The more we automate, the busier everyone gets

A document arrives from a colleague. The formatting is immaculate, the grammar faultless, the headings sensible. You start reading and something is not right, though it takes a paragraph or two to identify what. The observations are true but weightless. The recommendations would apply to any organisation in any sector. And section three, on closer inspection, contradicts section one.

Researchers at BetterUp Labs and Stanford's Social Media Lab gave this a name in 2025, and workslop is a good one: work that has the appearance of completion and none of the substance, which generates more effort downstream than it saved upstream.

Downstream is the word doing the work here. This material does not eliminate labour. It moves it, from the person who was supposed to do it to the person who receives it, and the transfer is invisible because nobody logs the hours spent working out what a document was meant to say.

The cost lands on someone with no way to bill it#

Follow what actually happens next.

The recipient has to determine what is wrong, which takes longer than finding an obvious error, because the surface gives no clues. They have to work out what was actually required, which the sender may never have established. Then they either fix it or start again.

If they have the standing to send it back, there is an awkward exchange and a delay. If they do not, and plenty of people do not, they absorb the hours quietly and it never appears anywhere. Multiply across an organisation and you have a productivity paradox that will not show up in any report: adoption metrics climbing, everyone visibly busy, and less actually getting finished.

The research put a substantial figure on the annual cost of this, and the number is arresting, but I find the mechanism more useful than the total. Money that goes missing invisibly is much harder to fix than money that goes missing loudly.

Almost nobody sending this is being cynical#

This is where I want to be careful, because the easy version of this argument is contemptuous and also wrong.

The person sending workslop usually believes they have done a decent job. They lack the expertise to evaluate what came back, so they cannot see the flattened logic, the missing context or the confident claim with nothing under it. What they can see is that it looks like the good documents they have received in the past, and looking right is the only signal available to them.

Someone who does not understand financial modelling cannot assess whether a generated model is sound. Someone without a marketing background cannot tell whether a campaign strategy is coherent or merely fluent. The tool produces output at a level of polish that exceeds the recipient's ability to judge it, which is a new problem, because previously the polish and the substance travelled together.

And the vending machine habit makes it worse. Type request, take output, send onward. No iteration, no interrogation, no second pass. The difference between that and a properly worked exchange is the difference between workslop and something genuinely useful, and nobody ever explained that to most of the people now expected to use these tools daily.

The leadership part, which is the actual problem#

MIT's Project NANDA reported in 2025 that around 95% of enterprise generative AI pilots produced no measurable return. That figure gets waved about as evidence the technology is empty. I read it as evidence of something more specific and more fixable.

We handed extremely capable tools to people, told them adoption was expected, and provided a webinar link. Then we set usage as a target.

That last part is the one that should worry any board. Measuring AI usage without measuring output quality mandates the production of workslop, in exactly the way that measuring a surgeon by counting incisions would produce a great many incisions. People respond to what is counted. If usage is counted, usage is what you will get, and quality will be somebody else's problem in a different department.

I have some sympathy for how this happened. Boards were told they were falling behind, adoption was the visible proxy for progress, and visible proxies are what get reported upward. But the accountability sits squarely with leadership rather than with the person who sent a poor document, and any framing that lands the blame on individual staff has misread it entirely.

What competence would actually require#

Not restriction. That ship sailed, and besides, the tools are genuinely valuable when used properly.

It requires teaching people how these systems actually work: that they predict plausible continuations rather than retrieve facts, that they will fill a gap rather than admit one, that fluency and accuracy are unrelated. None of that needs mathematics. It needs about two hours and someone willing to explain it honestly.

It requires standards for what constitutes acceptable AI-assisted work, and those standards have to include a person who understands the domain having actually evaluated the output. Not approved it, evaluated it, which is a different verb.

And it requires making it safe to say "I am not sure I can judge whether this is right". At the moment that admission looks like incompetence, so nobody makes it, so the material circulates. An organisation where people can say it is one where workslop stops at the first desk rather than the fifth.

The alternative is what a lot of places are living now. Impressive adoption statistics, flat productivity, and a shared unspoken suspicion that something is not adding up. The answer is sitting in everyone's inbox, beautifully formatted, and contradicting itself on page two.