Get ruthless about what's worth building

AI has made me faster at almost everything I do, but I’ve never been busier.

A job that took three hours now takes ten minutes. By any sensible accounting I should be swimming in free time. Instead, every task that collapses opens a door, and behind the door are ten things worth doing that were not on the table six months ago. The work did not shrink. It changed shape.

More of my week now goes on a question that barely existed before. Is this worth doing at all. Worth the tokens, worth the attention, worth the door it will open in turn. AI did not remove the hard part of my job. It moved it to the one place it cannot follow.

The most expensive way to learn this

Mark Zuckerberg appears to have discovered the same thing at considerably greater cost. At an internal town hall on 2 July, in a recording heard by Reuters, he told Meta employees that AI agent development over the previous four months “hasn’t really accelerated in the way that we expected”. This came six weeks after a layoff memo that framed AI as the defining technology of a generation.

Meta swapped people for capacity, 8,000 roles out, more than $100 billion of infrastructure in, and the candid admission is that the trade has not produced acceleration. That is not primarily a story about model capability. It is a story about where the constraint sits. Execution can be bought. The judgment that decides what all that execution is for cannot, and no amount of capex moves a human constraint one inch.

The workforce data says the same thing from the bottom up. BCG surveyed nearly 12,000 frontline employees this year and found that 42 percent save the equivalent of a full working day every week through regular AI use. Two thirds of them received little or no guidance on what to do with that day, and half admit the recovered time is not going into more valuable work. An eight-month ethnography at a US technology company, run by Berkeley Haas researchers, found that AI did not free up time at all. It expanded what people felt capable of taking on, so they worked faster, across a broader scope, into more hours of the day, without anyone asking them to.

The saved day is real. Almost nobody has decided what it is for. That is not a productivity problem. It is a judgment problem wearing a productivity costume.

The forty minutes that will not compress

I can watch this happening inside a single system we run.

This blog is produced by an agentic pipeline. One agent scans our repositories to understand what we have been building, searches the news for stories that connect to that work, and proposes three posts, twice a week. I pick one, or none. A second agent researches the chosen angle in depth and produces a brief. Then I sit down for a spoken session with a third agent, and we argue. We test the angle, fix the narrative, ground every claim in something we have actually done, and choose the title. That session produces a specification, a fourth agent drafts from it, and I edit the draft into final shape.

Before this pipeline existed, a post cost me two or three days. Nobody sensibly spends two or three days a week on a blog while running an engineering company, which is precisely the point. This system only became worth building because execution got cheap. It is one of the ten things behind the door.

Now watch which parts have compressed since we started. The final edit used to take an hour. The agents have accumulated a rich history of my writing, and it rarely takes more than fifteen minutes now. The research keeps getting better. The drafts keep getting closer. Every mechanical stage of the pipeline improves on its own.

The forty-minute discussion has not compressed at all, and it never will, because it is not execution. It is where the post takes its shape, where angles get tested, claims get grounded in what we have actually done, and the agent’s suggestions get overruled by an opinion only I can hold. It is the one stage of the pipeline that has to stay human, because it is where the value is added. And the deciding starts earlier still. Four or five of every six proposed posts die at selection, before I spend a single research token on them. The pipeline could sustain six posts a week. We publish one or two, because my judgment, not the machinery, sets the rate. The execution gets cheaper every month. The deciding has not shifted by a minute.

Yes, it rhymes with Jevons. That is not the lesson.

The obvious objection is that this is just Jevons paradox, cheap capacity gets consumed by rising demand, nothing new. Half right. Jevons tells you the saved time will be spent. It does not tell you who decides how, and that is the part that changed. The scarce resource has moved from hands to judgment, and most organisations are still managing, measuring and buying for the old bottleneck. The same answer covers the sharper objection, that Meta’s disappointment proves execution never got cheap. Zuckerberg’s frustration is with frontier agent autonomy, the far end of the curve. At the level of tasks, where the actual work lives, the savings are large, measured and boringly consistent. What Meta could not buy was the deciding.

Software development shows the same pattern from the inside. When code generation accelerated, the constraint moved downstream to review, and teams drowned in pull requests. That bottleneck yields to process. Structural annotation, deep test suites, automated review, security audits and pen testing remove it, and we have done exactly that. The stage that did not get quicker is the specification. If anything, we spend more time on specs than we ever have, and the more time we spend there, the faster everything downstream moves. The judgment is not the drag on the pipeline. It is the accelerant. You can automate the review of code. You cannot automate the decision about what the code should be, or whether it should exist at all.

We have made that decision at full scale before. We built the analytics product we had aspired to for twenty-four years, in two weeks, and then chose not to lead with it, because a two-week build is a category anyone can enter. The pipeline makes the same call twice a week at smaller stakes. The principle does not change with the size of the decision.

Measure the kill rate

We move faster than we ever have, and yet we have never been busier. What is important is what that busyness is made of. AI is absorbing the execution of white collar work, the code, the contract review, the research. What it cannot automate, and never should, is judgment. That is what humans bring to the table, and it is the only durable answer to where you now add value. The honest audit is not whether your team feels less busy. It is whether the hours AI hands back are flowing into judgment, into deciding what deserves to exist, or being swallowed by more execution by default.

Those that succeed the most in the next few years will not be the ones with the most AI. They will be the ones who made the best decisions about what’s worth building. Cheap execution puts ten worthwhile things behind every door it opens, and the discipline that matters now is choosing, out loud, on purpose, the nine that stay undone.

Four or five of every six ideas my own pipeline proposes die at my desk each week, and the blog is better for every one of them. So the question worth sitting with is not how much time AI has saved you. It is this. What did you stop doing, once doing it got easy?