Eldar Maksymov, an accounting professor at Arizona State University, has a new SSRN paper arguing artificial intelligence will quadruple finance employment rather than gut it. His main evidence is one piece of software: the spreadsheet.
The accountant math
Numbers from the Bureau of Labor Statistics do most of the work in Maksymov's SSRN paper. Roughly 339,000 US accountants in 1980, the year after VisiCalc shipped for the Apple II. About 1.4 million by 2022. The American population grew about 47 percent over those four decades. Accountants grew closer to 320 percent. Bookkeeping clerks, the role that actually got automated, shrank.
His framing is the Jevons paradox. British economist William Stanley Jevons argued in his 1865 book The Coal Question that James Watt's more fuel-efficient steam engine should have cut British coal use and instead did the opposite, because cheap coal made entire new industries economically viable. Maksymov's applied version: "the spreadsheet didn't replace the accountant" but exposed an appetite for financial analysis that companies couldn't previously afford. That sounds airtight, until you press on it.
What the paper is responding to
Anthropic's March 2026 Economic Index reported that Claude could theoretically perform 94.3 percent of tasks in business and finance occupations, tied with computer and math jobs for the highest theoretical exposure of any group. That number is the one driving the C-suite anxiety the paper is trying to talk down. Maksymov's bet is that the 94.3 percent is a productivity ceiling, not a headcount ceiling: cheap financial modeling becomes ubiquitous financial modeling, and demand expands to absorb the new capacity.
There's a demographic argument too. The AICPA's pipeline framing notes that roughly three-quarters of US CPAs are at or near retirement age, and CPA exam candidates dropped from about 50,000 in 2010 to around 32,000 in 2021. There is no horde of accountants for AI to displace, the argument goes. There is a shortage AI happens to be arriving in time to plug.
Where it gets shakier
Jevons is an empirical pattern, not a law of physics. The cleanest counterexample sits in the same country: per capita US residential electricity use has fallen about 5 percent since 2010, even as the average household acquired smart speakers, streaming devices, and at least one extra laptop. The EIA credits LED lighting and tighter efficiency standards. Cheaper, more efficient electricity did not conjure unbounded new demand. It made the existing demand cheaper to serve, full stop.
The bigger tension in the paper is internal. Maksymov leans on the line that AI automates routine while humans climb to higher-value cognition. But that routine-versus-cognition split is what gets harder to defend with every model release. Anthropic's own data shows software development and computer-mathematical work as Claude's single largest use category, and software engineering does not have an aging-out workforce. There is no demographic ramp for AI to fill there. There is also no obvious cognitive tier above the work being automated to retreat to. The accountant analogy assumes a higher floor exists. It might not.
What's worth keeping
The honest version of Maksymov's case is the boring one: AI has clearly enabled products and analyses that did not exist before. Cash-flow forecasting at a 50-person regional manufacturer was a Fortune 500 capability in 1979. The market for it now exists because the cost fell hard enough. That part of the spreadsheet story is real, and it's the most useful piece to carry into the AI conversation.
Everything else is comfort-reading dressed in nineteenth-century economics. Anthropic's next Economic Index update, expected later this year, will be a cleaner test of whether the displacement-or-expansion question is resolving in either direction.
