No one is sure, but the nineteenth-century British Prime Minister Benjamin Disraeli is said by some to have been first to declare that the three types of falsehood are lies, damned lies, and statistics. The renowned American writer Mark Twain (who famously used the phrase rather a lot) put it on Dizzy, for one.

I like this attribution because it would mark a venerable beginning to the astonishing ability of British politicians to survey the people and prove – inevitably – that public opinion supports their own agenda, no matter what it is.

I noticed this back when I was underwriting, in relation to issues that affected the job. The classic came in the debate over joining the euro. Supporters and opponents alike regularly referenced opinion polls that found near-universal support for their position.

Around the same time, the great British public was asked if they want mandatory ID cards. One poll found both that 80% support the plan, and that 58% don’t. The same poll! On ending the quaint English practice of the unspeakable pursuing the inedible on horseback with dogs, polls showed a majority in favour of banning hunting, and a majority against.

It’s fantastic fun to see the same twisting of survey findings when businesses are asked about AI.

A company called Sixfold (self-described as “underwriting AI that already knows the job”) surveyed 543 senior underwriting personnel in the US, UK, and EU who are using (71%) or piloting (29%) AI. Perhaps unsurprisingly, this AI company’s survey found respondents overwhelmingly believe that AI improves productivity, decision quality, and processing capacity.

A huge majority of 86% expect a positive return, as underwriters are freed to focus on “complex risks, relationship management, and higher-value judgement”. Alas, even the AI boosters admit that adoption barriers remain, including concerns about AI accuracy, integration challenges, data quality, insufficient training, and tools that do not fit underwriting workflows.

Contrast a survey released the same week by Glean, published through its pocket boffin group the ‘Work AI Institute’. Multi-sector Glean claims to “understand your business and what your business needs.” Their survey, The Work AI Index 2026: Global, is sub-titled “Botsitting, botshitting, and the hidden human labor of AI at work”. It presents a very different picture to that rendered by Sixfold.

They surveyed of 6,000 full-time digital workers in the US, UK, and Australia who said AI saves them 11 hours per week on average. However, the research found that th majority of their remarkable productivity gain is squandered on AI.

Of digital workers’ total AI interaction time, 37% goes to botsitting (described as “grunt work such as reloading context into different tools, catching hallucinations, and verifying outputs that sound confident, or, worse, flatter workers with the answers they wanted to hear instead of what’s true”). Another third (36%) is spent actually using the tool, and 27% on learning tools and building agents.

“Bottsitting” leads (forgive me) to “botshitting”. This Glean describes as “the act of shipping AI-generated work that employees haven’t verified, don’t fully understand, or can’t confidently stand behind.” Depressingly, they survey found that 69% of AI users admit to this, and states that heavy users are “64% more likely to botshit”.

The survey revealed three AI Paradoxes. First, of the workers surveyed, 75% said AI makes them more productive and 63% that it lets them do things they couldn’t do before, but 77% have corrected or redone AI-assisted work recently, and 30% do it every week (suggesting that perhaps they could have done it right the first time, all on their own).

The second paradox is that AI makes oversight more important, and strips away the cues that trigger a review of human outputs. “When everything AI produces looks polished, the appearance of the work gets decoupled from the substance,” the survey report explains. “Next is a slow surrender of agency… as the pile [of outputs] grows, [people] start to skim it, then wave it through.”

The final paradox shows a digital workforce trapped between dual AI threats: being replaced by AI, or looking obsolete. “Standing still doesn’t protect your expertise. It paints a target on your back”, the Work AI Institute concludes from the data. “So workers double down.” The unfortunate result is AI’s absorption of the tasks people find meaningful, in exchange for more botsitting.

Glean says the solution is to “treat AI as a work-design problem”. I’ve been arguing for process redesign since my first Diary entry. I also agree with the Sixfold survey finding that with AI adoption, underwriters are freed to focus on “complex risks, relationship management, and higher-value judgement”.

It takes me back to an article about the “complex realities” of AI adoption published in the Harvard Business Review back in February. Researchers reporting there found that using generative AI tools “didn’t reduce work”. Instead, the people they tracked for eight months reported that AI “consistently intensified” their workload. They worked faster and for longer on a broader set of tasks, typically on their own initiative. But when the glorious surge of productivity began to feel to them just like more work, they devolved to “lower-quality work, turnover, and other problems.”

The researchers’ prescription is for organisations to slow down, take pauses, sequence work deliberately (rather than reacting to each output), and – perhaps most important of all – ground AI work in human interaction by protecting time and space for listening and human connection.

These recommendations are based on what is essentially just another survey (albeit one led by a professor, and sufficiently lengthy and intensive to count as proper research), but it slips nicely over the other two surveys to deliver a nicely rounded picture of AI use and implementation.

I like this, because all three surveys support what I have been arguing here about AI for ages: it’s no panacea, but if used correctly, it can deliver superb benefits.

* Like every Insurance Technology Diary entry about AI, this one is accurate only to the best of my knowledge at the time of writing. The pace of AI progress is so great that I cannot guarantee it remains so now that it’s finished, let alone when you read it.

Guillaume Bonnissent is CEO of Quotech.