The Real Reason AI Isn't Paying Off

We have always chased the new tool. Fire, the printing press, the railroad, the computer. The human story is one technology after another, and reaching for the next one is the most human thing we do. We are doing it again with AI, and we should.

Right now the chase is enormous. Four companies plan to spend around 725 billion dollars on AI infrastructure in 2026, up from a record 388 billion the year before. It is the largest infrastructure buildout in modern history. In May 2026, OpenAI and Anthropic each created a dedicated team whose only job is getting the technology into companies, and raised billions to do it. They are not betting on just a smarter model, they are betting it on making the model actually work inside a business.

The sexy words are: AI architect. Agentic engineering. LangChain. Evals. Those are the words you put on a resume right now if you want to be part of the revolution. We love the machine. We love building it. We always have.

But the productivity is not showing up.

McKinsey's 2025 survey of nearly two thousand companies found that while 88% use AI in at least one part of the business, only 39 percent see any impact on their earnings. BCG looked at more than a thousand firms and found that 5 percent capture real value at scale, while 60 percent get little or none. Bain found that most gen-AI use cases met their technical expectations, yet under a quarter of companies could tie them to actual revenue or cost. The technology works. The value does not arrive.

But none of this would surprise an economist. What we are watching is not a failure of AI. It is the normal shape of how a powerful new technology gets absorbed, and it has happened the same way every time, with electricity, with the computer, with the internet. There is even a name for it.

The J-curve

Economists call it the productivity J-curve. A general-purpose technology arrives, and productivity dips before it climbs, because the gains wait on the slow work of rebuilding everything around the new tool. Brynjolfsson, Rock, and Syverson documented it. Electricity is the cleanest example. Factories first bolted electric motors onto the same layout built for steam power, and got almost nothing. The real gain came thirty years later, when they redesigned the whole factory around what electricity made possible. The motor mattered. But it was only ever half of it. The other half was the redesign, and without it the motor did almost nothing.

You can bolt a powerful technology onto a company as it is. People do it all the time. You just don't get much for it. The real productivity shows up only when you change the business itself to unlock the new tool, and that is true of every general-purpose technology we have ever adopted.

Same story, new machine

We hire the architect to build the agent, scope the work, and ship the tool, and we expect it to land. The trouble is in those last two words. It has to land in the business, which means dropping something built outside the company into years or decades of systems, some documented, most not, all tangled together in ways no one currently sees.

The research is blunt about what separates the winners. In that same McKinsey survey, the companies actually getting results were close to three times more likely to have fundamentally redesigned how the work gets done, and redesigning the work turned out to be one of the strongest predictors of whether AI reached the bottom line at all. Yet only about one in five companies had redesigned a single workflow. We are buying the motor and keeping the old layout, exactly like the factories did.

If it is a transformative technology, for it to transform the productivity of a business, it has to transform the business.

But you cannot redesign work you do not understand, and that is the part everyone skips. Before you can rebuild how a company runs, you have to know how it actually runs, which is harder than it sounds. Here is what that looks like up close. A procedure is written down, but one person does a step a little differently than the doc says. The person downstream knows it, so she quietly adjusts her own step to match. Multiply that by a thousand small adjustments that live only in people's hands and heads, and you have a system that works in practice and breaks the moment an outside tool assumes the documented version is the real one. The tool is correct on paper and useless in practice.

Two sides of the coin

If AI is going to touch the work of a company, someone has to understand that company whole, from the inside out. Not the technology stack. The human system.

Think of it as two sides of one coin. One side is the technical integration: connecting the AI to the data, the software, the infrastructure. This is the side that looks amazing in a demo. In a vacuum, with clean data and controlled inputs, it is phenomenal. The other side is the human integration. You have to understand how the work truly happens, the messy and undocumented way a business actually runs, and then rebuild it. That side is slow and unglamorous, with nothing to show off along the way, so it gets skipped, and skipping it is why the coin never pays.

The way out

None of this is an argument against the technology. The pull toward the new tool is human nature, and it is the right instinct. It is the same instinct that got us from fire to the printing press to the machine that can now reason. We should be excited.

The J-curve dip is coming whether we like it or not. The only question is how long we sit in the trough. The companies that climb out fast will not be the ones with the best model. They will be the ones who did the boring half of the work. They understood how their business actually runs, and then they rebuilt it.

The technology half is the one that leads every revolution. It comes first, it gets all the attention, and by itself has never produced the massive historical gains. We keep learning the same lesson, and we are learning it again now. The technology is never what has to change. The business does.

A fair question I am leaving for another piece: whether the technology is mature enough yet to redesign a whole company around, or whether for now the honest move is smaller, targeted rebuilds. That is a question of timing. It does not change the rule. Real gains still come from changing the business, not from bolting on the tool.