by Editorial Team, AdvisorAnalyst.com
Every wave of automation produces the same parlour game: rank the jobs, score the exposure, chart the carnage. Benedict Evans has played enough of these cycles to distrust the exercise entirely. In his May 24 essay, he states the thesis without hedging: "I think this is mostly impossible: I think this is an exercise in predicting something that cannot be predicted."
His method is not theoretical objection but historical autopsy. "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." Three case studies carry the argument.
The CPA test. Accounting absorbed a century of direct technological assault — "calculating machines, punch cards, mainframes, data processing, databases, PCs, spreadsheets, ERPs, cloud" — and Evans notes "in fact, we built half of the tech industry around automating this. Yet the number of accountants kept going up." He cites Dan Bricklin's account of CPAs running month-long projects on VisiCalc in days, the precise kind of productivity collapse that should have gutted headcount. It didn't. Evans extracts three reasons. First, regulation moved independently of technology — "changes in regulation produced new accounting requirements that led to a one-off surge in CPA hiring." Second, Jevons paradox: "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." Third, and most important, cheapness unlocks adjacent demand: "if you automate something that used to be expensive and time-consuming and it becomes cheap and quick, that probably unlocks other things... 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." The job title is stable; the job is not — and Evans flags the Census's own taxonomy churn (the vanished "billing, posting and calculating machine operator," the surviving but obsolete "data keyer") as proof that occupational categories themselves are unreliable instruments.
The newspaper test. The internet's real damage to journalism had nothing to do with reporting skill: "the job of journalism was paid for by a light manufacturing and trucking operation with (in the USA) a local monopoly on classified ads." The same logic broke the record industry, where the executive's salary depended on "manufacturing and shipping small pieces of plastic and aluminium foil." The skill was untouched; the business model underneath it was decoupled and destroyed. Evans's question for AI follows directly: "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?"
The Uber test. Mobile-industry veterans in the 2000s discussed location data endlessly without anyone flagging taxi medallions as at risk: "no-one was considering that this could totally change the nature of the job (and make a bunch of $1m medallion mortgages worthless)." Any 1995-era "internet exposure" model would have missed it entirely.
Beneath the historical cases sits a deeper epistemological objection: job descriptions themselves can't be specified completely. Evans likens ONET-style taxonomies to the old dream of expert systems — symbolic AI that tried to hand-code the steps of cognition before machine learning made that approach obsolete. "Sometimes, of course, the job really is just a task, that can be turned into a button, but that's actually pretty rare. Generally, the job is a complex mesh of things that we lack the capability to explain explicitly." He cites ONET's own absurdity — equating a family-trust administrator with a quant-fund desk, both requiring Lotus 1-2-3 fluency rather than Bloomberg — as proof the taxonomy is already broken before AI enters the frame.
He borrows Aaron Levie's framing of Gell-Mann Amnesia: "You have a pretty good sense of how complex your own field is, and how incomplete AI's addressability of that might be, but in other fields you forget this - you see a Claude template for a Powerpoint or a legal draft and you think 'wow, consultants and law firms are screwed!'" The deliverable was never the product — "when you hire Bain, BCG or McKinsey, they will give you some slides, but that's not what you're paying for."
His closing move pre-empts the "directionally correct" defense. Even granting that clerical-heavy roles are plausibly most exposed in aggregate, "you don't know if the exceptions are bigger than the rule" — and a 1995 prediction that the internet would gut physical media distribution would have been directionally right while meaning "totally different things for record companies, newspapers, TV companies and movie studios." The standard he proposes: any exposure model must pass "the newspaper test, the Uber test and the CPA test."
Takeaways for advisors and investors
This is a direct rebuttal to every "AI exposure score by occupation" chart now circulating in research notes and client decks. Evans's argument isn't that AI won't disrupt labor — it's that the disruption routes through business-model decoupling and unlocked demand, not through occupational task-replacement math. For portfolio conversations, the implication is to distrust precision: a sector-by-sector AI-risk heatmap built on job-description taxonomies is performing certainty it doesn't have. The more durable framework is to ask, for any holding, what cheap-and-fast version of its core task gets unlocked, and whether the company's moat secretly depends on an adjacent function that's quietly automatable — the BCG-slide problem, the medallion problem, the classified-ad problem, hiding one layer beneath the headline job title.
Footnote:
1 "Predicting AI job exposure." Benedict Evans, 26 May 2026
