Before the photocopier, there was carbon paper.
Every letter that mattered got a second sheet stuffed behind it, smudged and slightly blue at the edges. An extra copy, for the file, in case anyone ever came looking.
They called it a carbon copy. We still call it that now. We just shortened it to CC.
Then, in 1959, Xerox launched a machine that could copy a page in six seconds flat.
Before it, Americans made about 20 million photocopies a year. Not exactly overrun with paperwork.
By 1966, that number was 14 billion. That is overrun with paperwork.
Copying a memo suddenly cost next to nothing. So people copied everything, whether it needed copying or not. Articles, white papers, anything they reckoned everyone else ought to see, pushed onto more desks than anyone could ever get through.
One office worker at the time put it simply: “You’d have this huge pile of meeting documents, and nobody has read them.“
That was 1966.
Then it was reply-all, that same old “CC” still riding along in the header. Then it was the PowerPoint deck nobody read either, the one a US Marine general stood up and called “PowerPoint makes us stupid.”
Every time, the same pattern. The cost of producing something drops to nearly nothing. The cost of a human reading, judging, and deciding what to do with it doesn’t change at all.
We’ve built this machine before. We just keep giving it a new name.
Swap “carbon paper” for “AI,” and it’s last week.
One of our leadership team told me that she couldn’t see the wood for the trees any more.
Not because she had too little help.
Because she had too much.
Three separate reports had landed on her desk. Same topic. All AI-assisted. All from people genuinely trying to help.
All needing reading, reconciling, and turning into something you could actually decide on.
She said even Claude couldn’t get her through the pile fast enough. Which, given the pile was largely Claude’s doing in the first place, is a bit rich.
The promise was speed
Every vendor pitch says the same thing. Faster drafts. Faster research. Faster code.
Gartner reckons 40% of enterprise apps will carry task-specific AI agents by 2026, up from under 5% the year before. That’s not gradual. That’s a whole phase change in about eighteen months.
We’re only just getting round to asking the question that actually matters.
Faster at what, exactly? And who’s meant to be catching what comes out the other end?
The bottleneck was never typing
Generation was never the hard part.
Judgment was.
Someone still has to sit down, read the draft properly, decide if it’s actually right, and reconcile it with the other three drafts that say something slightly different.
AI made the first part almost free.
It did sod all for the second part.
Make one half ten times cheaper and leave the other half exactly as expensive, exactly as human, and you don’t end up with more output.
You end up with a longer queue.
There’s a name for this now
Stanford and BetterUp call it “workslop”. Rather a good bit of naming, that.
Work that looks finished. Reads like it’s finished. Isn’t, not really, not if you look closely.
It gets bounced back. Reworked. Quietly absorbed by whoever’s unlucky enough to be left holding it.
They put a figure on it: $9 million a year in lost productivity for the average large organisation, almost all of it burned on people cleaning up after each other.
HBR reckons it’s corrosive to trust as well, not just to output. Which tracks. Nobody sends you a bad report on purpose. They send you a perfectly plausible one, and you’re the poor sod who has to work out which bits of it to actually believe.
Progress or a Queue?
If you’re rolling AI out across a team right now, stop asking how much faster you can generate.
Ask who’s actually going to sit down and read it all.
Ask whether you’ve made their life easier, or just tipped more onto the pile.
That’s the only question that counts, and most of us aren’t asking it.
The bottleneck was never typing.
It was thinking.
AI didn’t shift the bottleneck an inch.
It just built a bigger queue in front of it.
Further reading:
- How the Photocopier Changed the Way We Worked, and Played — Smithsonian Magazine
- We Have Met the Enemy and He Is PowerPoint — Small Wars Journal
- Workslop: The Hidden Cost of AI-Generated Busywork — BetterUp Labs
- AI Workslop Is a $9 Million Issue: Stanford, BetterUp Study — Entrepreneur
- Harvard Business Review warns AI ‘workslop’ is rotting companies from the inside
- Gartner Predicts 40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026
- Tyk API Management Platform
