By mid-2026 I'd built a wide fleet of scheduled AI routines, briefings, outreach, content, audits. I ordered a full review of every one of them against what it actually still did, and then a second, adversarial pass that challenged some of the first pass's calls.
I killed or paused several routines outright, even ones already built and running: a social media growth bot, a dead outreach directory, a personal daily synthesis run that had never produced real output. I also refused to let a brand new weekly routine go straight onto a schedule. Kept it as a manual command until it proved itself by hand.
My own operating rule is automate last, and only add automation back for the small slice you actually miss doing by hand. A routine earns its place by being used, not by existing, and wiring a brand new process straight into a cron would repeat the exact automate first mistake I just named. Anything that could reach the team automatically also had to pass through me first. I'm the approval gate. Never an auto-broadcast.
The live routine count dropped. We stripped dead integrations and channels out of the survivors, folded a redundant weekly audit into a simpler daily check, and the new weekly review became a manual ritual first, proven by hand for two weeks before it was ever allowed to run on its own.
If a routine can't point to something it actually caught or delivered, it doesn't matter how clever it is. Cut it.
A system doesn't get to keep running just because it exists, and I stayed the one who approves anything before it reaches the team.