JIT-SA: a governed engineering method for volatile AI-driven service composition

Classical Service-Oriented Architecture (SOA), serverless orchestration, and microservice design all assume that service topologies are largely defined before execution. This assumption becomes restrictive in AI-native settings where user intents, resource needs, and service combinations are volatile and cannot be fully pre-enumerated at design time. In this paper, we introduce Just-in-Time Software Architecture (JIT-SA) as a governed engineering method for volatile AI-driven service composition. JIT-SA treats architecture as a transient artifact synthesized from intent and context under explicit operational and governance constraints. We contribute (i) a formal intent-to-topology transformation model, (ii) a layered engineering method spanning semantic compilation, bounded topology synthesis, validation, and ephemeral deployment, and (iii) governance mechanisms including policy validation, budget guardrails, clarification dialogues, and deployment constraints. We instantiate the method in a prototype using LangGraph for stateful reasoning and Akash SDL for infrastructure realization, and evaluate it across three burst-workload scenarios. The evaluation shows that JIT-SA can synthesize valid executable topologies without invalid deployment and that it occupies a distinct operating regime: it is advantageous for rare, compute-intensive, latency-tolerant workloads, but inferior to serverless or persistent orchestration for predictable or latency-sensitive tasks. The contribution of the paper is therefore not a universal replacement for SOA, but a governed engineering paradigm and artifact set for volatile service composition in AI-native information systems.