Ramp economist Ara Kharazian: 45% of US businesses use AI (not the Census's 8%), and SF's 996 culture is real

Summary

Ara Kharazian, economist at Ramp, argues the Census Bureau's headline figure — that roughly 10% of U.S. businesses use AI — dramatically understates actual adoption, and Ramp's own spend data tells a different story.

The Census question asks whether businesses use AI "to produce goods and services," a framing written in 2023 that maps poorly onto how AI has actually diffused: primarily through back-office tools, call transcription, marketing copy, and productivity software. Most respondents read the question as asking whether AI runs their factory floor, not whether someone on their team uses ChatGPT to draft a Facebook ad. Kharazian also flags methodological noise — the surveys run every two weeks with small sample sizes, and the summer timing likely shifts who within a given company fills them out.

The Ramp number

Ramp's spend data puts U.S. business AI adoption at 45%. Tech and finance lead at roughly 70% of firms on the platform, and Kharazian thinks even that is an undercount because it only captures paid subscriptions, not free tiers or personal accounts employees use at work. Even restaurants show 20% adoption — owners using AI-generated images or copy to speed up marketing work that would otherwise consume time they don't have.

The gap between the Census 10% and Ramp's 45% is partly measurement, partly scope. Ramp observes actual purchasing behavior across tens of thousands of businesses; the Census asks one person per company to self-classify.

996 in San Francisco

The second thread is Ramp's data on the "996" working culture — 9 a.m. to 9 p.m., six days a week — that has become a recurring claim in SF startup circles. Kharazian set out to check whether it was real or just social-media performance.

The signal he used: employee corporate card spend on business meals and late-night food delivery on Saturdays. The finding is geographically narrow. The Saturday spend shift shows up in San Francisco and nowhere else — not New York, not Austin, not Miami. New York shows a partial effect, roughly a quarter of the magnitude, and only after 8 p.m. The SF pattern starts at 9 or 10 a.m.

The timing is also striking. The shift is less than six months old and was essentially absent in 2024. That rules out a long-running baseline and lines up precisely with when 996 rhetoric started circulating in the SF tech scene, which Kharazian loosely ties to the intensification of the AI funding cycle.

Kharazian acknowledges the obvious objection — that Saturday meal expenses could be fraudulent card use — but dismisses it on geographic grounds. If it were fraud, you'd expect it everywhere. The fact that it's concentrated in SF, and only appeared recently, makes the fraud explanation hard to sustain. He notes separately that a credit-risk practitioner told him weekend business-software applications from SF had turned out to be mostly legitimate, while the same pattern elsewhere flagged fraud — an anecdote that directionally supports the thesis.

To test the 9-to-9 dimension properly, GitHub commit timestamps by hour would be the cleaner dataset, though Kharazian doubts Microsoft would publish internal data showing SF engineers outworking Redmond.