The Past , Present, and Future of Office work

Welcome back! Last week we talked about the economic importance of communication and the rise and fall of the telephone operator occupation, which mostly employed young women. Between 1910 and 1940, AT&T and its satellite companies rapidly adopted mechanical switching, which quickly put telephone operators out of work. Yet future cohorts weren’t harmed at all. They simply found other jobs, especially as typists and secretaries.

Why did typist and secretary jobs grow so rapidly in the early 1900s? Interestingly, the word “secretary” doesn’t even appear in the U.S. Census occupation descriptions until 1940. Occupation code number 236 in the 1940 Census is described as “stenographers, typists and secretaries.” From 1910 to 1940 it was just “stenographers and typists”, and before 1910 all office work was categorized simply as “clerks and copyists”.

New job titles often augur substantive changes in work, and this is no exception. The secretary occupation evolved over the first half of the 20th century alongside rapid technological advances that aided the transcription and communication of language from spoken to written form.

The economic importance of note-taking

The original technology for transcribing spoken words was shorthand. The practice of writing shorthand dates back at least to ancient Greece (the word stenography comes from the Greek stenos “narrow” and graphein “to write”), but scholars have found many examples of shorthand from Rome, Imperial China, Japan, and other ancient cultures.

Shorthand exists because people have highly imperfect memories, and because we can speak faster than we can write. Recording speech is especially important when words must be memorialized for legal reasons, such as in court proceedings. James Madison took notes on the Constitutional Convention in his own shorthand, making him the sole recorder of the founding principles of the newly formed United States of America.

Like language, shorthand existed in many forms, but some eventually became more dominant than others. Probably the best-known system is Gregg shorthand, which is still used today for court reporting. Gregg is about three times faster than regular writing, primarily because it records sounds rather than spelling (omitting silent letters) and uses abbreviations for common words (e.g. “k” for “can”). Here is as an example of a sentence written using Gregg shorthand – you can see how it saves time to write this way!

Gregg shorthand allowed stenographers to reach speeds upward of 100 words per minute, compared to about 30 for regular handwriting. That sounds impressive, but the average person speaks at about 150 words per minute, faster than most shorthand writers. However, like most tasks, machines help you do it faster. Miles Bartholomew invented the shorthand machine in 1879, pictured below.

As you can see above, the stenotype keyboard is much smaller than a regular keyboard. Words are written by punching the keys together in combinations (for example, pressing the K, the A, and the T together is how you write the word “cat”).

With modern machines, stenographers can reach typing speeds of up to 300 words per minute, about ten times the speed of normal handwriting and twice as fast as the spoken word.

If shorthand is so great, why isn’t everybody using it? The barriers to entry are too high. Gregg shorthand takes months to master, and as you can see from the drawing above, it is utterly incomprehensible to outsiders. Shorthand greatly lowers the cost of transcription, but not the cost of communication, because non-stenographers can’t understand it.

The typewriter revolutionized written communication The invention that truly unlocked mass communication was the typewriter. The typewriter was independently “invented” many times during the early 19th century, but early prototypes were mostly curiosities that didn’t pass a benefit-cost test relative to shorthand or regular handwriting.1 The first commercially successful typewriter was the Remington No.2, which sold about 100,000 units between 1874 and 1891.

Even though the Remington No.2 was very expensive (it cost about $100, which is roughly $4,000 in 2024 dollars), it was clearly a communications breakthrough, for two reasons. First, Remington showed in a highly publicized contest (see the poster below) that people could reach speeds of 100 words per minute by “touch typing (e.g. using most of their fingers and not looking at the keyboard, the way you are taught in typing class). This offered the possibility of near-perfect translation from spoken to written word, without the need for shorthand. Second, with carbon paper and typewriter stencils, typewriters could make multiple copies of the same document.2

The secretary occupation evolved alongside the mass adoption of the typewriter. Typewriter production exploded in the early 1900s, with major brands like Underwood selling an estimated 5 million units between 1900 and 1930. Growth in typewriter sales was driven by gradual improvements in quality and cost, including front striking (so you could see what you were typing) and eventually the electric typewriter.3 Large companies increasingly employed secretarial “pools”, groups of secretaries who were deployed to executives as needed for typing and other office duties. Gibbs college, the first secretarial school, opened in 1911.

Over a period where the typewriter was rapidly adopted, the occupation “typists and secretaries” grew sixfold, from less than 0.5 percent of all employment in 1900 to 3 percent in 1950.

How the typewriter fueled the rise of office work

Technological innovation in communication increased total demand for communication jobs. New technology didn’t destroy jobs because typewriters can’t write by themselves – they need a human operator. For the first time in history, we could inexpensively create accurate written accounts of meetings, phone calls, and other events. Businesses could also keep records of important transactions such as sales and expenses, and they could store and maintain customers and client information.

Over this same period, we also saw growing demand for other office functions related to information storage and retrieval. The figure below plots the trend over time in office and administrative support occupations. The solid blue line shows employment of financial clerks - people who keep records of finances, payroll and accounts and file and process paperwork from customers and clients. The red dashed line shows typists, secretaries, and administrative assistants and is identical to the chart from last week’s post. Finally, the dotted green line shows employment in other back-office jobs like proofreaders, office machine operators, data entry keyers, and general office clerks.

At its peak in 1980, office and administrative support work accounted for 12.7% of all workers in the U.S. economy. That’s one in eight jobs devoted entirely to the production, processing, storage, delivery, and retrieval of written information.

Yet since 1980, employment in all three occupation categories has declined rapidly, falling from 12.7% to only 6.8% in 2022. Today, secretaries and administrative assistants are as common as a share of all jobs as they were in 1920.

