Someone on your team is using AI to draft a campaign, they feel faster, and they are about to spend the next hour quietly fixing what it gave them: reloading context, correcting a claim, rewriting the lines that read like a spec sheet.
That hour has a name now. It is called botsitting. And the research says your team is doing far more of it than anyone is admitting.
Botsitting: the art of feeding AI context, checking outputs, fixing mistakes, and cleaning up the confident-but-wrong answers AI leaves behind.
The Glean Work AI Index 2026 surveyed 6,000 full-time knowledge workers across the US, UK and Australia. Here is the finding that should stop you.
Workers save around 11 hours a week with AI, then hand nearly half of it straight back.
Globally that runs to 6.4 hours a week per employee to botsitting. In the US, 36% of all the time people spend with AI goes on botsitting, and 36% admit shipping work they have not properly checked.
Everyone feels faster. Nobody moves.
Here is the paradox underneath it. Across the same research, 87% of workers use AI and 75% feel more productive. The share of organisations actually performing significantly better is thirteen percent.
That gap is the whole story.
A much-cited study from MIT's NANDA researchers put a harder edge on it. After tens of billions in enterprise spend, around 95% of generative AI projects show no measurable impact on the bottom line. The individual feels the lift, the P&L doesn't show it. Somewhere between the two, the gains get eaten.
Where the hours actually go
Some of it is workslop or AI slop: AI output that looks finished and is not. Researchers at BetterUp Labs and Stanford found that each piece of it costs the recipient almost two hours to sort out, and an estimated nine million dollars a year in lost productivity for a ten-thousand-person company. Upwork found 77% of workers say AI has added to their workload.
This pattern is consistent no matter the geography. In Germany, two-thirds of knowledge workers now use AI weekly and fewer than one in five organisations has scaled it past the experiment. McKinsey's 2025 State of AI found 88% of companies using AI and only 39% reporting any profit impact at all. The botsitting is structural, and it is everywhere.
Your AI never met your buyer
Ask why the hours pile up and you reach the same place every time. The model has no grounding in your actual market. It has read the internet. It has not read your buyer. So it produces something plausible and confident, and your marketer becomes the person who drags it back to reality.
The Glean data makes this almost embarrassingly clear. Workers in organisations that feed AI real context are 52% less likely to ship work they cannot stand behind, and they spend less of their time botsitting at all. Grounding is the lever, everything downstream of it is correction.
In life sciences the problem has a precise shape. In our own surveys of life science exhibitors, data quality is the number one barrier to getting value from AI, holding firm at 44% across two events, ahead of governance, skills and budget. Poor data is poor grounding. Feed the model thin, generic inputs and it gives you thin, generic output, then your team buys it back in hours.
Your buyer is built to catch it
For life science marketers this lands harder than for most. Your buyer is a scientist, trained to find the unsupported claim and pull on it. AI-generated copy aimed at that audience is the worst possible bet, because the reader is selected for catching exactly the thing the model gets wrong.
Generic confidence is what an ungrounded AI produces. It is also the fastest way to lose a sceptical scientist.
There is a fair objection here. Maybe botsitting is just early friction and part of the learning curve. The tools are new, the workflows fresh, and the supervision burden falls as everyone learns. Perhaps in eighteen months the time spent on botsitting drops.
However, the Glean research suggests that botsitting won't go away fast, as the highest performers botsit the hardest. They spend a greater share of their AI time checking, correcting and adding context. The difference is that they aim it well. They know when to keep AI out of a task, and they almost never give up and do the whole thing by hand. The loop is the discipline that makes AI worth having.
So the answer runs the opposite way from cutting AI back. Give the model something real to stand on, and put a human on the decisions that matter - particularly those that impact customer experience or your brand.
That is the whole argument for grounded synthetic customers and keeping the buyer in the loop. The buyer as live evidence the AI is grounded in, so it stops guessing and your team stops correcting.
Go back to that person on your team, an hour into fixing a draft. They are doing the only job the AI left them. They are being the grounding it never had.
