How would you explain OpenTelemetry versus Application Insights for a .NET API?

In production I avoid a big-bang observability rewrite. I start with server requests, dependency calls, exception telemetry, and a small set of service-level metrics, then prove that dashboards and alerts support real incident questions. Service names, environment and version attributes, propagation format, sampling, redaction, retention, and exporter failure behavior are defined before every team invents them differently. I also watch telemetry overhead and duplicate instrumentation during migration. Tool choice matters less than a stable semantic contract that lets engineers move from a user symptom to the responsible service and dependency quickly and safely.