Bounding the Machine: A Cross-Case Study of Governed Autonomy in Five Enterprise Agentic Systems

Abstract

Agentic systems built on large language models (LLMs) are entering regulated enterprise work faster than the theory that would tell us how their autonomy should be bounded. The presented study addresses that gap through an interpretive cross-case analysis of five agentic applications the author designed and built for a regulated financial setting - an orchestration router, an incident-intelligence engine, a programme-status summariser, an event-driven communications generator and a multi-agent document-compliance checker. Treating each artefact and its design rationale as a case, and reading the design decisions through a Gioia-style inductive coding, the study asks a single question: across systems built by one architect for one class of setting, what recurring moves bound machine autonomy. Five dimensions emerge - a deterministic floor beneath every probabilistic component, provenance made inspectable before the fact, human checkpoints placed at consequence rather than at convenience, scope confined by construction, and reversibility engineered into outputs. Together they compose a framework of governed autonomy that reframes the enterprise question from how capable the agent is to how tightly its freedom is drawn. The account contributes an empirically grounded vocabulary for a design problem the field has mostly discussed in the abstract.