From Chatbot to Agent Fleet: A Deployment Pattern Taxonomy for Enterprise LLM Systems

Enterprise LLM deployment is evolving from single-model chatbots to multi-component agent fleets, but the literature that could guide this transition is split into two camps that rarely talk to each other. Architecture papers propose idealized target-state blueprints; industry surveys report aggregate adoption percentages. Neither connects a specific architectural pattern to the organizational conditions required to operate it safely in production. We synthesize evidence from recent empirical studies of production agent systems and industry research (Gartner, Menlo Ventures, McKinsey, Deloitte) into a six-pattern conceptual taxonomy of enterprise LLM deployments, ordered by the governance surface each pattern introduces: prompt-and-respond chatbots, RAG-augmented assistants, workflow automation agents, multi-agent orchestration systems, embedded copilots, and autonomous decision agents. This governance-surface ordering tracks architectural complexity for five of the six patterns; embedded copilots are the deliberate exception, placed by the distinct integration-lifecycle risk they introduce rather than by the complexity of the underlying model call, which we make explicit where the pattern is introduced. For each pattern we characterize the architecture, the failure modes reported in the cited literature, and the governance implications of each added architectural boundary. Read through the lens of environment interactivity, the degree to which an agent acts, observes the consequences of its actions, and adapts before a human intervenes, the six patterns also trace a progression in how much of the enterprise's operational context the agent is exposed to and permitted to change. This paper is not an environment-design study; we do not construct or evaluate a training environment. We do argue that the enterprise IT and governance context we describe functions as a form of interactive environment in its own right, one whose feedback and verification infrastructure, audit trails, approval gates, compliance checks, becomes the binding constraint as agent autonomy scales. Drawing on this synthesis, we argue that governance overhead compounds non-linearly as systems add components, and that organizations adopting multi-agent orchestration without first operating simpler patterns in production court predictable failure modes rooted in missing observability and audit infrastructure. We illustrate the taxonomy with a hypothetical deployment scenario. We also propose a Deployment Pattern Maturity Model (DPMM): a qualitative diagnostic heuristic, not a validated predictive model, that maps organizational readiness indicators to recommended deployment patterns and gives practitioners a structured way to have the readiness conversation.