In July I wrote that AI is an extinction-level event, just not for the people everybody assumes. I hung it on the power loom, because the weaving numbers are about as clear as economic history gets: automation stripped 98% of the labour out of a yard of cloth, a single weaver ended up minding eighteen looms, and the number of weavers quadrupled anyway, because cloth got cheap enough that everybody wanted more of it.
The power loom was a British invention. Britain had around 37,000 cotton handloom weavers in 1780 and 240,000 by 1820, because mechanised spinning had made yarn cheap and somebody had to weave it. Then the power loom matured: 2,400 of them in 1812, 100,000 by 1832, 400,000 by 1860, and the share of British cotton cloth woven by power went from 5% to 95% between about 1815 and 1845. By 1860 there were 10,000 handloom weavers left. The trade grew more than sixfold and then all but vanished, in the same century, in the same county. Weaving itself multiplied. The handloom weaver did not survive it.
I have since gone looking for whether that was a one-off, because it is so relevant to where we are today with AI. It is not a one-off.
James Bessen, an economist at Boston University, took the 271 detailed occupations the US Census used in 1950 and followed every one of them through to 2010. Plenty of them disappeared. Boardinghouse keeper went because nobody wanted boardinghouses. Telegraph operator went because the telegraph went. But in exactly one case could the disappearance of an occupation be put down mainly to automation.
One! In sixty years! It was the elevator operator.
Sixty years spanning the mainframe, the minicomputer, the PC, the spreadsheet, the industrial robot, the barcode, the ATM/cash-machine, the internet and the smartphone. Every one of them arrived with a confident, specific, published list of the jobs it was about to erase. The count after six decades is just one.
Bessen's own conclusion is the part worth noting: that period saw enormous automation, and almost all of it was partial. Machines took tasks. They did not take occupations.
The ATM/cash-machine
The bank teller was the prediction everyone was most sure about.
A machine that dispenses cash and takes deposits does the visible part of a teller's job, faster, at three in the morning, for no wage. Over 400,000 of them went into the US. The number of tellers employed went up, from roughly 500,000 around 1980 to roughly 550,000 by 2010.
The mechanism is the thing, not the outcome. Bessen measured that the number of tellers needed to run the average urban branch fell from twenty to thirteen between 1988 and 2004. A branch that needs seven fewer people is a branch you can afford to open in places that could never previously carry one, so the banks opened them: urban branches up 43%. More branches needed more tellers in total, and those tellers spent their day on relationships and selling rather than counting notes, which was the part the machine could not touch.
Same arithmetic as the power loom, a century later, in a service industry, which is the harder case. Automation makes a unit of output cheap. Cheap output gets bought in far greater quantity. Greater quantity needs more people, doing the part that was never the bit being automated. The mechanism has a name and it long predates any of this: William Stanley Jevons pointed out in 1865 that every improvement in the efficiency of the steam engine raised Britain's total coal consumption rather than lowering it, because cheaper coal found more uses than the efficiency saved. Jevons paradox is the loom and it is the ATM, and it has been running underneath every automation panic since.
The end of the teller story is less comfortable. The effect ran out. US teller employment peaked around 2007 and is down to roughly 347,000, with another 13% decline forecast to 2034. What finally did it was online and mobile banking, not the ATM: a change in what customers wanted to do at all, rather than a machine doing the same job cheaper. Thirty years of automation grew the job. A change in demand shrank it. Those are different causes and it matters enormously which one you think you are facing.
None of which is American, incidentally. The first cash machine in the world went into the wall of Barclays in Enfield on 27 June 1967, built by De La Rue to John Shepherd-Barron's idea, dispensing £1 notes against a cheque impregnated with carbon 14. By the time the Bank of England reviewed twenty years of the major British banks in 1991, ATMs roughly equalled branches and about 60% of retail deposit withdrawals went through them, and the Bank's own account of what that did to staffing was this: rapid growth in staff numbers over most of the 1970s and 1980s, with investment in technology "helping replace staff in back-office processing tasks, releasing them for front-office sales tasks". That is Bessen's finding, in the Bank of England's words, a quarter of a century before he published it. The British version differs in one respect. Our branch numbers were already falling while American ones rose, so we got the reallocation without the branch expansion that drove the headcount growth over there. The ending is the same as well: bank and post office clerks are down 83,000 since 2001, and nobody blames the cash machine for that either.
The jobs are not the same jobs
So the honest version of what I argued in July is not "your job is safe".
David Autor and his colleagues at MIT went through eight decades of US occupation data and found that roughly 60% of all employment in 2018 sat in job titles that did not exist in 1940. Among professional occupations it was 74%. Not jobs that had evolved - job titles that had to be invented, because the work had not previously existed to be described.
