Somebody used to do this for a living

The strange thing is that clerical work didn't disappear. In May 2023, office and administrative support was still the largest occupational group in the United States, with 18.5 million jobs. Nobody abolished administration. What changed is harder to see.

I should say plainly where I'm standing. My company puts reporting and data processes onto phones and dashboards. We spend a lot of time trying to remove paper, duplicated entry and the little manual steps that accumulate around a process. We're one of the mechanisms of digitisation, so this isn't an argument from the cheap seats.

And one of the things we still receive, surprisingly often, is a photograph of a form.

Or a letter. Or a handwritten note. Someone has taken a piece of paper, photographed it and attached the image to a digital report.

Somebody still has to read it, work out what matters, decide what it means, classify it, connect it to the right record and perhaps type part of it into another system.

That distinction turns out to have a literature behind it. Researchers Lise Justesen and Ursula Plesner have a wonderfully unglamorous name for the work that appears around digital systems: invisible digi-work. They divide it into three kinds. Compensating work. Connecting work. Cleaning work.

That may be the most useful description of modern administration I've come across.

The machine does its job. Then somebody compensates for what it missed, connects it to something it cannot reach, or cleans what it produced.

Compensating, connecting, cleaning. The strange thing is that we once had a single word for somebody who did all three, and an official description of what it involved.

In 1959 the United States Bureau of Labor Statistics described what a secretary did. They handled "many business details for their employers on their own initiative", it said, "such as acknowledging correspondence, scheduling appointments and meetings, and obtaining a variety of information". Many "serve as representatives of their employers". The ones who were promoted were "given authority for making certain decisions".

In 2026 the same agency describes the same occupation as doing "routine clerical and organizational tasks".

The judgement didn't go anywhere. It was written out of the description.

This isn't new. In a House of Lords debate in 1953, peers were discussing research showing how much of a ward sister's day had been swallowed by organisation and clerical work. The complaint was that the ward sister was becoming "so little of a nurse and so much of a clerk". The proposed answer was familiar too: employ clerical or secretarial support to take forms and telephone calls away from trained nurses.

Thirty years later, the human-factors researcher Lisanne Bainbridge described one reason why. Automation does not simply remove work. It changes the collection of tasks left to the human. The designer automates what can be specified and leaves the operator with what could not be designed away.

That remainder can be small enough to disappear from the business case and large enough to fill somebody's afternoon.

More recent research makes the point less theoretical. A 2024 study of digital public administration in Germany found what its authors called "work for digitalisation": scanning, IT support, software maintenance and data validation. Less complex tasks were partly substituted, but new compensating, connecting and cleaning work appeared around them. The researchers found the workload could increase rather than simply evaporate.

There's an even stranger version when the work crosses the boundary of the organisation.

Self-service can be genuine automation. It can also be what researchers studying self-checkout have called pseudo-automation: the task is not performed by a machine at all. It is performed by the customer, unpaid.

You have probably done a version of this yourself. An organisation sends you a document. Later, another part of the same organisation asks you to upload a copy. You photograph the letter on the kitchen table, crop it, find the upload page and discover the file is too large.

Nothing in that sequence looks like a job. That's partly the problem.

The National Audit Office found something similar at a much larger scale when it examined the UK tax system. Between 2019-20 and 2023-24, HMRC's own cost of administering the tax system rose 15% in real terms. The NAO also found increasing administrative burdens on taxpayers despite the availability of digital channels, and noted that while HMRC estimates the cost to businesses, there is no equivalent estimate for individuals.

The organisation can count its administration. The person on the other side often cannot.

None of this means technology has failed. Quite the opposite. The useful warning is that successful automation can make the remaining work harder to notice.

The newest tools are already showing the same pattern. A Danish study linking surveys of roughly 25,000 workers to administrative labour-market records found widespread adoption of AI chatbots and reported productivity benefits, but no measurable effect on earnings or recorded hours large enough to exceed 2% two years after ChatGPT's launch. What did change was the structure of work, including new tasks around AI oversight and integration.

In medicine, a 2026 study of 8,581 clinicians across five academic medical centres found that AI scribes were associated with 16 fewer minutes of documentation per eight scheduled patient hours. That's real time. But after-hours electronic-record work did not significantly change, and clinicians delivered about half an additional visit a week. The saved time did not necessarily become empty time.

And the hours are not all rising. English teachers worked 48.1 hours a week in 2024 against 54.4 in 2016, and the fall came almost entirely out of the part that was not teaching. Roughly three quarters of them still say they spend too much time on general administration.

That's why I no longer think the useful question is simply, "What can we automate?"

What work will this remove? And what work will be left behind, moved somewhere else, or created around the edges?

The second question is the one most business cases are bad at.

It's also why the photograph of the form bothers me. Not because paper is old-fashioned. Sometimes taking a photograph is exactly the quickest and most sensible thing for the person standing there to do.

It bothers me because it's easy to count the digital submission and call the process digitised while ignoring everything somebody still has to do with the picture afterwards.

Ask what people are putting together by hand to keep somebody else informed. Ask what gets copied between systems. Ask what gets cleaned on Friday so the dashboard works on Monday. Ask what customers now type that an employee used to type. Ask what happens when the one person who quietly knows the workaround goes on holiday.

You will find administration in places where nobody has an administrative job title.

And sometimes you will find that the thing described as automated is simply a very efficient way of moving the work to somebody you stopped counting.