Predicting AI job exposure — Benedict Evans

Benedict Evans

Benedict Evans

Predicting AI job exposure

It would be really nice if we had some way to analyse which jobs, companies and industries were exposed to AI, and if we could assign scores, and build charts, and map that against the progress of large language models. We know, in principle, that like every other big wave of technology, AI is bound to destroy some jobs and create others. But which ones? In the last three years a bunch of people have been very busy crunching census data, making tables and building viral charts.<br>I think this is mostly impossible: I think this is an exercise in predicting something that cannot be predicted.<br>The simplest way to see the problem is to back-test this against other big technology shifts in the past. Some of the industries that should have suffered most ended up much bigger, and some of the industries that did suffer most should have been immune.<br>Hence, we spent a century automating accounting: we built calculating machines, punch cards, mainframes, data processing, databases, PCs, spreadsheets, ERPs, cloud… in fact, we built half of the tech industry around automating this. Yet the number of accountants kept going up.

View fullsize

This is high-level survey data, but you can see much the same thing at the micro level. 
The next chart is about as specific as it gets: 50 years of financial automation doesn’t seem to have hurt the market for CPAs. If you’d done any kind of analysis of professions exposed to automation from computing, this should have been at the top of the list. Dan Bricklin talks about CPAs in the late 1970s using VisiCalc to do one-month projects in a few days. And yet, look what happened.

I think there are three things to point to in this chart. The first is that technology was not the only variable: changes in regulation produced new accounting requirements that led to a one-off surge in CPA hiring (this is why economists say ceteris paribus). Second, within the automation conversation itself there is the Jevons paradox, which is really applied price elasticity: if you make it cheaper to do something, do you do the same for less money (or resources, or employees), or more for the same money, or does a new ROI mean you do more for more money? If a DCF takes a week and then it takes 30 seconds, you probably do more DCFs. ‘Exposure to automation’ might mean more work, not less.<br>But then, the more important story is that if you automate something that used to be expensive and time-consuming and it becomes cheap and quick, that probably unlocks other things. 
If analysis becomes cheap and easy, you do much more analysis, and mostly that’s also a different kind of analysis. Accountants today aren’t doing exactly the same work that they did in 1970 or 1980 ‘but more’ - they’re still called ‘accountants’ but the job is different. New technology often starts out being used for ‘the old thing but more’, but it rarely ends up like that.<br>Indeed, if you dig into the detail of the Census data, then ‘accountants and auditors’ itself is a fairly stable category, but all around that term there are lots of other finance job categories that appear and disappear over time. The job of “Billing, posting and calculating machine operator” appeared in the stats for a decade or so and then disappeared again. How often did that represent someone who started their career as a stock clerk, then became a ‘posting machine operator’ because that was how you did stock-keeping, and then retired as a stock clerk again when that was absorbed into software and the Census didn’t create a category for ‘PC operator’? Equally, there’s still a category for ‘data keyer’ but not for ‘ERP operator’. The same person doing the same actual job (or rather, serving the same business purpose) gets different job titles over time, while ‘accountants’ have the same job title while doing different things.<br>Then, I think there's a second problem that comes up in back-testing: the job might not change at all, but the business might change underneath you.

View fullsize

The internet didn't really change what it took to be a good journalist or a good A&R scout, but the job of journalism was paid for by a light manufacturing and trucking operation with (in the USA) a local monopoly on classified ads, and the record executive’s salary was paid by manufacturing and shipping small pieces of plastic and aluminium foil. 
That was a whole other thing that would not be captured in any analysis you tried to do of what it is to be a copy editor or a sound engineer. 
The internet decoupled a class of business where the product and the job were not affected by the internet but the business was.<br>It seems to me that we should expect the same thing to happen with AI: how many people have a job that has very low exposure to AI, but the business depends on some other job that is hugely affected by AI? How many people have a job doing something that’s very hard...