The British series runs longer and says the same thing. Deloitte's economists went through the census of England and Wales for every decade since 1871 and found agricultural labourers down from 6.6% of the workforce to 0.2%, and 200,000 people washing clothes for a living in 1901 against 35,000 in 2011, in a population that had gone from 32.5 million to 56.1 million. What replaced them was not nothing. Accountants went from 9,832 in 1871 to 215,678. There is now one hairdresser or barber for every 287 people in England and Wales, against one for every 1,793 in 1871. The caring professions went from 1.1% of employment to 12.2%, and the muscle-power trades from 23.7% to 8.3%. Weavers and knitters, since we started there, were down to 4,961 by 2014 from 24,009 in 1992, over a period when total UK employment rose 23%. And after a century and a half of all of that, the ONS puts the three highest employment rates in its entire 1861 to 2018 record in 1872, 1943 and 2018, all at 76% of the working-age population.
The work is safe and expanding. The job description is not.
And before anyone in telecoms takes any comfort from Bessen's number, I should point out that our industry ran the single largest counter-example in the data. Between 1920 and 1940 AT&T automated mechanical switching across more than half of the US telephone network and destroyed the telephone operator, which was then one of the largest occupations open to young women in America. In the cities that cut over, operator employment among 16 to 25 year old women fell by between a half and four fifths.
Feigenbaum and Gross went and looked at what happened to those people. The cohorts that followed were fine: the jobs came back as clerical and service work, including categories of work that had not existed before. The women already doing the job were not fine. A decade later they were disproportionately in lower-paid occupations or out of work altogether.
That is precisely the handloom weaver, and it is the reason I keep saying the technology is not what ruins anybody. Weaving boomed. Telephony boomed beyond anything the operators could have imagined. What got destroyed in both cases was the people and the businesses that kept doing it the old way while the world reorganised around them.
Where the new work actually comes from
Which leaves the question nobody in this debate ever asks. If automation keeps making things cheaper, and employment keeps going up, where is the extra work coming from? It cannot all be the same people buying more of the same thing.
Clayton Christensen named it in The Innovator's Solution in 2003: the growth does not come from competing for customers who are already buying. It comes from competing against non-consumption. The people who were not in the market at all, because the thing cost too much, took too long, or needed an expert they could not get hold of. Make it cheap and simple enough and an industry appears where there was not one, and that is where the unpredictable new job titles come from. You cannot forecast an occupation that serves a market which does not yet exist.
Which is the bit Jevons does not cover. Jevons is the same good, cheaper, bought in far greater quantity; non-consumption is demand that was never in the market to be counted.
It is also exactly why incumbents read all of this backwards. Christensen's real finding was never about technology. It was that thoroughly competent people, inside a business whose cost base demands a certain margin, will reflexively screen out the low-margin opportunity that creates the new market, and will explain very persuasively why they were right to. "AI is for our customers to build" is that sentence, dressed for work. My colleague Peter has written about the four AI strategies currently on display and they are all recognisable from this one insight.
The non-consumption in our industry is the conversation
Telecoms is sitting on one of the largest pools of non-consumption I have ever seen, and most of the industry is looking straight past it.
Email is indexed. Chat is logged. The CRM is structured and reportable. And the conversation, which is where customers state what they actually want, where compliance is met or missed, where fraud is attempted and where emergencies are declared, has historically produced a recording nobody ever listened to. Not for want of wanting to. Reviewing an hour of calls costs an hour of a person, so every business on earth reviews a token sample and bins the rest.
That is not a market being served badly. It is a market that has never existed. Almost every conversation your business has ever had went unexamined, and everyone quietly accepted that as the natural order of things.
Make it cheap and you do not get a smaller contact centre. That is the ATM mistake, it is the Klarna mistake, and it is sitting in a lot of board packs right now. What you get is a contact centre doing things it could not previously attempt at all: signals that fire while the call is still happening, guidance arriving in a handler's ear mid-conversation, the conversation becoming a record that compounds rather than a file nobody opens. The handler with an Operator watching their call is not a cheaper handler. They are a better one than they could ever be unaided, and the work that creates is new work, not recovered work.
We built it on that principle and we built it that way ourselves, right down to the internal plumbing nobody sees. Amplify the people. Do not replace them.
What the elevator operator was actually for
Come back to the one job that did go, because the reason it went is the warning, and it is not the warning people take from it.
The elevator operator did not lose to a cheaper elevator operator. They lost because the entire job was the interface. Their function was to be the human who pressed the button on the passenger's behalf, and the moment the passenger could press it themselves there was nothing underneath. Every other occupation on Bessen's list had judgement wrapped around a task. Automate the task and the judgement is still there, now applied across far more surface area than before, which is what happened to the weaver, to the teller, and to the generation that followed the switchboard.
So the question is not whether AI can do part of what you do. It can, and sixty years of census data says that has never once been enough on its own. The question is whether there is anything underneath. If your whole contribution is standing between a customer and a capability, passing the request along and adding a margin for the service of pressing the button, you are not a business facing disruption. You are an elevator operator, and I have written about your customers before.
Which is the point, and the reason July's post said what it said. Sixty years, one occupation eliminated. There is no equivalent register for companies, and if there were it would run to volumes. Automation has been extraordinarily kind to workers and utterly ruthless with firms, and everybody in this industry asking "how many people can this replace" is asking the question that has been reliably wrong since 1950, while ignoring the one that has been reliably right.
The asteroid was never aimed at your staff. It has your slower competitors' names on it, and if you stand still long enough, yours.
So: what is under your button?