Botsitting and botshitting: the AI adoption numbers nobody wants to say out loud

Eighty-seven percent of digital workers use AI at work. Seventy-five percent say it makes them personally more productive. Only thirteen percent say their organisation is performing significantly better as a result. That gap, between individual productivity and organisational uplift, is the single most important number in commercial AI transformation right now, and most rollout plans I've seen don't account for it at all.

A 2026 survey of six thousand workers across the US, UK, and Australia, run by the Work AI Institute with Stanford, Berkeley, Notre Dame, UCL and others, put a name to the gap. Two names, actually, and neither is flattering.

Botsitting is the unrecognised labour of making AI usable: feeding it context, supervising its output, debugging what it got wrong. The survey measured it at 6.4 hours a week per worker, split roughly across feeding context, supervising outputs, and cleanup. Thirty-six percent of AI sessions fail outright and need a full restart. Workers who botsit frequently are seventy-three percent more likely to be actively job hunting, which tells you something about how this labour actually feels day to day: it doesn't read as using a powerful tool, it reads as unpaid maintenance work nobody budgeted for.

Botshitting is the sharper problem, and the one I think most transformation programmes are structurally unprepared for. Sixty-nine percent of AI users admit to shipping work they haven't verified, don't fully understand, or couldn't defend if asked. Forty-one percent deliver AI-generated work they couldn't explain. Twelve percent knowingly ship output they believe is wrong. Twenty-eight percent have blamed the AI for a mistake that was actually theirs. Heavy AI users are sixty-four percent more likely to botshit than light users, meaning the people using the tools most aggressively are also the people most likely to be quietly passing bad work downstream.

The mechanism the survey names for why this happens is worth sitting with, because it isn't obvious. AI removes the disfluency cues that used to trigger scrutiny. A messy draft with typos used to signal "read this carefully." Polished AI output doesn't carry that signal any more, and reviewers, including the person who generated the output, read it faster and more trustingly precisely because it looks finished. The tool got better at looking done. It didn't get correspondingly better at actually being done, and the gap between those two things is where botshitting lives.

Here's the finding that should reframe how anyone designs a rollout: the thirteen percent of organisations getting real uplift are the ones that invested in what the survey calls human infrastructure, not tool deployment, and not simply the ones with the best tools. Concretely: they measure quality alongside productivity, not just speed. They build governance that's actually reviewed and enforced, not a policy document nobody reads. They ground AI in enterprise context deliberately, so the model knows which file is current and which workaround is actually live, not just what's technically searchable. And they spread adoption peer-to-peer rather than by mandate. Cross-functional teammates showing each other real workflows drove 5.6 times more adoption than top-down rollout announcements did.

The context finding is the one I keep coming back to, because it's the most actionable and the most commonly skipped. Workers whose AI genuinely has access to the information they need, context-rich in the survey's terms, show markedly lower rates of every bad outcome: less exhaustion, less weekly cleanup, less shipping-work-they-can't-explain, less using unapproved shadow tools, less hiding their AI use from managers. Context-poor workers show the opposite on every single measure. This is the difference between a rollout that solves the actual bottleneck and one that just adds a chat window on top of the same information-scattered reality everyone was already working in.

What I take from this, working on commercial AI transformation day to day, is that the tool selection conversation is almost never the hard part any more. The models are genuinely capable. The hard part is whether the surrounding system gives them the context to be right, whether anyone is actually checking the thirteen percent who are shipping wrong output on purpose, and whether the six-plus hours a week of invisible labour keeping the whole thing running is recognised as work, or quietly absorbed by the people already closest to burnout. Adoption numbers looked, for a while, like the whole story. They're not even the interesting half of it.