In the 1980s, people who knew how to use Lotus 1-2-3 for spreadsheets were the most expensive people in the office. By the early 2000s, if you didn’t know Excel, you couldn’t find a white-collar job. The shift from a few people’s advantage to everyone’s requirement took twenty years.
On June 2, OpenAI published a report on Codex and knowledge work. Buried in it is a recommendation that should make policymakers pause: treat AI fluency as basic economic infrastructure. Not “encourage people to learn AI.” Not “cultivate AI talent.” Roads, power lines, broadband. Something everyone has by default, without which you cannot participate in the economy. The specifics listed: funding hands-on AI training through schools, community colleges, public agencies, libraries, and employer partnerships.
This framing means OpenAI is betting on the same verdict: the shift from “knowing AI” as a competitive edge to “not knowing AI” as an employment barrier will be compressed to four to five years this time.
To understand why that bet holds, you need a piece of history that seems to have nothing to do with AI.
In 1987, economist Robert Solow wrote a line in a New York Times book review that became one of the most quoted quips in economics: “You can see the computer age everywhere but in the productivity statistics.”
He wasn’t joking. By the late 1980s in America, personal computers had entered offices at scale. Companies had spent tens of billions on IT systems. Banks and insurers processed transactions on mainframes. Word processors replaced typewriters. By any reasonable measure, these investments should have shown up in the productivity data. They didn’t. Between 1973 and 1995, annual US labor productivity growth fell from nearly 3% in the prior period to below 1.5%. The more computers spread, the slower economic output grew.
This phenomenon eventually got a formal name: the Solow paradox, or more broadly, the productivity paradox. It confounded economists for nearly two decades, until MIT’s Erik Brynjolfsson and his collaborators offered an explanation. Brynjolfsson’s core argument: general-purpose technologies are not plug-and-play. Electricity, steam, and computers are all general-purpose technologies. They can transform nearly every industry, but the transformation doesn’t happen by swapping old machines for new ones. It requires redesigning the entire production process, organizational structure, skill base, and business model around the new technology. These complementary changes lag far behind the technology’s arrival. Until they catch up, costs outweigh benefits, and the productivity data stays flat.
There’s a frequently cited case study here, one the report mentions but doesn’t unpack. It captures exactly why AI faces the same dynamic as electricity did.
In the 1880s, the first electric motors entered American factories. Factory owners did the natural thing: they removed the massive steam engine in the center of the building and replaced it with a massive electric motor. The floor layout didn’t change. All the machines still connected to the central power source through overhead line shafts and belts. One worker pulled a lever, and every machine on the floor started spinning. The only difference was the energy source, steam to electricity. This wave of retrofitting took decades, and throughout those decades, factory productivity barely improved.
Then, in the 1910s and 1920s, several things happened at once. Factories began installing small, independent electric motors next to each machine, the unit drive system. This meant one machine could turn on or off independently, without spinning the belts for the entire floor. Without line shafts and belts in the way, factory ceilings could be lowered — saving on lighting and heating. Machines could be arranged in the order of the production flow rather than by their distance from the power source — saving on material transport. Production lines could start on demand rather than running all at once — saving on energy and maintenance.
The electric motor itself never changed. The technology was identical. What changed was the physical layout of the factory, the way labor was organized, and management’s idea of what a production line should look like. Once those caught up, productivity jumped. Between 1899 and 1929, US manufacturing labor productivity more than doubled, with electrification as the single largest driver.
The lesson of this history isn’t about the motor’s capability. It’s about the speed of complementary change.
Let’s return to that report. OpenAI packed it with numbers: Codex now has over five million weekly active users, up more than 6x since the desktop app launched in February. Knowledge workers account for about 20% of users and are adopting Codex more than three times as fast as developers. Seventy-two percent of knowledge workers use it to produce documents and reports, 47% for engineering operations, 46% for code implementation, 41% for research. Data analysis is the fastest-growing task type, growing 110% week over week.
A product report including these numbers is normal. What isn’t normal is its narrative framework.
The report doesn’t start with product features. It starts in 1850, when 60% of Americans worked in agriculture. It traces the rise and decline of manufacturing, the emergence of knowledge work as the dominant employment category. Then it introduces the Solow paradox. Then Brynjolfsson. Then the electric factory case study. The report positions Codex as the factory redesign for knowledge work — not running old processes through AI, but redesigning work itself around what AI makes possible.
