Stop asking teams for a date. Ask for a range, and what would move it.
A single date sounds precise. Usually it's a guess with a confident face. Then everyone treats it as a promise, and the team spends the next month defending it instead of finishing.
What I ask for instead:
1. A range. "Mid to late March" is more honest than "March 14."
2. The two or three things that would push it to the late end.
3. The day the range gets tighter, and who will tell me.
Then I report the range upward, with the same assumptions attached.
Execs don't hate ranges. They hate surprises. A range with named risks gives them something to plan around. A single date just moves the surprise to later.
What do you ask for when someone wants "the date"?
@ConnorAbene Delivery has its own version of this. The status report stays green while cycle time stretches and the defect backlog climbs. By the time it turns amber, the team already knew for a month.
@buckleyplanet Dated, with a reason, is what keeps that one revision honest instead of quietly moving the goalposts. Who signs off on it in your pilots, the process owner or a steering group?
@staysaasy The tell is when you ask for one concrete example and it never comes. I've started asking "what would they want us to do differently?" Real team feedback usually comes with an ask. A cover story rarely does.
@gregorojstersek The hard part is the stepping back. When their first approach isn't the one I'd have picked, I ask what risk they're solving for instead of steering. If the context was clear, they usually saw something I missed.
@erickimberling Not too much control, but often too little visibility. What I've seen work: put the regulatory dates on the vendor roadmap review as named milestones, and keep a fallback plan you own if they slip. You can outsource the build, not the deadline.
@buckleyplanet Two sentences in the doc beats anyone's memory of kickoff. Do you hold that definition fixed for the whole pilot, or allow one revision if the first results show it was wrong?
Hypercare that ends on a calendar date isn't over.
It just hands the open incidents to a support team that wasn't in the room.
What has to be true before your team calls hypercare done?
The jump from measuring to controlling the trend is where most teams stall. I've seen dashboards with thousands of violations that nobody owned. What worked was one rule: no new violations in the code you touched this sprint, and the old backlog gets burned down on purpose, not by heroics.
@NilsSmith Naming what can wait is the half that gets skipped. I put it in the same message: "This is due Thursday, so the report slides to Monday, and I'll tell the sponsor." That last part matters. People won't drop something they think they'll be blamed for.
@Scrumdotorg In practice it means protecting the conversations speed squeezes out first: refinement where someone asks why we're building this, and retros where the quiet engineer names the risk. When AI shortens the build, I keep those on the calendar before I chase velocity.
@buckleyplanet That's the right split: the people who live with the result define the miss, and the team running the AI just measures against it. Do you get that owner to sign it off in writing, or is agreeing at kickoff enough?
@aakashgupta The fix that works for me is boring: one written definition per metric, with a named owner and a source table. Customer as of which date, active or billed, which system wins. Without that, faster dashboards just move the meeting from deciding to reconciling.
In January almost every PR at Linear had a second human reviewer. By October, 53% didn't. They didn't drop review. They decided where it goes.
Most teams are dropping it by accident.
On Nov 16, Linear said 53% of its PRs now ship with only the author plus an AI reviewer, up from almost none in January. They "became more deliberate about where human attention goes."
That's the part worth copying. Not the 53%. The decision.
A change can still pass review while the team only partly understands it. "Reviewed" is not the same as "we can explain this, and we know how to undo it."
Don't let coverage drift. Write it as a one-page rule:
1. Changes that touch money, customer data, access or security, or anything hard to roll back still get a second human.
2. Other changes can ship with the author plus an AI review if tests and checks pass.
3. Add one line to Done: the owner can explain what changed and how to undo it.
Take the rule to your next steering meeting so the risk call is made in the open, not by default.
Where does a second human still have to look on your team, and where have you never actually decided?
@rohanpaul_ai Already seeing that in delivery plans. Once code got cheap, the queue moved to review and UAT, and our estimates broke because they sized the build, not the proof. Now we size each item by how long it takes to verify, and name who signs off before work starts.
@asana@ArBose@AnthropicAI Clear roles is the part most teams skip with agents. The question that comes up first on my programs is who owns the output when an agent's change breaks something in UAT. Until that's written down, "teammate" is just a nicer word for tool.
@buckleyplanet Right, and without an end date the parallel run never really ends, it just gets expensive. Who signs off on the miss definition in the pilots you've seen?
@gregorojstersek You see the gap in the first incident. A builder re-prompts until the error goes away; an engineer can explain why it broke. We started asking for that explanation in review before AI-written changes merge.
Every "small change" request should come with one question: what are we dropping to make room?
Small changes aren't small. They're unpriced.
What's your go-to line when someone says "it's just a quick tweak"?
@erickimberling Helped most in build and test: short config sprints with the business in the room caught fit gaps early. Hurt when we tried to sprint data migration and cutover. Those need a fixed runbook and dress rehearsals, not a backlog.