A clerk’s filings agent can draft submissions but must not file documents without a human. Which responsible AI pattern fits?

A clerk’s filings agent that may draft but must not file without a human should use oversight mode that requires staff approval before executing filing actions. The pattern works because tool-calling agents can separate proposal from execution: the agent prepares content and a proposed tool call, then a clerk reviews and approves before the write reaches the court filing system. Human oversight is the responsible-AI control that keeps automation assistive for high-consequence judicial workflows.

Audit logging that records staff reviews after the agent executes filing actions fails by concept because it provides retrospective visibility rather than a prospective approval gate, leaving the agent able to execute filings without a human check and documenting the consequential legal act only after it has reached the filing system. Evaluation harness that scores test filings before the agent executes filing actions fails because it measures performance on test cases rather than enforcing a live human approval, so a passing score does not prevent the agent from executing live filings without a clerk's approval and provides no required human gate on the consequential filing action. Oversight mode that allows staff review after the agent executes filing actions fails because post-execution review comes too late to block the filing, so it allows the agent to execute the consequential filing action without prior approval and does not ensure a clerk reviews and approves before the write reaches the court filing system.

Exam caveat: oversight UX must prevent "approve all" rubber-stamping under backlog pressure, or the human gate becomes theatrical. Operational check: configure filing tools as approval-required, deny direct delete/write without oversight, test that unapproved calls do not mutate the filing system, and log both agent proposals and clerk decisions for audit.