IT HISTORY

Workers Got Measurably Faster and Company Revenue Did Not Move

Robert Solow spotted this gap in 1987 and computers took thirteen years to close it AI is now running the same clock

Photo by eowynring on Unsplash.

In February 2026, Stanford economist Nicholas Bloom put out a working paper most executives would rather not read aloud. Nearly 6,000 CEOs, CFOs, and senior finance leaders across the US, UK, Germany, and Australia had answered the same question, fielded jointly by the Federal Reserve Bank of Atlanta, the Bank of England, the Deutsche Bundesbank, and Macquarie University — what has AI actually done to your business over the past three years.

Eighty-nine percent said it had no measurable effect on labor productivity. Ninety percent said the same for employment. Among the small minority who did report a productivity effect, the average gain was 0.29 percent, not twenty-nine, zero point two nine.

Bloom didn’t reach for a new explanation. He reached for an old one, quoting a line from an economist who had died two years earlier — a sentence already thirty-eight years old when the paper hit the wire.

That sentence turns thirty-nine today.

A Nobel Economist Wrote One Sentence in 1987 and It Outlived Every Book Around It

Robert Solow. Photo by Olaf Storbeck via Wikimedia Commons, CC BY-SA 2.0.

On July 12, 1987, Robert Solow reviewed a stack of books about the information economy for the New York Times Book Review. One line from the review outlived every book on the list — “You can see the computer age everywhere but in the productivity statistics.”

He wasn’t being clever for its own sake. American businesses had spent over a decade wiring themselves with terminals and minicomputers, and increasingly with machines like the IBM PC, sitting on desks previously home to nothing more computational than an adding machine. Corporate IT spending had roughly tripled since 1970. Measured productivity growth, meanwhile, was stuck near its slowest sustained pace since World War II — a slump starting in 1973 with no sign of caring how many computers arrived to fix it.

IBM personal computer, 1981. Photo by Sailko via Wikimedia Commons, CC BY 3.0.

Six years later, economist Erik Brynjolfsson gave Solow’s observation a name still in use today. His 1993 paper in Communications of the ACM called it the productivity paradox, and it asked the question everyone had been avoiding — maybe the technology wasn’t the problem, maybe the way organizations were using it was.

He turned out to be right, but slowly. It took until March 2000, thirteen years after the original line, for Solow to walk it back in an interview with the same newspaper — “You can now see computers in the productivity statistics.”

What had changed wasn’t the hardware. It was everything built around the hardware: flattened management layers, re-engineered workflows, decision rights pushed down to the people who had actually learned to use the tools. Brynjolfsson gave this piece a name too, complementary investment. The computer was never going to pay for itself. The reorganization around the computer was what paid.

The Same Gap Opened Again and This Time the Numbers on Both Sides Are Better

Thirty-nine years and one technology cycle later, the shape hasn’t changed, and this time there’s stronger data on both sides of the gap.

The individual-level number is real, and it has been replicated more than once. In a 2023 randomized trial, ninety-five professional developers split into two groups were asked to build the same HTTP server in JavaScript as fast as they could. The group given GitHub Copilot finished 55.8 percent faster, and the effect was strongest among the least experienced programmers in the sample.

The same year, a study of 5,179 customer support agents given a generative AI assistant found productivity — issues resolved per hour — rose 14 to 15 percent on average, and 34 percent for agents in their first months on the job. One of the study’s three authors was Erik Brynjolfsson, thirty years after he had named the paradox he was now personally proving wrong at the individual level.

At the firm level, the number keeps landing exactly where it landed in 1987. MIT’s Media Lab surveyed the field in 2025 and put a figure on the divide — of the roughly $30 to 40 billion enterprises spent on generative AI in 2025, 95 percent of pilots produced no measurable effect on profit and loss. Deloitte’s 2026 enterprise survey found two-thirds of companies reporting efficiency gains but only one in five reporting any revenue growth from them. PwC’s global CEO survey put the share of executives seeing zero measurable financial benefit at 56 percent.

The individual math works. The company math doesn’t. This is the entire paradox, restated with 2026 numbers instead of 1987 ones.

