Automated tests are completely useless for finding (let alone solving) human interaction issues. To compare them with telemetry is a category error.
Always-on metrics are massively higher quality data. They don't collect the same kind of data in many cases, but they can reveal a lot of things that never get reported. They also don't suffer from the well-established pattern of people not accurately reporting their own behavior when asked / polled (stronger when asking about future behavior, but it applies in all cases).
Automated Error reporting does has it uses.