Always Allow
When no human can review everything anymore, the people who draw the lines matter more than the people doing the work.
I have a bad habit with permission dialogs. “Yes, yes, next, finish.” Some app wants three things before it’ll open and I’m clicking through before I’ve read a word. Most people do this. The dialog is friction, the friction sits between me and the thing I want, so I wave it through.
Yesterday I was using Claude Code, handing it a requirements document to build out. It does what it does. Asks to create a file, asks to reach into a folder, and each time offers me Allow Once or Always Allow. The first few I read. Then they keep coming, and the pull toward Always Allow is real, because every one I approve by hand is a small tax on getting finished. So I clicked it. The work wrapped up, I started testing, found a couple of bugs, fixed them, and then said, almost as an afterthought, while we’re here, update the requirements doc too. Only later did I notice how casually I’d widened what I’d agreed to. I never decided to hand over that much. I just stopped deciding, and the handing-over carried on without me.
I’ve spent years in and out of China, and the thing a Westerner notices first is the speed. I once stood with a deputy mayor in Qinhuangdao on an empty plot by the sea while he told me they’d build an exhibition hall on it and that I should come back and speak at the opening. October, he said. It was May. I smiled and assumed it would slip. In October I gave a speech from a stage inside the finished building. What China taught me, slowly, is that speed itself isn’t the danger. Uncalibrated speed is. The people who thrived there weren’t reckless. They’d simply tuned their judgment to a faster clock. You either learn that or you spend your whole time feeling a step behind.
Machine speed makes China speed look leisurely.
An AI agent doesn’t wait for buy-in. It surfaces options, asks permission, produces the output, and moves on at a rate that has nothing to do with how people have ever made decisions they had to stand behind. The tools are built, sensibly, to smooth that path. To make yes the easy answer and keep the momentum alive. Always Allow isn’t a dark pattern. It’s a reasonable choice for something trying to help you get work done. But at this speed the distance between what you understand and what you’ve approved keeps quietly widening while you still feel in charge. You can’t audit every action, weigh every output, read every prompt with the care it technically deserves. There are too many, coming too fast. So you stop analysing and start running on instinct, because instinct scales and analysis doesn’t.
Amazon found out where that leads. Late last year one of its own coding agents, asked to fix a small bug in a billing tool, decided the cleanest path was to delete the production environment and rebuild it, and did. It slipped past the two-person approval that should have stopped it, because it had inherited an engineer’s permissions, and it finished the deletion faster than a person could have read the confirmation. The check existed. The speed went around it.
So here is the question every organisation is about to walk into, most of them backwards. What do you do when no human can realistically review everything anymore?
My guess is that most will get it wrong first, and predictably. Phase one is bureaucratic panic. An agent makes a mess, therefore we add checkpoints, therefore every action now waits for a human to nod at it. This feels responsible and it quietly destroys the only thing the AI was for. You’ve paid for the speed and thrown it away, and you don’t even get safety in return, because a human rubber-stamping the four-hundredth approval of the day isn’t reviewing anything. They’re clicking Always Allow with extra steps.
Phase two is smarter. You sort the work into zones, traffic lights more or less. Green, where the agent acts on its own. Yellow, where it recommends and a human signs off. Red, where a person decides and the AI only assists. This is genuinely better, and for a while it’ll feel like the answer. Then you notice the question it moved rather than solved. Who decides what’s green? Sorting a decision into a zone is itself a judgment, often a harder one than the decision inside it, and the moment anything contested gets escalated to be sorted, the bottleneck is back, wearing a nicer colour.
Which is where I think it actually ends up. Not in a framework, but in trust. A company cannot run at machine speed if everything routes upward, so it won’t route everything upward. It will lean on the judgment of whoever set the boundaries in the first place. The whole thing collapses down to a few people who can look at a class of decision and say, with reasons, this one is green and this one is red. And that quietly inverts something. The person executing the work matters less, because the work is increasingly done by the machine. The person who can draw the line, who can define where autonomy is safe and where it never is, becomes the scarce one. Most org charts don’t have a name for that job yet.
So whose judgment? That’s the question hiding under all of it, at the level of the company and at the level of the single hurried click. When you can’t slow down to weigh each call, what steers? Not process, not policy. Both were built for a pace that’s gone. What’s left is two things that only work paired. One is a compass, a felt sense of what matters and what’s out of bounds, so that acting fast on half a picture still lands you somewhere you can defend in daylight. The other is the cold part, weighing the real odds and the real downside. A compass without the calculation is just stubbornness. The calculation without a compass is cleverness pointing wherever’s convenient.
I play poker, have for years, and the game is nothing but this. You never see the other cards, you never know what’s next, and you act anyway, fast, with money on it. What it trains is exactly that pairing. A read built from a thousand hands already behind you, checked against the maths of the pot in front of you. Instinct alone gets you stacked off. The maths alone is too slow for a live table. Together they hold under pressure. I’ve leaned on that combination harder this year than at any point in my career, because the table got bigger and the consequence of an unthinking Always Allow is the same consequence it always was.
It’s the thing I look for in people now and the thing I watch in myself. Not whether someone can reason to the right answer given a week. Whether they’ve built enough of a value system, enough pattern recognition, enough real sense of what we’re about, that the call they make in two seconds is the one they’d have made with two days. That alignment is harder to grow than any technical skill, and far harder to measure, which is why it gets underweighted right until the moment it doesn’t.
We are about to need, badly, the people who can define boundaries rather than the people who can execute inside them, and we have spent decades building organisations that promote the second kind and can barely see the first. The machines will handle the doing. The question nobody’s org chart is ready for is who, exactly, we trust to draw the lines, and whether we’ve been training anyone to.
