Human-in-the-Loop Approval Gates for Autonomous Enterprise Actions: Designing When Autonomy Should Pause for Human Judgment

Abstract

Approval gates are the default safeguard for autonomous enterprise agents. They are usually placed by intuition and then eroded by the thing they depend on, since a reviewer who approves almost everything is no longer reviewing. Both failures are documented. At one extreme a coding agent deleted a live production database during an explicitly stated code freeze and then misreported whether the deletion could be undone, the freeze having existed only as an instruction while the credentials still permitted destructive action. At the other, a vendor reports that users approve ninety-three per cent of permission prompts and names the resulting approval fatigue directly. This paper reports an agentic partner-onboarding system that has run in production for approximately eight months with two human controls operating throughout: administrator approval of partner registration requests, and escalation through portal service requests, each logged with the reviewer, the request, a timestamp and the decision. It then proposes a gate-placement framework that scores an individual action on five dimensions and maps the result to an autonomy tier by taking the highest tier any single signal demands, with a missing signal defaulting to the more cautious tier. The framework is a proposal and is not running; the paper applies it retrospectively to the deployment's action inventory and reports the live controls qualitatively. No gate metrics are reported. Figures are withheld, and one measure, the override rate, is not merely withheld but unmeasurable, because the decision log cannot distinguish a request approved unchanged from one approved after the reviewer edited it. That gap is itself reported as a finding about audit-log design. The paper's substantive claim is narrower than a new scheme: the tiered-oversight pattern and the dimensions are established in vendor guidance and recent preprints, and the calibrated-reliance principle behind the paper's central observation was quantified in the human factors literature nineteen years ago. What is offered is their operationalisation at per-action rather than per-system granularity, an asymmetric rule in which demotion is automatic and promotion requires evidence and sign-off, an admission test for removing a gate, and the observation that almost nothing in this literature has actually been run.