Telemetry debt is defined as the accumulated gap between what systems emit and what engineers can actually use to make decisions. Unlike technical debt, it rarely gets addressed because it doesn't throw exceptions — it just inflates platform bills and on-call fatigue. The piece identifies seven distinct forms: instrumentation debt, dashboard debt, alert debt, metric debt, trace debt, log debt, and ownership debt. Each has different root causes and requires different fixes. AI workloads are dramatically accelerating the problem by introducing entirely new telemetry categories (prompt traces, GPU metrics, vector DB monitoring, agent traces) at a pace that compressed a decade of traditional debt accumulation into 18 months. A six-level observability maturity model is proposed, where most organizations stall at Level 2-3 (centralize/standardize) and never reach Level 4 (prioritize). A practical ROI framework for evaluating whether any given signal should be kept or deleted is outlined, along with technical recommendations (OpenTelemetry, adaptive sampling, eBPF, data tiering, observability-as-code) and organizational ones (explicit ownership, quarterly reviews, rewarding deletion over accumulation).
