Bot-Sitting

For two years the enterprise AI pitch has been a clean subtraction: hours in, hours out, productivity up. The numbers coming out of the field this week say the subtraction is doing something stranger. Workers do feel faster. They are also quietly absorbing a second job they did not have before.

Rebecca Hinds, presenting Glean’s workplace research, gave the headline version employers want to hear: 73% of workers say AI makes them more productive, and on average report saving 13 hours a week, roughly a third of a work week. That is the number that gets into the deck. The number underneath it is that the same report finds people spending 6.4 hours a week feeding context to, correcting, and re-prompting the tools. About half the time saved goes right back in. Hinds calls the new labor “bot sitting,” and notes 36% of AI sessions simply fail.

The quality of what comes out the other side is its own problem. Hinds cites 40 to 41% of employees admitting they ship AI work they could not explain if asked, and a Stanford finding that 41% of YC AI startups are automating tasks people would prefer to keep human. Her sharpest line was structural: “the human becomes the integration layer.” The productivity gain is real; it is just being paid for in a currency the dashboards do not track.

This is where the turn matters commercially. Nathaniel Whittemore framed the pricing shift bluntly: the move from seat-based to usage-based pricing is the difference between Anthropic’s roughly $3 million run rate last year and its roughly $47 billion run rate now (his figures). Usage-based pricing bills the bot sitting. Every failed session, every re-prompt, every context dump is revenue. The 6.4 hours is not waste from the vendor’s side. It is the meter.

You can see enterprises starting to price this in. Jason Lemkin was flat about it: “the recovery is not coming if you’re not getting AI budget.” Pedro Franceschi at Brex said publicly that the CEO now has to be the chief AI officer because nobody else understands the bounds of the technology well enough. Both are describing the same admission: running AI inside a company is not a tool deployment, it is a permanent operational overhead someone senior has to own.

Which makes Rajiv Jain’s number land harder. Roughly a trillion dollars a year is going into AI capex, he said, against maybe 70 to 80 billion in AI revenue. That gap closes only if the productivity story is real on net, not on gross. The Glean data is the first honest look at the net, and the net is roughly half of the gross. Worth it, probably. Two-to-one, not ten-to-one.

The tell for next quarter is whether anyone starts measuring the bot-sitting hours the way they measure the saved ones. Hinds notes that the 13% of organizations actually capturing significant productivity gains do one thing differently: they measure a lot, and they put the data in employees’ hands. Everyone else is flying on the top-line 13-hours number and paying the usage bill at the bottom.