How Productivity Is Measured: A History

By Victor Hoang · Updated Oct 21, 2026

Picture an Egyptian inspector named Merer, sometime in the reign of Khufu. Each day he wrote down what his crew of boatmen did. His logbooks record days and missions, not blocks per hour. The distance between that logbook and a quarterly Bureau of Labor Statistics release is the subject of this article.

People have counted work for as long as they have written anything down. But they only learned to measure productivity, meaning output against the resources used, in the 20th century. Scribes in Mesopotamia and Egypt logged rations, deliveries and days of labor. Frederick Taylor timed tasks with a stopwatch in the years before his 1911 book. The U.S. Bureau of Labor Statistics (BLS) turned output per hour into a national statistic starting in the 1940s. Every era since has found that the measure it trusted missed something important.

That's the throughline of this history: each new way of measuring productivity answered one question and hid another. Counting motions couldn't see motivation. National statistics couldn't see the computer. And today's dashboards still struggle with the work most of us do, which is thinking. If you want the definitions first, start with what productivity is. For the wider story of how the idea itself developed, see the history of productivity.

Timeline at a Glance

The Tally Keepers (c. 3100 BCE-1900)

Scribes count rations and rank officials, but never output per hour

Start with a lump of clay. The Metropolitan Museum of Art holds a Sumerian tablet from the Jemdet Nasr period, around 3100 to 2900 BCE, described as an administrative account concerning the distribution of barley and emmer. It is an early account of goods handed out, the kind of count every productivity measure needs for its inputs. The same museum holds an Ur III tablet, dated around 2042 BCE, that is a record concerning laborers needed for irrigation work. Work and workers were already being written down in bulk.

Egypt left a different kind of record. In 2013, archaeologists at Wadi el-Jarf, a harbor on the Red Sea, found hundreds of papyrus fragments from the end of Khufu's reign. Pierre Tallet's 2017 publication describes the archive as the oldest papyrus archive found in Egypt, made up of two kinds of documents: accounts of commodities delivered to the workers, and logbooks of their daily activities over several months. The logbooks follow the inspector Merer and his team as they moved limestone blocks from the Tura quarries to the Great Pyramid at Giza. One set of papers tracks what went in (supplies). The other tracks what the crew did (days and missions).

China took a third route and measured the official instead of the output. Yi-Chun Huang's 2022 study of the Han dynasty describes performance quantified through points called suan: evaluators awarded points across several categories, then sorted officials of the same rank and post into nine grades. The top three grades were called zui and the bottom three dian, and days on duty were a basic factor. Merit went up or down with the grade, which makes it a close relative of a modern performance review.

Put side by side, the three systems solved different problems. Mesopotamian clerks accounted for goods. Egyptian inspectors accounted for crews and deliveries. Han evaluators accounted for people. None of them computed output per hour of labor, and nobody was trying to. Peter Drucker, writing in 1999, went so far as to say that in economic theory "Productivity did not exist" before Taylor. The ancient records were precursors, not productivity measures. What they share with later methods is the habit of writing work down so someone else can check it. For the origin of the word and the idea, see the origin of productivity; for the chronology across all of it, see the productivity timeline.

What this means for teams today: records come before metrics. If you can't say what was delivered, by whom and on which day, no dashboard will fix that.

The Stopwatch Years (1900s-1920s)

Taylor times the task and Gilbreth counts the motions

Frederick Winslow Taylor wanted to know how much pig iron a man should load in a day. His 1911 book, The Principles of Scientific Management, tells the story of a laborer he called Schmidt. The gang averaged about 12.5 long tons per man per day. Taylor wrote that a first-class handler ought to load between 47 and 48 long tons, and Schmidt was shown loading 47.5 tons in a day, for wages about 60 percent higher than men not on task work.

Taylor's measure was the task. Break the job into operations, time each one, set a daily standard, and pay for hitting it. That was a different unit from anything the scribes used. It measured output per worker per day, against a standard someone had derived from observation. Taylor also credited Frank Gilbreth's bricklaying work in the same book: Gilbreth had reduced the movements per brick from eighteen to five, and in one case to two. Where Taylor timed the job, Gilbreth counted the motions inside it, a split traced in our history of time and motion studies. Our explainer on time and motion study shows how the two methods combine, and scientific management covers the movement and its critics.

Drucker later credited Taylor with the first systematic look at work and estimated that manual-worker productivity had gone up 3% a year compound, fifty-fold since Taylor. The new problem was the one workers felt immediately: a measure of the task treats the worker as a unit of output. Whatever the stopwatch couldn't capture, such as skill, judgment and morale, simply didn't count.

Government statisticians were starting to ask the same kind of question at national scale. In 1922, BLS Commissioner Ethelbert Stewart signed an agreement with the Babson Statistical Organization for a productivity study of the construction industry. The tools were still industry by industry, though. A single number for the whole economy was a generation away.

What this means for teams today: a time standard works for repeatable, physical or procedural work. Use it where the task is stable, and be careful applying it to work whose value is in the judgment.

