Bounded Agentic AI: A Conceptual Design Framework and Measurement Proposal for Controlled Autonomy in Human-AI Work Systems

Agentic artificial intelligence shifts enterprise AI from systems that support human judgment toward systems that participate in organizational work by planning tasks, invoking tools, interacting with applications, and initiating actions. This preprint proposes bounded agentic AI asaconceptualdesignframeworkandmeasurementproposalforcontrolledautonomyinhuman-AI work systems. Building on prior research in AI governance, algorithmic accountability, human- AI interaction, socio-technical systems, runtime governance, bounded autonomy, enterprise architecture, and organizational control, the paper argues that model-level guardrails and generic human-in-the-loop designs do not sufficiently specify where enterprise agent autonomy is authorized, constrained, escalated, and made accountable. The framework defines eight autonomy boundaries: purpose, domain, data, decision, action, learning, escalation, and accountability. It further proposes a layered enterprise architecture control-plane model and the Agency Boundary Deviation Index (ABDI), a statistical construct for monitoring deviations from architecturally permitted behavior. A KYC onboarding workflow is used as an illustrative operationalization, not as empirical validation. The contribution is to make bounded agency explicit, measurable, and open to future empirical testing.