Autonomous AI agents in enterprise systems: A review of architectural gaps

Abstract

A central challenge in enterprise information systems is that the development of generative artificial intelligence has moved from conversational assistants and document generators to autonomous agents that initiate actions․ Agentic architectures have the potential to transform workflows but their probabilistic reasoning is not compatible with transaction integrity‚ security boundaries and governance models that support mission critical enterprise functionality․ This systematic literature review of literature published in 2023-2025 explores four key themes of security and authorization‚ observability and traceability‚ governance and compliance‚ and resilience engineering․ By analyzing the state of the art of 25 peer-reviewed works‚ the review finds that autonomous AI agents are not treated as first-class governed execution entities by existing enterprise architecture frameworks․ We detect four gaps in the literature: a lack of standardized isolation of agent activity‚ a lack of governing models for authorizing agent actions with uncertain outcomes‚ a lack of best practices for tracing the decision chain of the agent's reasoning‚ and a lack of resilience patterns for transactions based on these AI-driven agents․ We conclude by proposing a theoretical research agenda for Agent-Oriented EA․ The review concludes that without discipline‚ organisations risk increased security breaches‚ regulatory non-compliance and systemic operational failures as agentic systems increasingly become the norm.