The National Ledger (1940-1960s)

Washington turns output per hour into a standing statistic

By 1940, the question had moved from the factory floor to the national accounts. On June 7, 1940, Congress passed the statute that authorized and directed BLS to make continuing studies of productivity and labor costs. The Bureau already had a head start: in 1935 it had proposed a study of 50 industries to the Works Progress Administration, and it ran several surveys with the WPA's National Research Project. After the 1940 law, the work was permanent.

The results began to add up. In 1955, BLS published labor productivity measures for manufacturing as a whole, covering 1939 to 1953. In 1959, it published measures for the total private economy covering 1909 to 1958. The first report in a quarterly series, "Review of Wages, Prices, and Productivity, Second Quarter 1967," followed. Labor productivity, in the language BLS still uses, compares the growth in output to the growth in hours worked.

Meanwhile, an economist was asking what those numbers meant. In a paper published in August 1957, Robert Solow set out a method for separating the growth of output per hour into the part explained by more capital per hour and the part left over. Over the 40 years he examined, 1909 to 1949, output per man hour approximately doubled. By his calculation, about one-eighth of the increase was traceable to more capital per man hour and the remaining seven-eighths to technical change. Under that reading, most of the rise came from technical change rather than from more capital per hour, and that part wasn't directly measured at all.

That result created the new problem. If most growth came from something unmeasured, then output per hour could rise or stall for reasons no hours-and-capital count could explain. The 1980s would put that to the test.

What this means for teams today: a single productivity ratio tells you how much changed, not why. Pair any headline number with a view of what moved underneath it, whether that's headcount, tooling or process.

The Paradox Years (1970s-1990s)

Computers spread everywhere while the statistics stay flat

In 1983, BLS added a new yardstick: it began publishing multifactor productivity, now called total factor productivity, for the private business, private nonfarm business and manufacturing sectors. Where labor productivity divides output by hours, total factor productivity (TFP) compares output with a combination of inputs. It is the descendant of the question Solow posed in 1957. Six years later, in August 1989, BLS switched its main measure from hours paid to hours worked, so the numbers would better reflect labor actually used in production.

Then the computers arrived and the numbers refused to move. In 1987 Solow remarked that, in his words, "You can see the computer age everywhere but in the productivity statistics." The line stuck, and so did the puzzle. When economist Jack Triplett reviewed the debate in a 1999 Brookings paper, he laid out seven explanations. Among them: computers were still too small a share of the economy to show up; the price indexes for computers might be wrong; the sectors that used computers most, such as finance and insurance, had poorly measured output; some benefits went uncounted; and the effects might be real but lagged, as with electricity.

The late 1990s supplied part of the answer. In 2000, Stephen Oliner and Daniel Sichel of the Federal Reserve Board reported that U.S. labor productivity growth had rebounded after nearly 25 years of sluggish gains. Their abstract credits information technology use and computer production with about two-thirds of the one percentage point step-up between the first and second halves of the decade, and concludes that "information technology largely is the story." The paradox had been partly a timing problem, partly a measurement problem. If you want the full computing backdrop, see the history of computers. And the same question has returned with AI, as covered in the AI productivity paradox.

What this means for teams today: a tool you've bought isn't a change you've made. Give a new system a realistic time horizon, and measure the workflow it touches, not the license count.

The Knowledge-Work Problem (1999-2019)

Drucker names the gap that no ratio can close

By the end of the century, most workers in rich economies weren't loading pig iron. They were writing, deciding, teaching and selling, and the old yardsticks had little to say about them. Drucker's 1999 article in the California Management Review, Knowledge-Worker Productivity: The Biggest Challenge, put the problem plainly. He listed six factors that determine knowledge-worker productivity. The first is the question "What is the task?" The others include autonomy, continuing innovation, continuous learning and teaching, and treating knowledge workers as an asset rather than a cost.

The factor that matters most for measurement is the fifth. Drucker wrote that productivity of the knowledge worker is not, at least not primarily, a matter of the quantity of output, and that quality is at least as important. In manual work, he noted, quality is a floor: a minimum standard to be met. In most knowledge work, quality is the essence of the output. He gave examples. We judge a teacher by how many students learn, not how many fit in the class, and a medical laboratory by how many results are valid, not how many tests its machines can run. Some quality is measurable, like the survival rates surgeons are judged on. But for much knowledge work, he wrote, "we mainly have judgments rather than measures." The deeper difficulty, in his view, was disagreement about what the task is and what it should be.

Through these years, national measurement kept running. BLS reported that nonfarm business labor productivity grew at an annualized rate of 1.5 percent from the fourth quarter of 2007 through the fourth quarter of 2019, below the 2.1 percent long-term rate since 1947. At the level of a team or a company, though, a falling or rising national figure doesn't help a manager decide what to do on Monday. That gap is where team-level measures such as OKRs and KPIs do their work. Our piece on measurement and continuous improvement shows how teams fill it.

