Small and medium-sized enterprises (SMEs) that sell through messaging platforms often handle customer enquiries manually, leading to fragmented conversations, delayed responses and inconsistent service. This article proposes a workflow-aware conversational support assistant that combines a large language model (LLM), retrieval grounding, merchant API tools and context-preserving human hand-off. Using a design-science-oriented literature synthesis, in which sources were identified through keyword searches of academic databases, screened for relevance to messaging-based SME support and LLM-agent design, and analysed thematically, the study derives three design priorities: grounding customer-facing claims in current merchant data, separating model reasoning from validated execution, and transferring accumulated conversational state when escalation occurs. The resulting five-layer architecture comprises messaging integration, agent orchestration, retrieval and grounding, store action, and escalation and hand-off. Volatile data such as price, inventory and order status are retrieved live; state-changing operations require explicit customer confirmation and schema validation before execution. The article also specifies an evaluation protocol based on scripted end-to-end customer journeys using journey completion, factual agreement, action correctness, escalation precision and recall, context preservation, and response latency. The contribution is architectural and methodological; implementation and empirical validation remain future work.
Read more