In 1987, Nobel laureate Robert Solow looked at the American economy — awash in personal computers, networking equipment, and corporate IT spending — and said something that should have ended every productivity argument for the next forty years:
"You can see the computer age everywhere but in the productivity statistics."
He was right. Despite trillions invested in information technology throughout the 1970s and 1980s, productivity growth had actually slowed — dropping from 2.9% annually to 1.1%. The computers were everywhere. The gains were nowhere. Economists called it the Solow paradox, and it haunted the IT industry for nearly two decades before the numbers finally turned around.
This week, a study published by the National Bureau of Economic Research surveyed nearly 6,000 CEOs, CFOs, and senior executives across the US, UK, Germany, and Australia. The finding: roughly 90% said AI has had no measurable impact on productivity or employment at their firms. Seventy percent of those companies are actively using AI. It is not that they have not adopted it. They have. It is just not doing anything.
Apollo chief economist Torsten Slok put it bluntly: "AI is everywhere except in the incoming macroeconomic data."
Solow is back. Same paradox. Different technology. Same explanation.
The Last Time This Happened
The original IT productivity paradox was not eventually resolved by better hardware. The 486 did not suddenly unlock what the 386 could not. What changed was how organizations structured themselves around the technology.
It took until the mid-1990s — nearly two decades after the first wave of corporate computing — before productivity growth surged again. And the difference was not faster processors. It was Walmart redesigning its entire supply chain around real-time data. It was financial firms rebuilding their trading floors around networked systems. It was companies that stopped bolting computers onto existing workflows and started rebuilding the workflows from scratch.
The gains came when the humans adapted. Not when the technology improved.
1.5 Hours a Week
Here is the number buried in the NBER survey that tells the whole story. While two-thirds of the executives surveyed said they personally use AI, their average usage was 1.5 hours per week.
One and a half hours. Per week.
That is not adoption. That is tourism. These are people who opened ChatGPT, asked it to summarize an earnings report, nodded at the result, and went back to doing everything the way they did before.
You cannot bolt a language model onto a 1990s workflow and expect a productivity revolution. You cannot use AI for 90 minutes a week, within the same org chart, the same approval chains, the same meeting cadence, the same reporting structures — and then publish a survey saying the technology does not work.
The technology works. The integration does not.
Slop In, Slop Out
There is a phrase that keeps running through my head when I read these studies: slop effort in, slop effort out.
I wrote about the Janitor Effect a few weeks ago — the dynamic where executives feel productive because they are delegating ideation to AI, while their staff spend days cleaning up the vague, hallucinatory output that lands on their desks. That is slop in, slop out at the organizational level. The executive puts in a lazy prompt, gets a lazy strategy, hands it downstream, and someone else absorbs the cost of the laziness.
But the problem runs deeper than bad prompting. It is structural. Most companies are using AI the way most companies used computers in 1983. They took the typewriter off the desk and put a word processor in its place. Same job. Same workflow. Same output. Slightly faster typing. And then they wondered why productivity did not double.
AI in 2026, for most organizations, is a faster typewriter. It is a slightly more capable search engine. It is an autocomplete for emails. It is a summarizer of documents that people were not reading in the first place. It is a tool shoved into a process that was never redesigned to take advantage of what the tool actually does.
What Adaptation Actually Looks Like
The people getting real leverage from AI right now are not using it within their existing workflows. They are rebuilding the workflows around the capability.
A developer who uses AI to autocomplete lines of code gets a 5% improvement. A developer who uses agentic AI to run four parallel workstreams, review its own output, and iterate autonomously gets a 20x multiplier. The difference is not the model. It is the architecture of how the human interacts with it.
A CEO who asks AI to summarize a report saves fifteen minutes. A CEO who restructures their entire information pipeline so that AI pre-processes, triages, and surfaces only the decisions that require human judgment — that is a different operating system for the business. But that requires rethinking the org chart, not just installing a chatbot.
The pattern from the IT era holds: the technology is the easy part. Rewiring the humans and the institutions around it is the hard part. And we have barely started.
Five Shifts That Separate Typewriter Mode from Infrastructure Mode
I have been operating with AI as infrastructure — not as an accessory — for months now. Here is what I have learned about the difference between the 90% who see nothing and the people who see everything.
1. Audit the workflow before you touch the tool.
Most people start by asking "what can AI do?" Wrong question. The right question is "what do I actually do all day, and which of those tasks should not exist anymore?"
Map your actual workflow. Not the idealized version in the employee handbook — the real one. Every step, every handoff, every approval chain, every meeting that exists because someone needed context they could have gotten from a document. Then ask: which of these steps exist because a human was the only option? That is where AI goes. Not on top of the workflow. In place of the parts that were never supposed to be manual in the first place.
