The last time a machine was supposed to erase a job, it made the job better. This time can be the same.

The short version

• Between the late 1970s and about 2010, US bank teller employment rose, through every year the cash machines were supposedly eating it.
• The count peaked near 600,000 around 2010. The sustained decline came later, and it came from banking on a phone, not from the ATM.
• James Bessen’s research says why: the ATM cut the cost of running a branch by roughly a quarter, banks opened more branches, and each one still needed people.
• Inside the job, the machine cleared the queue. Tellers spent the hours on the talk instead of the counting, and the branch became somewhere people could talk to someone.
• The pattern is three steps of arithmetic, not magic: cheaper input, cheaper service, more demand, and more people needed to serve it.
• This October looks the same, but faster. AI took months where the ATM took decades, so the redeployment has to happen while the machine arrives, not afterwards.

No one asked the bank teller what they thought. That is the part worth recovering now.

In the late 1970s a machine appeared in the wall outside the bank. It counted cash. It took the deposit. It handed money over at eleven at night, on a Sunday, in the rain. Across the counter sat someone whose whole day had been built around those tasks. Every forecast in the country drew the same straight line: one machine, fewer people needed per branch, an occupation quietly filed away.

The forecast was wrong. Not subtly wrong. Teller employment in the United States went up. It went up through the 1980s, the decade the machines stopped being a novelty. It went up through the 1990s. It was still going up in the 2000s, when every branch on every high street had one in the wall and no one queued for cash any more. It peaked near 600,000 around 2010. And when it finally turned, the thing that turned it was not the cash machine. It was banking on a phone.

The fear was reasonable

That deserves saying plainly, because no one was being stupid. The transaction was the visible part of the job, and by 1980 a large share of a teller’s hours went on counting and processing. If a machine does the counting and the processing, arithmetic says fewer people are needed. Every doom-laden column of the era rested on a mechanism that was completely real. None of it was invented.

What the prediction missed is that a job is a bundle, not a task. Take one item out of the bundle and the bundle does not shrink by the size of that item. It changes shape. The branch kept its counter, its door, its opening hours, its place on the high street, and gained a machine that took the dullest half of the day and did it without tiring.

So the 1985 columns were right about the mechanism and wrong about the outcome. The lesson lives in that gap, and it is the most useful thing in the whole story.

What actually happened

The record is not ambiguous. The US Bureau of Labor Statistics has counted tellers for decades, and its occupational employment series climbs through the years the machines were supposed to eat the job. It levelled off near 600,000 around 2010 and then began a sustained fall, when online and mobile banking changed where people did their banking, not when the cash machine did.

Boston University economist James Bessen found the reason, and his explanation is now standard in economics. The ATM cut the operating cost of a branch by roughly a quarter. A cheaper branch is a branch worth keeping open, so banks opened more of them: more places, more corners, longer hours. And every one of those branches still needed people.

Cheap work did not mean less work. It meant more of it, in more places, done by more people.

The branch became somewhere to talk

Then something quieter happened, inside the job itself.

A counter is not only a place where money moves. It is where someone sits down with a letter they do not understand. It is where a small business owner works out whether they can cover payroll on Friday. It is where a person who has never borrowed anything in their life asks the question they are slightly ashamed to ask. Those conversations were always part of being a teller. They were always squeezed out by the queue at the door.

The machine cleared the queue. The freed hours went somewhere obvious in hindsight: into the advice talk, the problem solved in one visit instead of three, the local knowledge that is the difference between “no” and “here is how we can do this&rdquo. The counting tier of the job went into a wall. The human tier got its hours back, and the branch became what it had always claimed to be: somewhere you could go and talk to someone.

Put the two versions of that day side by side. In one, half your hours disappear into a machine and you go with them. In the other, the machine takes the counting and hands you the talk. Only one of those is a worse job. That second one is what happened.

The machine did not take the job. It took the part of the job nobody wanted, and left the part nobody else can do.

The arithmetic underneath

None of this is mystical, and it is not technology being benevolent. It is arithmetic, and it runs in three steps. A cheaper input makes the service it feeds cheaper to provide. Customers respond to cheaper by buying more of it. Someone has to serve the extra volume, even though each unit of it needs fewer people than it did before.

The telephone exchanges of the early twentieth century ran the same script. Automatic switching let one operator handle many more calls, and the prediction was mass unemployment among operators. What happened, as Bessen files, is that calls got cheaper, subscriptions rose, and operator employment grew alongside them, for decades.

That story has an ending worth being honest about. The operator job did go, once the technology matured far enough to remove it entirely. Not when anyone said it would, and not before a generation of people had done the better version of the work in the meantime. Which is the truest thing history offers here: not a guarantee, a reprieve, and a direction.

Automation rarely erases a job on the day it arrives. It changes what the job is, changes the economics of the business around it, and only then, sometimes, changes the count.

The number that matters

The whole shape of it, in one table. Numbers only where the sources support them.

Direction and the approximate peak are read from the BLS occupational employment series for bank tellers. Read the series itself for exact year-by-year figures.

This October, the same fear

It is October 2026, and the fear has the same shape it had in 1985.

