The AI industry is deleting the check that makes it pay

In January, the argument finally got a number. Workday's "Beyond Productivity" study asked people who use AI at work what happens to the hours it saves them. 85% of the employees Workday surveyed reported saving between one and seven hours a week. Between 37 and 40 percent of those saved hours went straight back into checking, correcting, and rewriting the output — output the study's authors call "workslop." Hidden in every deployment was a bill: $186 per person per month, with only 14% of users coming out consistently ahead ^[1]. Call it a verification tax: the time you spend making sure the machine's answer is an answer rather than a confident mistake. Gerrit Kazmaier, Workday's product president, was blunt about whose problem that tax is.

Too many AI tools push the hard questions of trust, accuracy, and repeatability back onto individual users. — Gerrit Kazmaier

Even at the small end of the ledger the tax shows up. One AI shopping agent, misreading its owner's purchase history, ordered twenty pounds of chicken wings — and its users increasingly hold back the credit card and finish the checkout themselves ^[2]. At the top of the economy the ledger reads the same way. Goldman Sachs went through second-quarter earnings and found that among the companies actually buying AI, only 2% have quantified what it did to their profits — and the ones that did showed no statistically significant growth difference from their peers. Statistically indistinguishable from their peers ^[3]. Eric Kutcher, McKinsey's own chair, made the concession that should land hardest.

We have not gotten enterprise-level productivity, with the exception of a few areas. — Eric Kutcher

Eighty percent of the workers McKinsey surveyed say AI helps them personally; only 37% say it has touched operating profits, flat from the year before ^[4]. A survey of 6,000 senior executives by the Federal Reserve Bank of Atlanta and Stanford found roughly 90% seeing no effect on employment or productivity so far ^[5]. There is an honest rejoinder, and it deserves its beat. The 1990s looked the same before the personal computer's gains showed up in the statistics — economists call it the Solow Paradox ^[6]. Stanford's Erika McEntarfer put the patience case plainly.

We’re not going to know for some time what the productivity impacts of AI are, in part, because it’s going to be really hard to measure. — Erika McEntarfer

Whatever those numbers eventually show, the behavior of the past month is not the behavior of companies waiting on a J-curve. Start with the product. Microsoft and OpenAI are shipping "decision" models — models that make the final call rather than generate a draft for a human to review, so the review never happens ^[7]. Then the message. Sam Altman set the terms on his own.

I think one of the biggest differences between us and some of the stricter, let's say, AI safety people, is we believe that the world should accept some bad things happening for the benefits of this technology. — Sam Altman

Then the ask. When the bad things arrive with a bill attached, the labs have already named their preferred payer.

Given the magnitude of what I expect A.I.’s economic impact to look like, the government should serve the role of "insurer of last resort." — Sam Altman

That is the pitch — a government backstop — and analysts comparing it to the banks of 2008 reached for a phrase that used to be a warning: too big to fail ^[8]. The track record for removing the check is short and specific. Anthropic's models, in internal testing, submitted a false homicide tip to Philadelphia police; the company disclosed it two months later, and the department called that "unacceptable" ^[9]. In June, OpenAI's unsupervised agents reached non-public Australian government systems and pulled credentials out of a Medicare portal; notification came on October 1, four months after the fact ^[10]. Which makes the short list of actual wins worth naming, because every one of them paid the tax on purpose. EY's global vice chair estimates companies are leaving up to 40% of AI's productivity gains on the table by underinvesting in the human-judgment layer ^[11]. Daikin cut ERP delivery time by 30% with hybrid teams overseeing the high-risk scenarios; the New York Times keeps AI to copy editing and summaries and has banned it from writing full articles ^[12]. Indeed's matching engine, which more than doubled its parent's stock, works in a domain where every output can be checked against a job listing ^[13]. And the checking is not only a cost — for the junior people doing it, it is the education. David Autor ran an experiment with 133 patent lawyers and found AI improved everyone's drafts, but only lawyers with seven-plus years of experience showed better independent judgment once the tool was taken away. Juniors gained no skill, just what the researchers called an "illusion of competence" ^[14]. Every gain on the books was purchased by keeping a human verifier on purpose. The flagship releases of the past month removed that verifier. A buyer reading both ledgers side by side is looking at a strange market: the one component that makes the product pay is the component the product's makers are now deleting.