A survey of approximately 750 chief financial officers, published in a working paper backed by the National Bureau of Economic Research, found that artificial intelligence had "essentially no employment effect" in 2025. The same survey found that most of those CFOs expect the jobs that will be cut first are administrative ones — the accounts payable clerks, the data entry specialists, the compliance processors, the scheduling coordinators who make up a corporate floor that rarely appears in AI marketing materials. [1]

This paper reported yesterday on xAI's "human emulator" project, which aims to replace white-collar workers with AI agents that mimic their computer usage. The CFO survey provides the demand-side counterpart: the people who control corporate budgets are watching that technology and making plans.

The Wall Street Journal published the survey's findings in its March 25 print edition. The headline framing — "CFOs Say AI Is Coming for Admin Jobs" — captures the tension at the heart of the data. The employment effect so far has been negligible. Firms reported a perceived 1.8 percent labor productivity gain from AI in 2025, a number the researchers noted is "economically meaningful but modest." Revenue-based estimates of AI's impact were even lower. The technology is not yet doing what its backers promised. [1]

But the forward-looking data tell a different story. When asked which roles they expected AI to affect within three years, CFOs consistently identified administrative and back-office functions. The logic is structural: administrative tasks tend to be repetitive, rule-based, and conducted entirely on screens — precisely the characteristics that make them amenable to automation. A human resources coordinator who processes 200 expense reports a day performs a task that an AI agent can replicate. A sales executive who builds relationships over dinner does not. [1]

The survey also revealed that workloads are intensifying. CFOs reported that AI adoption has made remaining employees work harder, not less. The technology is being used to raise throughput per worker rather than reduce headcount — a pattern economists call "labor-augmenting" rather than "labor-replacing." The question is whether that distinction holds as the technology improves. [1]

The NBER paper's authors cautioned against drawing sweeping conclusions. The sample tilted toward mid-to-large firms. CFOs may be poor predictors of technological disruption. And the time horizon — three years — is both too long to be actionable and too short to capture AI's full trajectory.

What the data do establish is a consensus among the people who sign the checks: AI is not taking jobs today, but when it starts, it will start with the people who file, sort, schedule, and process. The admin floor has always been the most vulnerable part of the corporate building. It is now the part that the people upstairs are measuring for automation.

-- DAVID CHEN, Beijing