On 12 October the IMF’s Kristalina Georgieva described AI as a “slow-moving tsunami&rdquo, with about 40% of jobs worldwide exposed to it and about 60% in advanced economies. On 7 October Bluevine reported that 53% of the small business owners it surveyed say AI has already absorbed at least one role’s workload, and 75% plan to absorb more than hire. In the same survey, 65% of owners described their teams as exhausted.

Those two stories are being read the way the 1985 columns read the ATM: exposure equals extinction. The honest response is the one the teller data teaches. Exposure is not extinction. The second-order effect, the cheaper service, the expanded demand, the work that moves up a tier, is where the outcome gets decided. No one in 1985 could see that the cash machine was quietly building a kind of bank that needed more humans, doing better work, for another twenty-five years.

The one place the analogy strains is speed, and it would be dishonest to skip it. The cash machine took decades. AI took months. A teller had a generation to grow into the newer version of the job. The administrator watching AI absorb their inbox this autumn does not get a generation.

That is not a reason to throw the pattern away. It is a reason to hurry the part of the pattern that matters: move people up the tier while the machine starts carrying the volume, not five years after it already does.

What is left is the part only people do

Here is where the old story turns into the new one, and where it becomes something to look forward to than endure.

This time the machine does not stop at counting. It takes the repetitive knowledge work too: the copying between systems, the chasing, the drafting, the remembering, the file no one has opened since March. Every one of those jobs contains a quiet fear that this is all the job is.

It never was. Strip them out and what remains is the work people were always the only ones able to do. Judgement. Relationship. Care. The phone call no one wants to make. The decision you make when the rulebook runs out and someone is waiting. That is not the residue of a job after the machine has eaten it. It is the job, with the packaging removed.

Some roles will genuinely shrink, the teller count did fall, and honesty requires saying so. What the history suggests is not that the numbers hold. It is that what people are left doing gets more human, not less, and that firms which redeploy their people early do better than the ones who wait to be forced.

Every generation has been certain the machine was coming for the work. Every generation has been right about the change and wrong about the ending. The weavers were told this. The telephone operators were told this. The tellers were told this, across twenty-five years in which their numbers went up and their days got better.

The people who feared the machine in 1985 were not foolish. They were early, and they were looking at the mechanism instead of the arithmetic. This time we can do both at once: hand the machine everything it is good at, and give the humans back the part they were always meant to do.

Where Twinthos comes in

That is the whole idea behind a Twinthos digital employee. The repetitive work, the follow-through, the reminders, the routine answers, the chasing that depends on someone remembering to push it, goes to software. The people stay for the part that needs a person. Above the Authority Line a human decides; below it the digital employee acts on its own.

The goal was never a business with fewer people in it. It is a business where the people finally do the work only they can do, and go home with their week intact.

If you want to know what your business’s version of the counting-and-dispensing is, the specific hours a machine could carry from tomorrow, while your team moves up a tier, that is what the £999 Operations Audit measures.

Next step

Start with the £999 Operations AuditBook a discovery call

Twinthos is a managed digital employee service, not software you install. It works across the email, diary and accounting systems your business already runs. Below the Authority Line it acts on its own; above the line a person decides.

Common questions

Did the ATM really fail to cut teller jobs?

It cut tellers per branch, yes. But the Bureau of Labor Statistics series shows total teller employment rising through the 1980s, 1990s and 2000s, peaking near 600,000 around 2010. The sustained fall came with online and mobile banking, not with the cash machine.

Who is James Bessen, and what did he find?

A Boston University economist. His work on automation and jobs found that the ATM cut branch operating costs by roughly a quarter, banks responded by opening more branches, and teller employment rose with them. He documents the same demand-expansion effect in the early telephone exchanges.

But didn’t tellers eventually lose out anyway?

Yes, and the honest version of this story says so. From about 2010 the count fell, because online and mobile banking changed where people did their banking. The machine that did the damage was not the machine everyone was afraid of. Twenty-five years of better work happened in between.

Does this mean AI will not take jobs?

It means exposure is not extinction. Bessen's own research includes technologies that did cut employment, so this is not a promise. It is a reason to look at the second-order effect - what the freed capacity gets redeployed into - rather than reading the technology alone as a verdict.

What is different this time?

Speed. The cash machine took decades to reshape banking; AI has moved through offices in months. A teller had a generation to grow into the newer version of the job. Someone whose inbox is being absorbed this autumn does not. That is why moving people up the tier early matters more than it did.

What does this mean for a small business in October 2026?

This week's numbers - about 40% of jobs worldwide exposed to AI according to the IMF, and 53% of small firms already letting AI absorb a role according to Bluevine - describe exposure, not an ending. The practical move is to decide now which hours go to the machine and which work you want your people doing instead.

Sources

1. Bank tellers employment data, US Bureau of Labor Statistics Occupational employment series for bank tellers, the primary data behind the employment trend.
2. Automation and jobs: when technology boosts employment, Federal Reserve Bank of Boston working paper (Bessen et al.) Primary research on why the ATM increased branch economics and teller employment; the demand-expansion mechanism.
3. IMF managing director on AI and jobs, CNBC, 12 October 2026 The “slow-moving tsunami” remarks and the 40%/60% exposure figures that this article engages with.
4. Bluevine Lean Team Economy Report, PR Newswire, 7 October 2026 The 53% absorbed-a-role and 75% plan-to-absorb findings, with methodology.