John McIntyre
Founder, HotPMO
Every so often a report hands you a word you didn’t know you needed. This summer’s word is botsitting, and once you’ve heard it you will see it everywhere that AI is in use in your organisation
It comes from the Work AI Index 2026, a survey of 6,000 digital workers across the UK, US and Australia by Glean’s Work AI Institute, with academics from Stanford, Berkeley and UCL among the authors. Botsitting is their name for the unrecognised, unbudgeted, untracked labour of making AI usable: feeding it the context it’s missing, supervising its output, debugging its mistakes and cleaning up when it’s confidently wrong.
Workers report saving around eleven hours a week through AI. And that would be great, were it not for the fact that they also report spending 6.4 hours a week botsitting: 2.3 hours feeding context, 2.2 hours checking output, 1.7 hours debugging. So most of a working day, every week, spent minding the machine, and 36% of AI sessions fail outright anyway.
Across all the time people spend with AI, 37% goes on botsitting against 36% actually producing work. We now spend as much time supervising the tool as using it.
Those hours appear in nobody’s business case, and they certainly don’t appear in the hype pieces about AI on social media. They just eat the savings the business case promised. If you’ve been wondering why your organisation’s AI spend keeps rising while delivery looks much the same, then surfacing botsitting time is a good place to start.
For the strategic version of that puzzle, I’ve written separately about the AI productivity paradox and what an industrial revolution can teach us about it in a separate article.
From botsitting to botvomiting
Botsitting is at least honest work. The problem is what people do when they tire of it, and the report has a ruder word for that which I won’t repeat here – I’ll rename it to a more accurate description: botvomiting, which is defined as shipping AI-generated work you haven’t verified, don’t fully understand, or couldn’t defend if asked.
69% of AI users admit to doing it. It scales with usage: half of light users admit it, 82% of heavy users. 41% say they have delivered AI work they couldn’t explain. And when that work fails, heavy users are 3.4 times more likely to blame the tool.
Think about that in a project context. Unverified content flowing into status reports, business cases and board papers, read by people who have no idea nobody checked it. The report describes a cycle: AI gets deployed, botsitting load rises, people tire, corners get cut, unverified output moves downstream, cleanup piles up, and the organisation responds by deploying more AI. Round and round, faster each time.
Why good people ship unchecked work
It is tempting to label botvomiting as laziness. But it isn’t that simple. Harvard Business Review published research in May on the psychological costs of adopting AI, surveying over 1,200 employees in the UK and US. The finding that matters here is what the researcher calls competency debt. AI hands you a polished, confident answer in seconds, and that output carries a sense of clarity you often don’t have in your own work. Felt capability rises while owned skill erodes, so leaning on the tool gets easier every week. Add in models trained to agree with whatever you bring them, and the checking muscle atrophies precisely when it matters most.
As a leader of a PMO, you can quickly find yourself in a situation where your team are churning out convincing looking reports that they do not understand, leaving your credibility exposed, while your team’s competence declines. Â
Competency debt, as HBR describes it: the two dials move in opposite directions.
One more HBR finding worth highlighting: psychological debt was almost twice as high among people who rarely use AI (a score of 60) as among fluent daily users (36). The anxious, arms-length, occasional user carries all the dread and none of the fluency. Half-hearted adoption of AI in the PMO turns out to be the most expensive kind.
Four moves for PMO leaders
First, measure it. Ask your team to log botsitting hours for one month: time spent feeding, checking and correcting AI. Put your number next to the 6.4-hour average and you have the beginnings of a business case nobody can wave away.
Second, impose context discipline. The Work AI Index found context is the dividing line: workers in context-rich organisations were 64% less likely to feel worn out by AI and half as likely to ship unverified work, while 53% of workers say the information their AI needs is locked somewhere it can’t see. Five minutes briefing the tool before a task changes what comes back. That habit costs nothing and it’s the cheapest fix on this list. In the medium term – work with peers in Info tech to explore how AI models can safely gain direct access to data that would improve quality and their usefulness.
Third, make verification active rather than passive. Skimming AI output for errors is exactly the mode where your brain switches off. Compare instead: write your own recollection of the meeting actions, or your own view of the risk position, then set it against the machine’s. The points of disagreement are usually where the problems live.
Fourth, name the accountability. KPMG’s Global AI Pulse Q2 2026 found only 24% of organisations can name who is accountable for decisions made using AI, while those with clear accountability report established ROI at three times the rate of those without. Someone must own what “verified” means before AI output moves downstream. In most organisations the natural owner of that standard is the PMO.
None of this needs a tool purchase. It needs somebody to treat AI working practice as a process worth designing, which is a sentence PMO people have been waiting years for the rest of the business to say. If you want working materials for the four moves above, our PMO Success Hub is where we keep them.
Sources
- Work AI Index 2026, Glean Work AI Institute.
- Harvard Business Review, “The Psychological Costs of Adopting AI”, May 2026.
- KPMG Global AI Pulse, Q2 2026.
