Supervision and evidence

Firms should approve the questions, disclosures, routing, follow-up and test scenarios. Providers should retain workflow versions, tool results, access history and incident records. Public claims should describe tested behavior, not a desired architecture.

This article is operational information, not legal advice. A firm should review professional-responsibility, privacy, recording and supervision requirements with its own counsel in each relevant jurisdiction.