Large Language Models (LLMs) are increasingly used across many software engineering activities, but their role in architectural refactoring remains largely unexplored. At the same time, decomposing monolithic systems into microservices continues to be a difficult and error-prone task, particularly when determining meaningful service boundaries.

This exploratory study investigates how LLMs can support architects through a multi-step decomposition process that applies seven summarisation strategies and four LLMs to generate architectural views and candidate decompositions. We assess these decompositions using six established metrics across four open-source systems that provide both monolithic and microservice versions, enabling comparison with reference architectures.

The results show that LLM-based decompositions outperform traditional techniques on several architectural quality metrics, including Cycle Independence (+15%), Data Transaction Purity (+8.5%), and Business Context Purity (+7.3%). However, established automated approaches remain markedly stronger in recovering the ground-truth service structure (MoJoFM +62%) and retain a modest advantage in modularity. This indicates that while LLMs can produce architecturally clean and domain-coherent service boundaries, they still struggle to reproduce the exact target decomposition achieved by specialised algorithms.