Human-in-the-Loop Control for High-Risk Legal Agents

Abstract

Agentic artificial intelligence can decompose legal objectives, retrieve authorities, invoke external tools, draft documents, communicate with stakeholders, and execute multi-step workflows. In high-risk legal settings, a final approval click is not an adequate control: errors can propagate across a trajectory, evidence can be withheld from the reviewer, confidential data can cross an unauthorized boundary, and a human may be asked to approve an artifact that no longer matches the version later executed. This paper proposes Human-in-the-Loop Control for Legal Agents (HILC-Legal), a riskadaptive governance architecture that separates language generation from authorization and separates authorization from execution. The architecture combines hard policy constraints, a calibrated action-risk score, evidence packets, competenceaware reviewer routing, version-bound authorization tokens, a least-privilege execution broker, provenance logging, and post-action monitoring. We define six control invariants for legally consequential agent behavior and a four-tier oversight policy ranging from bounded read-only assistance to human-in-command decision making. We also specify an evaluation protocol covering unsafe action escape, authority and evidence fidelity, confidentiality preservation, approval-version consistency, reviewer correction yield, automation bias, workload, calibration, and recovery. The proposal is a design and evaluation framework; it does not claim deployment results. Its central argument is that meaningful human oversight is an engineered control property - requiring information, competence, authority, time, and effective intervention mechanisms - rather than the mere presence of a person in a workflow.