Law firm intake without pretending the model is a lawyer

Firms buy 'legal AI' and then ask it to summarize case law. That can be useful later. It is not the bottleneck that costs most small and mid-size practices money every week. Intake is.

A potential client sends a PDF or fills a form. Staff retype names, dates, and opposing parties into the matter system. Conflict checks happen late. The good file has already hired someone else.

An operations agent can extract parties, jurisdiction, dates, and the requested work. It can flag missing conflict information and draft a receipt: we have the file, a lawyer will review, here is what we still need. That is not advice. It is logistics.

Hard stops matter. Anything that looks like a legal opinion, a fee quote outside an approved range, or an instruction to destroy documents should queue for a person. The audit trail should show what the model saw and what it proposed.

Confidentiality is a design constraint, not a footer. Use approved accounts, the minimum necessary data, and no training of public models on client documents. If that cannot be guaranteed, do not start.

Measure time from first contact to a complete intake file, lag before a conflict check, and matters opened without a second round of retyping. Those are operations metrics. Win rates still belong to the lawyers.