Digitization and the nonrivalry of data

What explains the decline of office work since 1980? If you’ve been reading regularly, you already know the answer. It was the personal computer. Like the typewriter, the personal computer facilitated the recording and storage of information in written form. However, unlike the typewriter, computers can record, store, and manipulate information in digital rather than physical form.

Digital information storage has several advantages. First, you can more easily make changes to a document without having to reproduce the whole thing. Second, filing and organizing is much easier because documents can be sorted on multiple characteristics (you can search the files on your hard drive by keyword, date, or folder - but physical documents can only be in one place). Third, the absence of a physical form means that digital information can more easily be copied, delivered, and preserved. In the language of economics, digital information is nonrival, meaning usage by one person does not crowd out usage by another person. A physical document can only be in one place at a time.

Once information became easy to store and manipulate, we no longer needed so many people to transcribe, codify, and organize it. Secretaries focused increasingly on other duties, like scheduling and coordinating meetings and personal assistance. Yet office calendars are increasingly digitized and synced up within organizations, and only high-level executives have their own assistant. The job of secretary/administrative assistant is likely in permanent decline.

The nonrivalry of digital information (e.g. data) has had broader impacts on the overall organization of the economy. Digitization has lowered to zero the cost of reproducing information, and it has greatly lowered the cost of manipulating and editing documents and other forms of digital data. Basic economic reasoning tells us that falling costs leads to an increase in supply. At the dawn of the 20th century, the U.S. economy was data-scarce – just codifying information alone had economic value. A hundred years later, we are data-drenched, to the point of drowning.

By some estimates, the internet in 2024 collectively stores about 147 zettabytes (ZB) of data.4 One ZB is equivalent to the storage capacity of 250 billion DVDs! Ex-Google CEO Eric Schmidt estimated that only about 0.05 ZBs of information was created over humanity’s entire history up to 2003.

When information is abundant, the ability to make sense of it becomes especially valuable. This explains why managerial, professional, and technical occupations have grown since 1980, as rapidly as clerical work has declined. These jobs require you to go beyond collecting and storing information.

Jobs with titles like “business analyst”, “consultant”, and “solutions architect” require workers analyze and synthesize information in ways that improve business decision-making. In some ways, it is data compression, not collection – distilling a sea of information down to its most critical elements. When you have access to more than 250 billion DVDs worth of data, it’s important to know what you are looking for!

Artificial intelligence - information processing made easy

We can think of the large language models (LLMs) underpinning generative AI tools as performing incredibly sophisticated operations on data (primarily words). Each time you ask ChatGPT a question, it is effectively compressing all 147 zettabytes of the internet in a way that delivers a highly customized response to your specific query.

Generative AI commodifies the manipulation of digital information. It may take half a century, but I believe it will eventually lead to the extinction of office and administrative support jobs like administrative assistants and financial clerks. The entire purpose of these jobs is to lower the cost of transmitting and storing information. In the long-run, AI will drive the cost of “routine” information processing down to nearly zero, eventually eliminating the need for most human labor in those jobs (although we will need a lot more energy efficiency to get there!)

There is a clear analogy here to the impact of mechanization on farm labor. For most of human history, the bottleneck to increasing food production was physical power. Steam and electricity eventually relaxed that constraint, and farm work mostly disappeared because we only need so much food.

Similarly, the key bottleneck in business decision-making for most of modern history was a lack of information. Advances in information collection, storage and retrieval eventually relaxed that constraint, and now we are awash in data. Routine office jobs were created in a time of information scarcity, and they may no longer be needed.

A harder question is whether AI will eventually replace jobs that analyze and synthesize information to improve decision-making. I am less certain that will happen, for two reasons. First, there is an essential complementarity between the ways that LLMs and humans analyze information. Generative AI models excel at any task for which there are many existing examples in their training data. They can access information about anything that has ever happened anywhere repeatedly, a feat far too difficult for the human mind. People, on the other hand, are excellent guessers. We reason remarkably well with very little data. Because AI models and humans approach problems differently, I can imagine lots of situations where people who know when to use AI and when to overrule it will do better than either party acting alone.

Second, economic interactions are often strategic, meaning the right decision depends on what you think your competitors will do and how they will respond. Because LLMs reason from training data, they are never fully up to date. Concretely, imagine that two companies are developing strategies to outcompete their opponent for market share. They use a frontier AI model to tell them where they should locate new stores, but the right answer depends on what their opponent is doing, which in turn depends on what AI model the opponent is using. If you both have access to the same AI technology, the winner will be the company with a better human in charge.5

Still, the frontier of AI technology is advancing rapidly. I am making predictions based only on what I see today and some quasi-linear extrapolation to the near future. If the day comes that an AI agent can run a Fortune 500 company without human assistance, then I, for one, will welcome our new robot overlords.

1 The Early Office Museum contains some wonderfully quirky examples of antique typing machines, including the Kaligraph, Charles Thurber’s Patent Printer, and the Hansen Writing Ball.

2 Two other important innovations were 1) the shift key, which moved a different part of the typebar to contact with the ribbon, allowing for both upper- and lowercase letters to be used without changing the typebar manually; and 2) the QWERTY keyboard layout, which minimized typebar jams by spacing out frequent letter combinations.

3 Electric typewriters were much more reliable and allowed for other complementary improvements like proportional spacing and the typeball or “golfball” design, which reduced jams and allowed multiple fonts to be used in the same document.

4 Caveat – I have no idea how they arrived at this number! It seems like a hard thing to estimate.

5 The idea that what I do depends on what you will do, which depends on what I do, and so on is called “level-K reasoning”. People have already programmed AI agents with level-K reasoning capabilities, but I am not aware of any evidence on whether such agents can reliably outperform people in real-world strategic interactions.