Botsitting

There is a kind of work that does not feel like work and does not show up anywhere, and most of my time with AI is spent on it. Deciding what to ask. Loading enough context that the answer has a chance of being good. Reading the output closely enough to trust it. Catching the place where it went confidently wrong and steering it back. Doing the whole loop again because the first result was almost right, which is the most expensive kind of result there is. It feels like producing. It is mostly supervising.

The workplace researchers have started naming this, which is how you know it is real and not just my temperament. Glean's index of digital workers found people spending around six hours a week "botsitting," feeding and correcting and babysitting their AI tools, more time than they spent using those tools to produce the actual work. Six hours. That is most of a working day, every week, going into a column that no budget has and no dashboard tracks. The subscription is the line item everyone debates. The botsitting is the bill.

The attention tax

I have written before about the part of my own mind the agents took over, the long flat middle of a task that a machine is glad to grind through. That part is real and it is a gift. But it came with a tax, and the tax is attention. Every hour the tool saves me on the doing, it bills back some fraction on the deciding and the checking, and that fraction is invisible because it does not look like effort. It looks like reading. It looks like thinking. It is the cost of being the one responsible for output you did not personally produce.

The dangerous thing about this tax is that you cannot feel it from the inside, because the felt experience of using AI lies to you in a specific direction. A 2026 study found what the authors called a speedup illusion: people expected the tool to make them much faster and reported lower effort using it, even on tasks where it saved no time at all. Less strain is genuine and worth something. But less strain is not more output, and a tool that feels like a shortcut while costing you the same hour is the easiest expense in the world to undercount. I have trained myself to distrust the feeling of speed and look only at what actually shipped, because the feeling and the shipping have come apart, and the feeling is the louder of the two.

There is a second move buried in the attention tax that almost nobody counts: deciding when not to use the tool at all. That decision is itself work. Every task now carries a small upfront question, is this inside the frontier or just outside it, is this faster delegated or faster done by hand, and answering it costs a little attention before any work happens. Multiply that micro-decision across a day and it adds up to a real load, the cognitive overhead of running a second worker who is fast, tireless, and occasionally, fluently wrong. I have argued before that attention is the scarce resource, the thing to spend on purpose. Botsitting is what happens when a tool spends it for you without asking.

The coordination tax

The attention tax is the part I pay myself. The coordination tax is the part the system pays, and it is larger, and it falls on other people.

Here is the pattern the field experiments keep finding. AI makes you faster at the parts of your job you can change alone: the writing, the summarizing, the first draft, the search. In one randomized study across dozens of firms, the clearest effect was that people spent meaningfully less time on email. But the same study found the broader shape of work mostly unchanged, because the rest of the job runs on coordination, and coordination did not get faster. Meetings did not shrink. Approvals did not speed up. Handoffs and ownership and the review bottleneck all sat exactly where they were. The individual got faster. The system did not, because the system was never limited by how fast any one person could draft.

And then there is the version where the local speedup makes the system slower. Stanford and BetterUp gave it a name that is hard to forget: workslop, polished AI output with nothing underneath it, passed to a colleague who then spends real time discovering it is hollow and redoing the work. Around forty percent of desk workers reported receiving some in a single month, costing roughly two hours each to sort out. Look at the arithmetic of one incident. The sender saved ten minutes generating a clean-looking draft. The receiver lost two hours. The company is down an hour and fifty minutes and it shows up as productivity, because the sender's ten-minute win is the only part that got measured. The cost moved downstream, out of sight of the person who created it, which is exactly why the behavior keeps repeating. I have written about owning the word slop and what separates craft from waste in machine-made work. Workslop is that waste, dressed up to look finished and handed to someone else to find out it is not.

This is the real reason the individual can feel transformed while the organization stays flat. The gains are local and the costs are distributed. One person banks the time saved. Another person, somewhere down the chain, pays it back with interest, and no ledger connects the two.

What the bill actually says

So when someone reduces the cost of AI to the subscription plus the tokens, they are reading the cheapest line and ignoring the meter. The honest bill looks more like this:

subscription, plus context preparation, plus verification, plus rework, plus the cleanup of work that looked finished, plus the trust you have to rebuild every time the confident output turns out to be wrong.

None of that is an argument against the tools. I would not give them up, and the time they save me on the doing is real even after the tax. It is an argument for honesty about where the money goes, because you cannot manage a cost you refuse to name. The financial line is the one the finance team can see. The attention and coordination lines are the ones that decide whether you actually came out ahead, and they are the ones the real bill for AI is mostly made of. Count them, or keep wondering why the productivity you can feel so clearly never shows up in the numbers.