On July 13 the Stanford Digital Economy Lab released a statement running to eighty-eight words. It carried the title We Must Act Now , organized by Erik Brynjolfsson, Director of the Stanford Digital Economy Lab and co-founder of Workhelix, with Ajay Agrawal, Anton Korinek and Tom Cunningham.

The statement is austere. AI may become radically more powerful over the next decade, producing an economic transformation larger than the Industrial Revolution on a far shorter timeline, with large-scale job displacement alongside major gains in living standards. Then the instruction: act now to build the incentives and institutions that steer AI toward complementing people rather than merely imitating them.

The document is short. The signature list is the argument.

It launched with more than two hundred economists and AI researchers, sixteen of them Nobel laureates, and is now approaching two thousand signatures. A few things about the names matter.

The skeptics signed. Daron Acemoglu and Simon Johnson, who took the 2024 Nobel, put their names to a statement that says the transformation could exceed the Industrial Revolution — this after Acemoglu's own work put AI's total factor productivity gains at well under one percent over a decade. When one of the field's most rigorous skeptics signs the alarm, the disagreement is no longer about whether.

The coalition does not line up politically, which is the point. Krugman, Ferguson and Cowen do not agree about much. Jeff Dean at Google, Jack Clark at Anthropic and the chief economists of OpenAI and Anthropic have signed a warning about the thing their own companies are building. So has Yoshua Bengio, who has spent recent years warning that the risks need governing. Getting that group onto the same eighty-eight words means the claim is not political.

I sat down with Brynjolfsson recently to understand what sits behind those eighty-eight words. He did not reach for nuance. "There is a tsunami coming at us of technical capabilities," he said. "And we're not prepared in terms of the organizational changes."

"When I first started talking about these issues years ago, I felt like a bit of a lonely voice," he told me. Now the skeptics are on board, and he singled out the last three or four months. He mentioned, almost in passing, that the laureate count had moved again, to seventeen, because one had emailed after publication asking to be added.

The profession has stopped arguing about whether this is happening. That is the part that dates a leader's own hesitation.

Statements signed by economists get read as macroeconomic commentary, which is probably how most executives filed this one when it hit the news. But steering AI toward complementing people rather than imitating them is not a regulatory lever. There is no bill that does it. It is a capital allocation decision, made repeatedly, in business cases and investment committees and the specific choice about which line on a P&L an AI project is expected to move.

The economists have named the outcome. The mechanism for producing it sits inside companies, and there are three things the economics is now saying clearly that are mostly invisible from an ordinary executive dashboard.

One: You Are Measuring The Wrong Side Of The Ledger

GDP measures what is bought and sold. With few exceptions, as Brynjolfsson put it to me, if something has zero price it has zero weight in the official statistics.

His correction is GDP-B (the ‘B’ stands for Benefits), built with Avinash Collis, Erwin Diewert, Felix Eggers and Kevin Fox. It inverts the question. Not what would you pay for this, but what would we have to pay you to give it up.

His team has now run the method on AI itself. Working with Avinash Collis, Felix Eggers, Sophia Kazinnik and David Nguyen, he asked people what they would need to be paid to give up AI chatbots for a month. The average answer was about $124. Most of them pay $20, or nothing. Aggregated across 115 million American adults, the team puts the total consumer surplus at roughly $172 billion a year — value received and never transacted, comfortably more than the AI industry collected in American revenue over the same period.

Most of what this technology produced last year is invisible not only to the companies buying it but to the companies selling it. And that is a consumer measurement, not a corporate one. Nobody has run the equivalent exercise inside a company.

Now bring that inside a company. A firm counts what it charged and what it paid. When AI makes an analyst three times faster, no transaction occurs. When it stops a support case escalating, the value shows up as an absence, and absences are not recorded.

I have run into the consequence often enough to expect it. Ask a team using these tools what has changed and the answer is emphatic. Ask finance about the same period and the answer is very little. Early on I would have assumed one side was overstating. After three decades of measurement work I treat that contradiction as information about the instrument rather than about either group.

Both accounts are usually accurate. Only one has a form to be entered on.

Two: Your Instrument Cannot Tell Adoption From Transformation

The second blind spot runs the opposite way, and it flatters you. Organizations count usage and treat all usage as equivalent.

Gallup, where I work, finds the most common AI applications are writing and editing at 51%, search and research at 49%, and general assistance at 39%. These are real conveniences. They are not transformative in any sense a shareholder would recognize. They make existing tasks pleasanter, and existing tasks were rarely the constraint.

Which is why the breadth finding matters more than the adoption headline. Among employees using AI for one or two purposes, 45% say it has had a positive impact on their productivity. Among those using it for seven or more, 90% do. Gallup is careful to note the correlation does not establish cause, and the arrow may run both ways. But whichever direction it points, the variable moving alongside returns is not access. It is range.

So the criteria for judging AI success have to change, and the questions should get harder as they go.

Did it make an existing task faster? The measured gains here are real and sometimes large. They are also the category most likely to vanish, because the saving gets distributed into everyone's week rather than landing anywhere finance would find it.

Did it create a capability that did not exist before? Something newly visible, competence in someone who lacked it, a decision the organization could not previously make. This is where the value actually is.

