A veto only catches what you can already see

You’ve got a recurring meeting you’re pretty sure is overhead. A status sync, a planning review, a standing check-in. AI can run it now, and the pitch for letting it is clean. As long as every action gets logged and a person can veto anything before it takes effect, automating the meeting is safe. You keep the record. You keep the override. What’s left to lose?

That pitch has a working example. Linear is the project-tracking tool a lot of software teams run their work in. This summer it shipped Loops, small automations that watch for a change and then act on their own. When a project’s target date slips, a Loop updates the launch plan, files new issues against the new timeline, and posts a Slack message spelling out what changed, why, and who needs to act [verified]. Every step is recorded. A person can object. It has been running in production since July, on the paid tiers [verified]. This is the safe-by-audit design made real, and it works.

There’s a catch the log can’t catch, and it isn’t a flaw in the logging.

A log proves who decided what and lets you undo it. It can’t surface a disagreement nobody has said out loud yet. Automating a meeting is safe when the meeting only moved around information people already had, and unsafe when the meeting existed to drag a hidden conflict into the open.

Go back to that slipped date. The Loop treats it as a scheduling fact. The date moved, so update the plan and tell everyone. But why did it slip? Often because two teams were holding incompatible assumptions about the scope, and the calendar is where that finally showed. In the old meeting, the slip forces a conversation. Someone says the date was never realistic. Someone else says the requirement was smaller than that. In the back and forth, an assumption nobody knew they disagreed about gets said out loud for the first time. The Loop skips the conversation. It records the symptom, the date, smooths it into an updated document and a notification, and the disagreement that caused the slip stays buried.

This is John Cutler’s argument, made in September. Cutler writes TBM, a widely read newsletter on how product teams actually function, and he’s worth reading because he digs for the structural cause under a surface complaint. His September 24 piece, “AI, the Loss of Positive Friction, and What to Do About It,” lands on a claim that sounds backwards at first. Some of the friction AI is busy removing was doing real work [reported]. When design, engineering, and product have to negotiate a requirement, the negotiation itself exposes conflicting mental models and unshared assumptions. Take the negotiation away and the misalignment doesn’t leave with it. It gets baked silently into whatever the automation ships [reported].

That’s where the veto stops being the safety net it looks like. A veto only works on what you can see. Async review shows you the proposal and the reason attached to it. It doesn’t show you the assumption you and the proposer don’t share, because neither of you has said it, and saying it was the entire point of the meeting you automated. So you skim a clean Slack summary, nothing looks wrong, and you don’t object. The Loop runs. The paper trail is perfect. The misalignment ships anyway.

Now the strongest reason to think none of this is a problem, because it’s real. The two ideas never actually touch. Linear’s Loops automate coordination: date changes, document edits, status. They don’t run your retrospective. The careful version of the automation case agrees, and it singles out the retrospective as the one ceremony that shouldn’t compress, since teams that skip it reliably watch velocity fall and technical debt climb within a few sprints [verified]. So the tidy split holds. Automate the coordinating meetings, protect the judgment ones, and Cutler is describing the second pile. No collision.

The split holds for the examples and breaks on the rule. The automation case doesn’t say coordination meetings can be automated when they’re logged. It says governance meetings can, and it treats audit structure as the whole question of whether that’s safe. Governance is the big bucket. It includes the meetings where conflict is supposed to surface. Stated that broadly, the rule licenses automating exactly the meetings Cutler is warning you about.

The scope guard makes it worse, because it guards the wrong thing. The instinct is to automate the small, low-stakes stuff and hand-review the big calls. But a slipped launch date is small and low-stakes by that measure, and it’s precisely where a hidden disagreement hides best, because nobody calls an emergency meeting over a calendar update. The size of a decision tells you nothing about whether it was carrying a buried conflict.

Then there’s what this does to the retrospective, the one meeting the careful version fought hardest to keep. It saved the right meeting for the wrong reason. The case protects it by calling it governance for the execution loop. But look at what a retrospective is for. It exists to make visible what no single person saw while the work was happening. That’s a surfacing meeting, the exact kind Cutler says a log can’t stand in for. So the case already shields one surfacing meeting. It shields that one by accident, and the rule it states would put every other meeting like it on a Loop.

Hold this at its real weight. Cutler’s argument is one essay, and finding hidden structure is his signature move, the thing he does to every subject. A writer who always finds a deeper cause will sometimes find one that isn’t load-bearing, and the multi-company comparison that would test this claim hasn’t been run. He says so himself. So this isn’t proof that auditable automation is unsafe. The Linear facts are solid, and so is the skipped-retrospective data. The dispute is over how to read them. What it forces is smaller and firmer than a reversal. It’s a scope correction. “Automate it if it’s logged and vettable” is true for meetings that pass information around, and unproven, maybe wrong, for meetings that exist to make people argue.

So before you put a ceremony on autopilot, the audit question isn’t the one that decides it. The one that does is quieter. What was this meeting actually for? If it moved information people already carried in their heads, automate it, log it, and the veto will catch what matters. If it existed to force a disagreement into the open, the log will faithfully record a decision that was never really contested, and the veto nobody thought to cast won’t save you. There’s a clean test in there, and it’s worth running before the next Loop goes live. Could the disagreement have shown up in the written proposal? If it could, review catches it. If it only ever comes out when people argue, you can’t automate the arguing away and keep what the arguing was for.