Every analyst crunching census data to score AI job exposure is doing something that cannot be done. That is the central claim of Benedict Evans in this piece, and he backs it with a century of accounting automation as the test case. From punch cards to VisiCalc to ERPs to cloud, the tech industry spent 100 years automating accounting. CPA headcount went up anyway.
The piece does real work explaining why. Two mechanisms are in play. First, regulation creates demand that overwhelms automation effects, breaking any clean causal story. Second, the Jevons paradox applies directly: when a DCF model drops from one week to 30 seconds, firms run more DCF models, not fewer. Dan Bricklin documented CPAs using VisiCalc in the late 1970s to compress month-long projects into days. The profession grew. 'Exposure to automation' can mean more work, not less.
The more important argument is buried in the final section and worth reading in full. Census categories like 'accountants and auditors' stay stable while dozens of adjacent finance job titles appear and disappear around them. The label persists, the job changes entirely. Evans argues this makes AI exposure mapping structurally useless, not just imprecise. The jobs that get created are in categories that do not yet exist.
[READ ORIGINAL →]