Every analyst building AI job-exposure scores and exposure charts is doing something that cannot be done. That is the core argument here, and it is backed with a century of data. The accounting profession is the sharpest example: from punch cards through mainframes through VisiCalc through ERPs through cloud, the entire tech industry spent 100 years automating accounting work. CPA headcount kept rising anyway. Dan Bricklin watched CPAs in the late 1970s compress one-month projects into a few days using VisiCalc. The CPA population grew.
The reason the prediction fails is structural, not incidental. Two forces wreck any exposure score. First, regulation introduces independent demand shocks that have nothing to do with automation. Second, Jevons paradox applies directly: cut the cost of a DCF model from one week to 30 seconds and firms do not buy the same number of DCFs for less money. They do more DCFs. Census data on accounting confirms this at the micro level. The stable category label hides a rotating cast of sub-roles that appear and disappear across decades.
The original piece goes further than the accountant case and works through what this means for reading any AI exposure study published in the last three years. The argument about cheap analysis unlocking different analysis, not just more of the same analysis, is where the piece gets genuinely useful. Read it for the methodology critique, not just the conclusion.
[READ ORIGINAL →]