Some 80,000 tons of pig iron lay in small piles beside the works at Bethlehem, waiting to go onto railroad cars. A gang of about 75 men carried it, one 92-pound pig at a time, up an inclined plank and into the car. Frederick Winslow Taylor called them “good, average pig-iron handlers.” They loaded about 12 and a half long tons per man per day.
Taylor’s team studied the work and came up with 47 to 48 long tons a man per day. “This task seemed to us so very large,” he wrote, “that we were obliged to go over our work several times before we were absolutely sure that we were right.”
They picked one man from the gang. Taylor’s book calls him Schmidt, and the conversation began, as Taylor wrote it down: “Schmidt, are you a high-priced man?” A high-priced man would earn $1.85 a day instead of $1.15. He would also do exactly as he was told.
The next day a man stood over Schmidt with a watch. “Now pick up a pig and walk. Now sit down and rest.” At half-past five, by Taylor’s account, Schmidt had 47 and a half tons on the car.
Nearly four times what an average man had loaded. How he got his numbers would follow Taylor all the way to Congress.
Fifty ways to do one thing
Workers, Taylor argued, learned by watching the people nearby, so each trade had “perhaps forty, fifty, or a hundred ways of doing each act.” One of them, he believed, was always quicker and better, and only careful motion and time study could find it.
For shoveling, his men varied the load on the shovel and watched two or three first-class shovelers for weeks. A first-class man, Taylor reported, did his biggest day’s work at about 21 pounds a shovelful, more than with 18 pounds or with 24.
Time the work instead of guessing at it, and a better method often shows up.
Most people take the same route to work for years. Time three routes for a week, and the fastest isn’t always the one you’d have bet on.
Then Taylor took the step people still argue about. The science of the work belonged to the management, he wrote, because the man suited to handling pig iron “cannot possibly understand it.” Planners would write the method down; workers would follow it. That is the separation of planning and doing. “In the past the man has been first,” he wrote in The Principles of Scientific Management in 1911; “in the future the system must be first.”
Twenty to twenty-seven percent
In January 1912 Taylor testified before a special House committee set up to investigate “the Taylor and other systems of shop management.” The chairman, William B. Wilson, pressed him on the stopwatch.
A timed figure was never the final one, Taylor explained. For turning railway axles, his studies added 20 to 27 percent for unavoidable delays. Wilson asked: “By what scientific formula or mathematical calculation did you arrive at an addition of 20 to 27 per cent to the time which you have determined by that stop watch?” Careful, repeated study, Taylor said, “not the arbitrary judgment of anyone.”
The pig-iron numbers drew the same doubts. Taylor’s own book says only about one man in eight in that gang was physically able to keep the new pace. In 1974 the management historians Charles Wrege and Amedeo Perroni re-examined the story; the summary of their paper calls it “more fiction than fact.” A 2000 follow-up, built almost entirely on primary sources, called the anecdote “erroneous.”
Taylor has defenders. A 2001 study judged the discrepancies “relatively minor” beside the lesson, and a 1982 review by Edwin Locke argued that Taylor’s views on time study, standardization, goal setting and rest pauses were “fundamentally correct.”
Taylor chose pig iron, he told the committee, because it was “the simplest kind of human effort.” For a run of only 10 or 100 axles, he added, a man would need as much as 70 percent extra time, because the job was new to him. The 20 to 27 percent was for axles made by the thousand, “year in and year out.”
A shovel for a model
With an AI model, the instruction is the shovel, and you can vary it as freely as Taylor varied the load.
In 2023 researchers at Google DeepMind did roughly that. One model wrote candidate instructions; a second solved a sample of grade-school math word problems with each (3.5% of the GSM8K training set); the best were kept. With PaLM 2-L solving, the winner was “Take a deep breath and work on this problem step-by-step.” It scored 80.2% on the test set, against 71.8% for the familiar “Let’s think step by step.”
The “one best method” turned out to be narrow. “Let’s work together to solve this problem step by step,” which merges two instructions that scored 71.8 and 60.5, scored 49.4. Other pairs of models turned up different winners, and scores on the problems used in the search often ran higher than on the test set.
So treat a winning prompt the way Taylor’s critics treated 47 tons. Set aside ten past cases with known answers that you never use while tuning. Run the old and new instructions on them three times each, and read the answers as well as the scores. If the new one wins only on the cases you tuned it with, you’ve sized the shovel for one man.
Four minutes an invoice
Imagine you run accounts payable, with eight clerks matching invoices to purchase orders, each with a method of her own. You time a week. One clerk averages four minutes an invoice; the team averages eleven. You sit with her, write down how she sorts and checks, and make it everyone’s method.
The average drops. So you set the target at four minutes.
Two months later the slowest clerk on the sheet is the one who handles the old supplier accounts with disputed prices, the invoices the fastest clerk never had to match.