What this means for teams today: before you track the volume of knowledge work, define its quality. If two people on a team disagree about what "good" looks like, the metric will only measure the disagreement. Our Peter Drucker profile goes deeper on the ideas.

The Age of Signals (2020s-present)

Remote work, developer metrics and AI test every yardstick

The 2020s turned measurement into a live argument. Start with where people work. An earlier randomized experiment at Ctrip, a travel agency in China, ran call center volunteers at home or in the office for nine months. Home workers' performance was about 13 percent higher: around 9 points from more minutes worked per shift and around 4 from more calls per minute. They were also promoted less often relative to their performance. The newer test, published in Nature in 2024, followed 1,612 employees at Trip.com in a six-month randomized trial of hybrid work, two days at home and three in the office. The authors found no significant difference in performance reviews or promotion rates, and attrition fell from 7.2 percent to 4.8 percent. Our guide on remote work productivity covers the working habits behind those results.

Software teams faced the same problem in a sharper form, since developers' output is easy to count and hard to judge. In 2021, Nicole Forsgren and five co-authors published the SPACE framework in ACM Queue, arguing that developer productivity cannot be measured by a single metric or dimension.

Then AI arrived and made the gap between feeling and measuring visible. In the "Generative AI at Work" study, Erik Brynjolfsson, Danielle Li and Lindsey Raymond tracked 5,179 customer support agents. Access to an AI assistant raised issues resolved per hour by about 14 percent on average and 34 percent for novice and low-skilled workers. A very different result came from METR's 2025 randomized trial: 16 experienced open-source developers worked on 246 tasks, and tasks took 19 percent longer when AI was allowed. Beforehand the developers expected AI to speed them up by 24 percent, and afterward they believed it had sped them up by 20 percent. The METR authors are careful about what that shows: only 16 developers, one kind of work, early-2025 tools, and they don't claim it applies to most software work.

Both findings can be true at once, because the tasks, the tools and the people were different. That's the point. At the national level, BLS reported that nonfarm business labor productivity rose 1.4 percent in the second quarter of 2026 at an annualized rate and 2.2 percent from a year earlier, and that the current business cycle's 2.1 percent annualized growth since the fourth quarter of 2019 matches the long-term rate since 1947. Whether AI is changing that is a question the statistics can't yet settle, and a team trying to decide where to deploy AI needs its own measurement. Our guide to AI productivity ROI metrics gives a framework for building it.

What this means for teams today: don't trust how productive a tool feels. Run a before-and-after on one workflow, with a measure of quality alongside speed.

The measures side by side

Each method in this history measures a different thing. Mixing them up is where most arguments begin.

Key People

Key Facts

Key Facts: Measuring Productivity

• Solow's 1957 paper found that of the doubling in output per man hour from 1909 to 1949, about one-eighth came from more capital and seven-eighths from technical change.
• In Taylor's pig-iron story, the gang averaged about 12.5 long tons a day and a trained handler loaded 47.5 tons.
• BLS has measured labor productivity for the total private economy back to 1909, and the nonfarm business rate since 1947 is 2.1 percent a year.
• Oliner and Sichel credited IT use and computer production with about two-thirds of the one-point step-up in late-1990s productivity growth.
• In METR's 2025 trial, developers using AI took 19% longer, yet believed they had been 20% faster.

The Pattern Behind the History

Across 5,000 years, five patterns repeat. We call this the Proxy Ladder, our own synthesis rather than a published model: each era measures the thing it can see, then climbs a rung when that proxy fails.

1. Count the visible input first. Clay tablets counted barley, Merer's crew logged days, Taylor timed tasks. The first measure is always whatever can be seen and written down.
2. Aggregate when you can't see inside. BLS and Solow moved from the job to the nation, trading detail for comparability.
3. The measure lags the work. Computers spread for years before the statistics showed them, and the same debate has returned for AI.
4. Quality escapes the count. Output per hour handled pig iron well and teaching badly. Drucker's answer was to define the task before counting anything.
5. Feeling drifts from measurement. METR's developers believed they were faster and weren't. Self-reports and dashboards need a controlled comparison beside them.

What It Teaches Teams Today

Treat measurement as a set of questions, not a score. A few practical rules follow from the history.

• Write the task down first. Drucker's "What is the task?" comes before any metric.
• Pick the right rung. Use task standards for repeatable work, outcome measures for judgment work, and ratios like output per hour only to see change over time.
• Pair speed with quality. A volume measure alone rewards the wrong behavior, in a call center or a code review.
• Run before-and-after tests. Compare one workflow before and after a change, as the Trip.com and METR studies did with controls, instead of trusting impressions.
• Expect a lag. A new tool or process can take time to show in results, and that's not always a sign it failed.

Measurement has always been a way of deciding what to pay attention to. The teams that do it well keep the number honest about what it can and can't see.

Productivity measurement has moved from clay tablets to controlled trials, and each step added precision without removing the need for judgment. Start with what the work is, count what you can, and keep an eye on what the count misses.