2. Write the spec before the prompt.
The single highest-leverage thing I did was writing a standard operating procedure before I ever asked AI to do real work. A document that describes my standards, my voice, my process, my file structure, my checklists. Not a prompt — a spec.
When most people prompt AI, they are giving it a blank canvas and hoping for a Rembrandt. When I prompt AI, it already knows my codebase, my style, my quality bar, and the seven steps that constitute "done." The difference in output quality is not subtle. It is the difference between a temp on day one and a team member on month six.
If you are getting slop out, look at what you are putting in. Vague prompt, vague result. Detailed spec, detailed result. This is not a technology problem. It is a management problem.
3. Kill the task, not just the time.
The 90% are measuring "time saved on existing tasks." That is the wrong metric. The right metric is "tasks that no longer exist."
When Walmart adopted IT in the 1990s, they did not ask "how much faster can a clerk count inventory?" They eliminated manual inventory counting entirely and replaced it with real-time data from the supply chain. The task itself disappeared. That is where the productivity showed up.
With AI, stop asking "how much faster can I write this report?" and start asking "why does this report exist, who reads it, and can the information be delivered in a way that does not require a human to compile it at all?" If the answer is yes, the report is gone. Not faster — gone. That is adaptation. Everything else is a faster typewriter.
4. Manage it like an employee, not a search bar.
I wrote about the $20 employee and the delegation framework: My Job, Our Job, Your Job. The people who fail with AI are the ones who skip straight to "Your Job" — dump a vague request, get a vague result, conclude the technology is broken.
You would never hire a new employee and on day one hand them a sticky note that says "make the company better" and walk away. But that is how most organizations use AI. No onboarding. No context. No feedback loop. No escalating trust based on demonstrated competence. Just a prompt and a prayer.
The onboarding matters. The iteration matters. The feedback loop — where you explain why the output missed, not just that it missed — is what turns a chatbot into a collaborator. Skip the management layer, and you will be in that survey next year telling the NBER that AI does nothing.
5. Redesign the org chart, not just the toolchain.
This is the one nobody wants to hear. The reason the Solow paradox lasted twenty years in IT is that companies kept putting computers into organizational structures that were designed for filing cabinets and carbon copies. The computers were fast. The org chart was slow. The org chart won.
The same dynamic is playing out now. Companies are giving every employee a ChatGPT login and keeping the same approval chains, the same meeting cadence, the same reporting structures, the same layers of middle management that existed before any of this technology arrived. The AI is fast. The org chart is slow. The org chart is winning.
The companies that will break through the paradox are the ones that ask the uncomfortable question: if we were building this organization from scratch today, with AI as a given, how many of these roles, these meetings, these approval layers would exist? The answer, for most companies, is significantly fewer. And that is exactly why it is so hard — the adaptation requires people to redesign themselves out of the process.
The Uncomfortable Implication
If the Solow paradox holds — and every piece of macro data says it is holding — then the uncomfortable implication is that AI will not produce broad economic gains until organizations undergo the same kind of painful, structural redesign that the IT revolution required in the 1990s.
That means new management practices. New org structures. New definitions of what constitutes work. New skill requirements that have nothing to do with "prompt engineering" and everything to do with systems thinking, task decomposition, and the ability to direct autonomous agents the way you would direct a team.
The executives in that survey are predicting AI will boost productivity by 1.4% and output by 0.8% over the next three years. Those numbers are modest to the point of embarrassment for an industry that has attracted trillions in investment. But they might be right — not because the technology is limited, but because the adaptation has not happened yet.
The 1990s productivity boom did not start until companies stopped treating computers as accessories and started treating them as the foundation. The AI productivity boom will not start until companies stop treating language models as toys and start treating them as infrastructure.
The Signal in the Noise
Meanwhile, the people who have adapted are operating at a level that makes the survey results look absurd. One developer building a 180,000-star GitHub project. Small teams shipping products that used to require fifty engineers. Solo operators running consulting practices that compete with agencies.
These are not anomalies. They are the leading edge of the adaptation curve. And the gap between them and the 90% who report zero productivity impact is not a technology gap. It is a thinking gap. The technology is identical. The results are not.
That is the Solow paradox in its purest form. The tool is the same. The outcome depends entirely on whether you redesigned the work around the tool, or just set the tool on top of the old desk and hoped for the best.
The Protocol: The Solow paradox was never about the technology. It was about the humans refusing to change how they work. Ninety percent of executives say AI has not moved the needle. They are right — for them. Because they put slop effort in and got slop effort out. The tool is not broken. The integration is. Adapt the work to the tool, or the tool is just a faster typewriter.
