The output was never the problem. The schedule and provenance were.
If it had told me it had skipped a week, I would have kept it.
Staff engineer, data platform · Rovna Logistics (400 staff, mid-market)
Why teams turn a scheduled agent off in month two. Ana ran a Monday metrics job for seven weeks, then disabled it.
The report was fine. The problem was that on Tuesday nobody could tell me whether it had actually run.
Every churned team so far describes the same gap: the output was trusted, the schedule was not. Ana checked manually for three weeks, decided that was the job she had been automating, and stopped.
I forwarded the first one to my director and she asked me to just paste it into the doc. So I was doing it by hand anyway.
The artifact was not the unit of sharing in her organisation — the weekly doc was. A correct output that cannot travel the last ten metres fails where no output fails. Second interview with this shape (see 03).
We turned it off when the numbers moved and nobody had changed anything. I could not see what it had read.
Trust broke on provenance, not accuracy — the figure was right and she still could not tell which query produced it. Three unprompted mentions in six interviews, and no question about it in the guide.
I set it up in about ten minutes. That part was genuinely great.Setup was never raised as friction in any interview.
I'd pay for it if it posted into the channel itself. That is the whole feature for me.Unprompted willingness to pay; the fourth.
Nobody on my team reads a link on a Monday. They read what is in front of them.Supports the finding at 21:40.
Put a run-status line on every scheduled artifact: last run, next run, skipped.
Kaisa Lehtinen · into the 18 September synthesisTest a plain-text export against interviews 7–9 before building for it.
Marco Bell · by 12 SeptemberAdd a direct provenance question to the guide; three unprompted mentions is a signal we are not asking about it.
Kaisa Lehtinen · before interview 7