Efficient performance testing for microservices is essential for engineers to ensure that deviations of performance and resource usage metrics from expectation are promptly identified within the rapid release cycles microservice development. To this aim, engineers would need to explore the space of possible workload configurations and focus only on the most critical ones, namely those that unexpectedly cause performance issues, such as low-load configurations. This requires a huge effort, and can be infeasible in short release cycles. We present CALLMIT, a framework using Large Language Models (LLM) enhanced by causal reasoning to automatically generate critical workloads for microservices performance testing. Engineers query CALLMIT to generate workload configurations expected to expose deviations from performance requirements, so as to actually run only tests that trigger critical configurations. We present the experimental evaluation on three subjects, with comparison to a conventional Retrieval-Augmented Generation technique. The results show that causal models improve the correct identification by LLM of performance-critical workload configurations.