Where the Missing Gains Actually Go

Economists have a polite name for the missing piece — complementary investment. It’s the right term, but it describes an absence rather than a mechanism. It says what didn’t happen without saying what happened instead.

Work meeting. Photo by Amtec Photos via Wikimedia Commons, CC BY 2.0.

I have sat in enough of these industry rooms in 2026 to see the mechanism directly, and it isn’t idle capital sitting on the sidelines waiting for a quarterly plan. It’s an internal economy built to convert individual AI speed into something resembling output without being output. I run an AI-assisted pipeline myself, and the one rule I never relax is a mandatory human check before anything goes out — not courtesy, insurance against exactly this failure mode.

It tends to start several rungs above the people doing the work. The managers under the most pressure to show ten-times gains are often the ones furthest from a keyboard — the last IDE they opened may have been a decade and two job titles ago. A one-month build cycle compresses to a week, then to a day, each cut justified by a headline about AI-driven velocity nobody in the room can independently check.

The compression doesn’t stop with scheduling. A product manager who once had to sketch a flow and think through an edge case can now hand a vague paragraph to a model and get back a seemingly complete document — organized, illustrated, confident — while answering none of the questions an engineer actually needs answered. Unable to get a straight answer from the person who wrote it, the fastest way to close the loop is to prompt the same model into a demo and ship it.

Quality assurance faces the same volume problem from the other side, and increasingly delegates the first pass of testing to the AI itself. If the model reports no issues, the verdict often stands in for a human ever touching the product. The status report reaching leadership is, not infrequently, generated by another model from whatever raw material the team produced during the week — a document with excellent production values and one clean number attached, the kind that photographs well in a slide deck.

Every layer in that chain is individually rational. Each one is simply handing the volume problem downstream to the next layer, and the layer with the least room left to hand it anywhere is the one actually writing the code.

What looks like a productivity story from the boardroom looks like a shrinking cushion from the terminal — shorter timelines, requirements that multiply because a feature nobody used to ask for now appears to cost nothing, and monitoring tuned to read a pause at the keyboard as a discipline problem rather than a sign the engineer just found something worth thinking about.

What eventually breaks isn’t effort. It’s coherence. A system patched by successive rounds of AI-generated changes, layered on by people responding to different pressures with no one holding the whole design in their head, accumulates failures that resist explanation by design or by model. Some of what gets built this way never reaches a real user. It gets quietly deleted before anyone has to explain what it cost.

The Part No Survey Question Can Measure

None of this shows up in an NBER survey question. “Has AI affected your firm’s productivity” has no box for “yes, and then we spent the gain building things nobody can maintain to prove it.”

The individual gain in the Copilot study and the customer-support study was real precisely because someone was still measuring the actual outcome — tickets resolved, working code. Move one level up the organization, into the layer that decides what gets built and reported, and the incentive to measure the real outcome gets weaker at every rung, right up until the thing being optimized is the appearance of the number itself.

Solow’s paradox resolved in thirteen years, and it resolved through organizational rebuilding, not a faster chip. If the current version resolves the same way, it won’t be a smarter model doing it — it will be some manager, somewhere in the chain, still current enough on what the work actually costs to say no to a report that photographs too well to be true, and being believed when they say it.

References

• Robert Solow, “We’d Better Watch Out” — citation investigation, Standup Economist
• Productivity paradox — Wikipedia
• Bloom, Davis, et al., “Firm Data on AI” — NBER Working Paper 34836
• 6,000 execs struggle to find the AI productivity boom — The Register
• Thousands of executives aren’t seeing an AI productivity boom — Fortune
• Peng, Kalliamvakou, Cihon, Demirer, “The Impact of AI on Developer Productivity: Evidence from GitHub Copilot” — arXiv
• Brynjolfsson, Li, Raymond, “Generative AI at Work” — NBER Working Paper 31161
• MIT report: 95% of generative AI pilots at companies are failing — Fortune
• Deloitte, “The State of AI in the Enterprise” — 2026 report
• PwC 2026 Global CEO Survey — press release

Support writers. All our nonprofit’s offerings here.

Click here to subscribe to the IT Chronicles newsletter.

Click for Wordsmith, Mystery Writing, Write Like Stephen King, more

By the EIC Susan Brearley with Ideogram