Within this framework, the report lays out four policy recommendations. They aren’t a parallel wish list. They form a continuous chain of reasoning, anchored in Brynjolfsson’s proposition.
First, modernize public sector workflows. Second, treat AI fluency as basic economic infrastructure. Third, put workers at the center of AI adoption — the people closest to the work should decide where AI intervenes. Fourth, update public procurement around jobs to be done, not software licenses.
Taken together, these four mean: AI will produce large productivity gains not through better models, but through organizations redesigning work around AI. This redesign cannot be top-down. Just as a factory owner couldn’t get electrification’s gains by swapping a steam engine for an electric motor, an organization cannot get AI’s gains by handing old processes to an AI to run. The redesign has to start from the people closest to the work. A nurse knows which forms slow down patient care. A caseworker knows where benefits delivery breaks. A teacher knows how much administrative work eats into teaching time. These people don’t need to wait for IT to schedule a project. They need tools and permission, then they can act.
Two data points in the report directly support this line of reasoning.
First, the task mix among knowledge workers. Forty-six percent of knowledge workers using Codex are writing code. Not because they suddenly wanted to learn programming, but because the threshold for “writing a script to finish the task at hand” has dropped below the cost of waiting for a professional developer. A product manager builds the dashboard herself. A researcher writes the data-cleaning script himself. An executive constructs an internal tool that reconciles files and produces a weekly report. As the report puts it: “The boundary between software work and knowledge work has blurred, because AI enables people to reach beyond their formal role and build what their goal requires.”
Second, the shift to parallel tasks. Roughly 50% of users now run multiple Codex tasks simultaneously during the day, up from less than a third in mid-April. Users have moved from “let AI help me with one thing” to “let AI handle three things at once.” This is the same kind of leap as going from using Excel to make a table to using Excel to manage a department’s budget. The tool didn’t change. The dimension of the work did.
Together, these two data points point to the same conclusion: AI tools are shifting from productivity tools for specialists into a broader baseline capacity for getting work done.
When you place this historical thread alongside the current usage data for AI tools, you see an enterprise software adoption model that runs in the opposite direction from the past three decades.
The old way: companies pushed software to users. The CIO selected it, ran procurement, deployed it, trained people department by department. Users waited to be trained and granted access. This path worked for thirty years because it moved slowly enough to match organizational tempo.
Now the opposite is happening. Knowledge workers found AI tools on their own, started using them on their own, and only then did organizations realize they needed guardrails and security policies. Codex’s knowledge worker growth is over three times that of developers’. IT departments didn’t drive this adoption. People discovered something that let them bypass queues and handoffs.
The worker-led AI adoption proposed in the report is an attempt to systematize this bottom-up dynamic. Let nurses, caseworkers, teachers — the people closest to the work — decide where AI should intervene. Give them tools and training to build their own solutions. Not to replace their professional judgment, but to automate the administrative overhead that consumes their time.
The report includes a case study. Taiyo Inoue, a mathematics professor at California State University, used Codex to generate scripts that maintain course information in the Canvas learning management system. It saves him four to five hours of administrative work every week. He reinvested that time into redesigning his classes around collaborative problem-solving, giving students more opportunities to work through mathematics together and in person. A professor writing his own scripts to save time. The students were the ultimate beneficiaries.
Computer skills went from advantage to requirement in twenty years. Smartphones went from a few people’s toy to everyone’s essential tool in ten. How fast AI fluency will make the same transition, nobody knows for certain. But the data points in this report — knowledge workers adopting three times faster than developers, 50% of users running parallel tasks, data analysis growing 110% week over week — together point to a window of four to five years, not twenty.
For individuals, this means: the time you spend now learning to use AI tools to accomplish your actual work is time spent building a capability that will be the default expectation in a few years. The earlier you build it, the more the intermediate years pay off. Build it late, and you get no extra return once it becomes the baseline.
For organizations, this means: the highest-return investment isn’t tool selection. It’s pushing the people closest to the work in front of the tools first. Let them discover what AI can eliminate from their day, then let them act on it. The organization only needs to do two things: grant access, and build guardrails.