Seventy years of trying to get the work done without doing it
Everyone was wrong
about which part
was the job.
Not wrong in an interesting way. Wrong in the same way, four times in a row.
Each time a machine took over some piece of business—the arithmetic, the finding, the running of servers, the first draft—the people doing that work mistook the repeatable part for the valuable part.
What follows is that pattern, and then a fifth version happening now.
↓ 7 min · 5 eras · one pattern
ERA 01 · THE RECORD · 1954—1990
The first business computer was bought to do payroll.
In 1954, General Electric became UNIVAC's first commercial customer. The machine went into Appliance Park in Louisville to handle payroll, budget analysis and inventory control. It did not design a refrigerator or make a management decision. It was aimed at the work where being interesting had no value and being right had all of it. The installation was difficult enough to unsettle other buyers; the system became fully operational in 1956.1
Burton Grad, who helped design GE's payroll system, later described a machine with 2,000 memory locations and no programming or data-management tools. Multiplying two numbers took three instructions. Correctness was not a feature layered onto the work. It was the work.2
The immediate result did not fit the clean job-loss story. Clerical occupations kept growing as a share of American employment until 1980, according to the Bureau of Labor Statistics. Holding a number in one place became cheaper long before office work disappeared. What grew valuable was deciding which numbers mattered, how they related, and what should happen next.3
ERA 02 · DISTANCE AND CONTEXT · 1995—2013
Craigslist broke the price. The iPhone broke the place.
In early 1995, Craig Newmark began emailing friends about events around San Francisco. The list became a website; the website accumulated jobs, apartments and things for sale. Craigslist stayed mostly free and deliberately plain. By 2004 it was drawing about 4 million visitors and 1.5 million postings a month with a staff of 14.4
Newspapers saw online classifieds coming and built listings sites of their own. What they could not do was match a price of zero on the line item that helped pay for the newsroom. United States newspaper classified revenue fell from $19.6 billion in 2000 to $5 billion in 2011. Calling the industry slow misses the mechanism. Many publishers were structurally unable to be right.5
Then, on January 9, 2007, Steve Jobs introduced the iPhone. Apple sold the millionth unit 74 days after it went on sale. Finding a customer was no longer just cheap; location, identity, search and payment could travel with the customer. A business no longer waited for someone to arrive at a desk and describe the situation.6
Reach became free. Being believed did not.
ERA 03 · THE SERVICE · 2006—2020
The server became a line on a credit-card statement.
On March 14, 2006, Andy Jassy announced Amazon S3 with a proposition that would have sounded irresponsible a few years earlier: storage should be taken for granted. There was no minimum fee. A gigabyte cost 15 cents a month to store and 20 cents to transfer.7
The launch announcement named a Berkeley science project that planned to distribute 60 million microscope images to 100,000 volunteers. The numbers mattered because the customer no longer needed to buy for the peak in advance. Infrastructure became a service that could expand after demand arrived instead of before.7
This solved one problem and manufactured another. Okta reported in 2018 that the median number of apps in its customer base had grown 24 percent in two years; 87 percent of customers had at least one custom integration, and 64 percent had more than four. Cloud made starting cheap. It did not make agreement automatic.8
THE FAILED PREDICTION · 1975—1995
The office became digital.
It did not become paperless.
In a 1975 BusinessWeek article, Xerox PARC director George Pake predicted that by 1995 office workers would call documents onto desk terminals instead of reaching for paper. The terminal arrived. The clean disappearance of paper did not.9
The prediction matters because it was right about the mechanism and wrong about the outcome. New tools rarely delete a whole system of work. They remove a cost, invite more use, and expose the next constraint.
Until this point, every wave on this page automated a step. Somebody still had to decide which step came next.
ERA 04 · FROM DRAFT TO TASK · 2023—2025
A draft became a task before teams learned to review either.
OpenAI released ChatGPT as a research preview on November 30, 2022. The first broadly useful business effect was mundane: a memo, job description, landing page or function no longer had to begin from nothing. By 2025, Stanford's AI Index found that 88 percent of surveyed organizations used AI in at least one business function. Generative AI had reached 53 percent population adoption in three years.1012
By March 2025, OpenAI was defining agents as systems that independently accomplish tasks for users and shipping tools for web search, file search, traces and evaluations. The unit was no longer only a paragraph. It could be a sequence: find the record, compare the policy, update the system, report the exception.10
DRAFTS: 4,096 / REVIEWED: 3
The surprise is that visible activity is not the same as useful throughput. In a randomized study of 16 experienced open-source developers completing 246 real tasks, METR found that early-2025 AI tools made completion 19 percent slower. The developers believed afterward that AI had made them 20 percent faster. They accepted fewer than 44 percent of AI generations and spent measurable time reviewing and cleaning the output.11
That result does not prove AI makes work slower in general. The developers knew their repositories unusually well, and the tools have improved. It does prove something operators cannot afford to ignore: generation can get cheap faster than judgment does.
ERA 05 · THE OBJECTIVE · 2026
The agent can keep going. The business still needs an owner.
The 2026 AI Index records the capability jump plainly. On OSWorld, a benchmark of real computer tasks, agent success rose from roughly 12 percent to 66.3 percent. That is a large gain. It is also a system still failing about one attempt in three under benchmark conditions.12
Adoption tells the same two-sided story. Eighty-eight percent of surveyed organizations reported some AI use in 2025, but agent deployment remained in the single digits across nearly every business function. Most production systems are still narrow, and the distance between an impressive demonstration and a workflow trusted with money, customers or reputation remains enormous.12
That is the strongest objection to calling this a new era, and it is partly right. Capability alone has never been the whole economic change. The shift becomes real when a business can give a system a standing objective—keep the ledger matched, keep the pipeline current, surface the exception—and rely on it to return on a cadence with evidence.
The human work is deciding what "done" means in writing, which systems may be touched, what must be reversible, and when the machine has to stop and ask. Agentic work does not remove the owner. It makes ownership impossible to hide.
The work left the machine.
The responsibility
did not.
The payroll clerk gave way to the analyst. The classified salesperson gave way to a harder question about trust. The server room became an operations problem. The blank page became a review queue. Each wave made an old action cheaper and moved value toward the person who could define the outcome and recognize the exception.
For a small business, the practical move is narrow: choose one recurring workflow, measure its time and error rate, automate the reversible parts, and keep a named person at every decision involving money, customers or reputation.
In 2026, the new job is deciding what "done" means before anything runs.
Sources & method
Dates and quantitative claims are linked below. Company announcements are used for launch details and prices, not as independent proof of business impact. The machine-share bar remains illustrative. No first-person loss or client result has been invented to satisfy the copy format.
- Computer History Museum, “Making UNIVAC a Business.”
- Burton Grad / Computer History Museum, “The First Commercial Computer Application at General Electric.”
- U.S. Bureau of Labor Statistics, “Assessing the Impact of New Technologies on the Labor Market.”
- Craigslist, mission and history; archived 2004 profile for visitor, posting and staff figures.
- Pew Research Center, State of the News Media 2012.
- Apple Newsroom, “Apple Sells One Millionth iPhone.”
- Amazon, 2006 Amazon S3 launch announcement.
- Okta, Businesses at Work 2018.
- Society of American Archivists, citation from Sellen and Harper on the paperless office prediction.
- OpenAI, “Introducing ChatGPT”; “New tools for building agents”.
- METR, randomized study of early-2025 AI and experienced open-source developers.
- Stanford HAI, 2026 AI Index Report.