1987, Again
In 1987, Robert Solow wrote one of the most quoted sentences in economics.
“You can see the computer age everywhere except in the productivity statistics.”
Computers were everywhere. Offices full of them. Companies spending billions. And the productivity statistics showed nothing. Economists called it the productivity paradox. It lasted a decade.
This February, Fortune published findings from a survey of thousands of CEOs. Their verdict on AI: no measurable impact on employment or productivity.
The economists they interviewed recognized the pattern immediately.
They’re calling it the Solow Paradox. Round two.
Why the original resolved
The first productivity paradox eventually ended.
By the mid-1990s, productivity growth came back. And when researchers went back to understand the delay, the answer was consistent: it wasn’t about the computers. The machines had been capable the whole time.
The gains arrived when companies did two things they hadn’t done in the 1980s. First, they redesigned their processes — not just digitized the old workflows but rebuilt what work looked like given what computers could do. Second, they built coordination infrastructure: shared networks, standardized protocols, systems that let people and tools exchange information without constant human intervention in the middle.
Economists gave this a name: complementary innovation. The technology alone doesn’t move the needle. The technology, plus the organizational infrastructure it requires, does.
The computers were already there. The decade of lost productivity was the gap between deploying the technology and rebuilding the layer above it.
The same gap, a different decade
In 2026, the capability story is fully written.
96% of organizations run AI agents. AI tools have proliferated across individual workflows. Every major vendor has shipped. The models are powerful, accessible, and cheap. The Stanford AI Index confirmed it: AI adoption is outpacing both the PC and the internet.
And yet: thousands of CEOs, across industries, saying no measurable impact.
The Solow Paradox doesn’t require ignorance. It doesn’t require failure to adopt. It requires deploying technology without rebuilding the layer above it. The 1980s companies bought computers. The 2020s companies deployed AI. In both cases, the coordination infrastructure wasn’t rebuilt to match.
Here’s what’s different this time: we know the playbook. We watched this movie. We have forty years of research on how the original paradox resolved.
We’re making the same mistake anyway.
The 8% number
There’s a data point circulating this spring that hasn’t gotten enough attention.
Only 8% of the time savings from AI tools are being reinvested into activities that actually benefit the worker. The other 92% is absorbed by organizations as expanded output demands.
Read that once more.
You saved an hour with AI. You kept eight minutes. Your organization captured fifty-two.
This is not a productivity story. It’s a distribution story. The efficiency gains are real — they’re just moving in the wrong direction. Upward, not forward.
The Jevons Paradox explains part of it: efficiency gains get converted into expanded commitments. The ratchet explains the rest: organizations, once they see the gains, reset expectations accordingly. The productivity dividend doesn’t accrue to the person doing the work. It accrues to the system around them.
There’s a specific mechanism here worth naming. In the 1980s, the productivity gains from computing eventually reached individuals — but only after the coordination layer was rebuilt. The gains started flowing when the infrastructure changed who held information, who had to route it, who had to translate it across systems. When coordination became infrastructure rather than human labor, the arithmetic changed.
Until then, every efficiency gain got absorbed by the overhead of coordination done manually.
Same pattern. Same mechanism. Same missing layer.
What got rebuilt in the 1990s
The companies that captured productivity gains from computing didn’t just buy more computers. They rebuilt around them.
They networked the machines so information moved without people carrying it. They standardized data formats so handoffs didn’t require translation. They redesigned processes so the coordination cost of getting things done dropped alongside the execution cost.
The net result: efficiency gains stopped being absorbed by coordination overhead. They compounded instead.
What’s missing from the AI deployment of the 2020s is exactly this.
Every AI agent runs. Most run alone. Outputs land somewhere and wait. Loops open and don’t close. Context doesn’t transfer between tools. The person in the middle holds it all — and the overhead of holding it is exactly where the productivity gains disappear.
94% of organizations are worried about “agent sprawl.” Only 7-8% have what researchers call mature agent governance. That gap is not a compliance problem. It’s a coordination infrastructure problem — and it’s where the gains are going.
The humans filling that gap are the 8% keeping their time savings. The rest went to the system.
The resolution is the same
The Solow Paradox resolved when companies stopped deploying computers and started rebuilding the layer above them.
The AI paradox will resolve the same way.
Not by switching models. Not by buying another productivity app. Not by prompting better or adopting faster.
By building the infrastructure that sits above the AI — the layer that knows what’s committed, what’s open, what’s been handed off and not returned, and what the gains should actually accrue to.
Without that layer, AI efficiency gains get absorbed before they reach the person doing the work. They fill commitment stacks. They get extracted by organizational output expectations. They disappear into supervision overhead that no one designed a system to handle.
The 1990s didn’t solve the productivity paradox by making computers faster. They solved it by rebuilding coordination as infrastructure — so the humans in the middle weren’t absorbing costs that belonged to the system.
That’s the pattern. It’s forty years old.
The CEOs in the Fortune survey aren’t failing. They’re in 1987. The capability is deployed. The coordination layer hasn’t been rebuilt yet.
The question is how long this decade’s paradox lasts.
In the 1990s, the answer was roughly ten years from widespread deployment.
We’re at year two.
