Honest. The month AI made our standups worse, and the two changes that fixed it.
We are a 7-person team across 4 timezones, so we have never done a live standup. Everything is async in Slack, written each morning when your day starts. In early 2026 we wired up a setup where everyone could auto-generate their update from Linear activity with one click. The pitch was obvious: nobody likes writing standups, the data is right there in the tickets, let the machine do it. For about a month, it did.
Then the reply rate quietly collapsed. The updates got longer (the model is happy to expand) and smoother and completely interchangeable. Every one opened with a clean list of yesterday's closed tickets and ended with a vague "continuing on the same today." They were accurate. They were also unreadable, in the specific sense that nobody read them. We only noticed because a teammate stayed blocked on a code review for three days, mentioned it in his AI update all three days, and not one person responded, because not one person was reading past the first identical line.
That was the lesson: the problem with AI standups is not accuracy, it is that perfectly accurate, perfectly bland updates are invisible. The machine had removed the friction of writing and, with it, the signal that a human cared enough to flag something. A standup is not a status log. It is a request for attention on the one or two things that need it. Automate the log and you delete the request.
We changed two things, and only two. First, the Blocker-First Rule: every update has to open with what is blocking you, named to the person who can unblock it, or with an explicit "no blockers" if there are none. The activity log got demoted to one line at the bottom. Second, the One-Human-Sentence Rule: after the AI compresses your tickets, you add a single line of judgment in your own words. We literally added a Slack reaction the lead drops on any update that is pure machine output with no human sentence, as a gentle nudge.
The numbers moved on the thing we cared about. Before the change, a named blocker got a same-day reply maybe half the time. After, it was the strong majority, because the blocker was now the first thing in the message instead of the last. Time to write barely changed; the AI still does the recall and compression, which is the part that was actually annoying. What changed was that the human kept the two jobs a human is there for: deciding what needs attention, and saying the one thing the model could not know to say.
Where we still do not use AI: anything that touches how a specific person is doing. A standup is about work, not about people, and the moment an update drifts toward "is this person keeping up," that is a private conversation, not a prompt. We use the manager-rollup workflow to spot who might be quietly stuck, and then a human goes and talks to them. AI points at the pattern. It never gets to write the message about the person. That line has not moved in 90 days and it is not going to.