Can you put a number on that capability? Most organizations stop here, because the honest answer is no. Changing that answer is the entire discipline Brynjolfsson has spent thirty years building.

Does the number reach the ledger? Almost nobody passes this one. A figure that lives in a slide deck is not on the books. Until it sits somewhere capital gets allocated, it will lose every argument to the payroll line.

An honest portfolio review runs every deployment down that ladder and notes where it stops.

Three: Cost Removed Has A Number. Capability Created Does Not

The third is about what the statement actually asks for: AI that complements people rather than imitating them. That is not a moral preference, and it is not new. Brynjolfsson named it four years ago.

In a 2022 essay called The Turing Trap he argued that Turing's imitation game made human-likeness the goal of the field and has kept it there for seventy years. As machines become better substitutes for human labor, he wrote, workers lose bargaining power and grow more dependent on whoever controls the technology. His diagnosis of why it happens anyway names the audience directly: there are excess incentives for automation rather than augmentation among technologists, business executives and policymakers.

Ask why so much AI investment aims at reducing headcount rather than expanding what people can do. The honest answer is not that executives believe substitution produces better returns. The honest answer is that substitution produces a number. A reduced payroll line books immediately, appears in the quarter and requires no interpretation. Augmentation produces capability, faster judgment and more surplus in customers' hands, almost none of which carries a price.

Ram Charan put the same point to me recently in blunter terms. Companies are delayering for cost, he said, not for capability.

The distinction matters because the two produce different organizations. Strip out a layer to take out expense and the remaining layers do the same work with less help. Remove one because decisions can now be made closer to the work, and you have redesigned something. Both land identically on a payroll line. Only one of them builds anything.

AI did not create that preference. It supplied a more respectable justification for it.

Brynjolfsson was direct with me about the pattern. Almost everybody overemphasizes buying the technology and chasing the easy-to-measure things, headcount reduction among them.

His own field research makes the cost concrete. Studying 5,172 customer support agents working alongside an AI assistant , he and his co-authors Danielle Li and Lindsey Raymond measured an average productivity gain of 15%. The least experienced agents improved most, in both speed and quality. The most experienced gained slightly in speed and got marginally worse in quality.

The technology was not replacing capability. It was manufacturing it, and almost all of the return came from the people who had least of it. An organization reading only its payroll line would remove exactly those workers and delete the gain in the same motion.

The same study found three other effects. Customers grew more polite. Fewer conversations escalated to a manager. Agents were likelier to stay. None of those has a line on a P&L either.

The instrument recommends the weaker strategy, and it recommends it in the language of discipline.

What The Fix Actually Looks Like

He gave me one example of an organization that got this right, and its value is that it required no new data collection whatsoever.

The company wanted to know whether its call center was performing. It had net promoter scores: one question, a scale of one to ten, one number at the end. What it also had, and was not using, was millions of transcripts.

They ran sentiment analysis across them. Not a new question put to customers, but a reading of what customers had already said, in their own words, while the experience was still happening. Where in a conversation the tone turned. The result was a far better picture of how customers felt than any score built on one question could produce.

Nothing was created. The value had been sitting in the record the entire time, unpriced and therefore unread.

His conclusion went past the metric. New measurement, he said, ultimately drives different ways of doing business.

What To Do Before The Next Approval

I asked him what he would tell a chief executive about to commit serious money to AI. His answer was about measurement rather than technology. Move past measuring the technology investment, he said, and start measuring which specific tasks and functions you expect to change, and be hard-headed about connecting those to value. It is the task-by-task approach Workhelix, the company he co-founded, was built around.

He was unimpressed by the alternative. Too many chief executives issue a broad instruction that the company needs to use more AI, he said. They literally send memos that just say that.

Two moves follow, and neither requires new software.

Instrument the tasks before funding the tools. Name the specific tasks expected to change and state how the change would be detected. If the answer is that overall productivity will improve, there is no measurement, only a hope. A task you cannot instrument is a task you will never prove you improved.

Count time and capability, not just cost. The standard business case is built on cost removed, because cost removed has a number. Time returned to a senior person, decisions made with better information, work that no longer escalates. These are the corporate equivalent of consumer surplus and they need a recorded form, or they will keep losing arguments to the payroll line.

The eighty-eight words asked for institutions that steer AI toward complementing people rather than imitating them. Everything above is what that request looks like once it stops being a policy sentence and becomes a budget decision.

His own line on the statement was that AI capabilities are advancing far faster than our understanding of the economic implications, and that the opportunities of the era sit inside that gap.

That is more true inside a company than at national scale, because a company has fewer excuses. It is far easier to observe what is changing in a business you run than in a continental economy, and the continental economy is currently doing a better job of it.

The institution with the most direct leverage over the outcome is the firm, and it is running an instrument that registers cost removed with great precision and capability created not at all.

Brynjolfsson calls himself a mindful optimist rather than an unconditional one. He believes a much better world is available and does not believe it arrives on its own. "We really do have a lot of agency," he told me, and he meant it as a warning as much as an encouragement.

Agency requires being able to see what you are deciding about. Right now most companies cannot.

The instrument will not warn you. It will report steady state right up to the point the numbers move, and then it will describe the loss with impressive accuracy.