I kept seeing articles about a new study on AI adoption. Fortune ran one. Futurism ran one. I’d skim one before bed and then not sleep. These numbers confirmed what I’ve felt over the past two years as companies race to adopt AI to increase speed of delivery. The first one is anecdote. The second one is data. At Amazon I was taught that you needed both — and that when they conflict, trust the anecdote. This time they didn’t conflict. That’s what kept me up.
So I read the actual paper. Nearly 6,000 executives surveyed across the US, UK, Germany, and Australia. Not a paid online panel where bots answer questions about AI — central bank panels, the Fed, the Bank of England, that the researcher had spent a decade helping build. CFOs responding because the Federal Reserve was asking. The data was as clean as this kind of measurement gets.
Three out of four companies are using AI. The productivity gain so far is approximately zero.
I read that and knew I wanted to talk to the person behind it. His name is Nicholas Bloom. He’s a Stanford economist — William Eberle Professor, co-director of NBER’s Productivity, Innovation, and Entrepreneurship program. He built the survey methodology the Census Bureau uses to measure management quality. When serious people want to know what’s actually happening with AI in the economy, they look at his data.
He’s measuring the problem from the outside. I’ve been building inside it for two years. We were looking at the same thing from different angles, and I wanted to know if we were seeing the same thing.
We talked for thirty minutes.
When the electric motor was invented, it was obviously better than the steam engine. But factories had been designed around steam. Steam power ran off a single rotating shaft — one massive engine turning one pole, with belts and pulleys transmitting force to every machine in the building. Factories were tall and thin because the power had to travel down from the top through mechanical transmission.
The electric motor didn’t need any of that. Each machine could have its own motor. The optimal factory was a single flat slab. But getting there required tearing down the tall thin building and constructing something completely different.
That took thirty years.
Economists call this the J-Curve. A new technology initially makes things worse — you’re investing in change while the old system still runs — before the gains arrive. The gains don’t come from the technology. They come from the reorganization that the technology makes possible.
Bloom thinks AI is a J-Curve on steroids. Cloud computing, pioneered by AWS around 2010, has reached about 50% adoption in sixteen years. AI went from the launch of ChatGPT in November 2022 to 75% adoption in three years. The adoption speed is unlike anything in the economic record. But adoption isn’t reorganization. Adoption is plugging the electric motor into the tall thin building and wondering why the factory isn’t faster yet.
Everyone in the industry knows that actual AI use outpaces what any organization will officially report. People use personal devices. They copy results back into approved systems. The measurement captures what institutions permit, not what individuals do.
But if all that covert use were producing real gains, the gains would show up even with underreporting. They’re not. Nobody knows how to reorganize their work yet. Most people are using AI the way they used to use Google — a souped-up search bar. Faster answers to the same questions, plugged back into the same workflows, inside the same tall thin building.
The electric motor is running. It’s just powering the same old belts and pulleys.
One of Bloom’s students used Claude Code to build a complicated economic model — the kind of work that normally takes two months of careful coding. Done in days.
But, Bloom told me, you still have to make sure the model produces sensible results. That the assumptions aren’t absurd. That AI hasn’t introduced errors you’d have caught naturally during two months of slow, careful construction. The speed means you skip the checking that used to happen as a byproduct of the effort.
We share the same worry. Our names still go on anything we ship.
The execution collapsed. The verification didn’t. The student’s name is still on the paper. This is the same pattern I keep finding everywhere I look — in education, in enterprise software, in my own building process. Speed without verification is not progress. It’s exposure. And as a recent industry analysis documented, the pattern is showing up at production scale — AI agents deleting environments, bypassing guardrails, causing outages. The friction that slowed things down was also the friction that kept things working.
I came into the conversation expecting to have to make the case for judgment as the bottleneck. I didn’t.
When I described the pattern — that AI collapses execution costs, that everything then routes to the human for decisions, that the human’s judgment and attention become the binding constraint — Bloom’s response was immediate. He called it a bottleneck on management and judgment.
Then he added a prediction: that’s bad for new hires and good for management.
We started building on this together. Seniority correlates with judgment, but the correlation isn’t tight. I’ve worked with people who built entire careers around executing well — process, bureaucracy, institutional scaffolding — but whose authority was never built on judgment. When execution commoditizes, those people don’t suddenly become decision-makers. They become exposed.
Bloom picked this up and described his own work. Hours of meetings. Students presenting research. Him giving feedback. “That’s all judgment,” he said. And it is. But he’s a tenured Stanford professor whose entire career selected for exactly that skill. The question is what happens to the people whose careers selected for something else.
His World Management Survey — the instrument that made his reputation — measures whether firms have good process. Structured targets, performance monitoring, talent management. It does not measure whether firms decide well. AI can replicate process. Nobody has built the instrument for judgment quality.
The man measuring AI’s impact on the economy cannot yet measure the thing that matters most.
Here’s what I took from the conversation.
Bloom sees the J-Curve from thirty thousand feet. I see it from inside the building. We arrived at the same conclusion from completely different starting points: the technology works, the institutions haven’t reorganized, and the bottleneck is not capability but the human judgment required to use capability well.
The tall thin building is everywhere. Your company’s planning process. Your team’s review cycle. The way you write documents, approve decisions, allocate attention. All of it was designed for a world where execution was expensive and the structure existed to manage that expense.
The electric motor is here. It has been here for three years. The productivity gains are waiting on the other side of the redesign that almost nobody has started.
The question is not whether your organization uses AI. Seventy-five percent already do. The question is whether you’ve started tearing down the building.
Nicholas Bloom’s paper, “Firm Data on AI,” is NBER Working Paper 34836. I had the chance to discuss his findings in a wide-ranging conversation in March 2026, and he reviewed this essay before publication.
Thank you for your insights on AI. I am finding them fascinating as I try to understand AI and it’s impact moving forward.