Distributed tracing is widely used to debug microservices, yet most workflows either inspect single slow traces or compute aggregate percentiles without clearly identifying which component is responsible for tail latency. We study whether trace data alone—without node, network, or profiler access—can reliably attribute sporadic, random slowdowns to specific components in a service chain. Our approach treats each span’s latency as a feature and the end-to-end latency as an outcome, combining (i) conditional tail analysis that focuses on the slowest traces, (ii) variance-normalized influence scores that account for service fan-out and concurrency, and (iii) lightweight causal tests using trace-local timing only. We evaluate across bursty and steady loads, with confounders such as autoscaling, Garbage Collection (GC) pauses, and downstream dependency jitter. Experimental results demonstrate that the proposed method identifies the primary slowdown contributor with high precision. The method remains effective even in the absence of explicit error logs. That part stands out because it shifts tracing away from just checking after the fact. Now its more like attributing the root cause online, in near real time. It helps fixing issues sooner and also makes enforcing Service-Level Objectives (SLOs) fairer, and rollouts safer in multi-tenant clusters.