Forty percent of U.S. desk workers say a colleague sent them "workslop" in the past month: AI-generated work that looks polished but doesn't hold up. Each one took an average of one hour and 51 minutes to untangle, according to research by BetterUp and the Stanford Social Media Lab covering 1,004 full-time workers.
The researchers put the cost at about $186 per employee per month. For a 10,000-person company, that's over $9 million a year in lost productivity. None of it shows up as an AI expense. It shows up as someone else's afternoon.
This is the side of the AI productivity story that rarely makes the pitch deck. Generating work got close to free. Checking it didn't.
The Bottleneck Moved
MIT economist Christian Catalini has a sharp name for where this ends up. When AI makes execution cheap, he argues in a new HBR piece, verification becomes "the scarce, valuable capability in organizations." The companies that win, in his framing, will operate as "verification factories."
It's a useful reframe because it explains a pattern many teams already feel. Drafts arrive faster. Decks, briefs, and code multiply. But the number of people qualified to say "this is right" hasn't changed — and those people are now buried in review.
The BetterUp data shows where the load lands. Managers were more likely than individual contributors to report receiving workslop: 54% versus 38.5%. The people with the least spare time are absorbing the most checking.
What Unchecked Output Actually Costs
The hours are only part of the story. In the same research, recipients felt annoyed (54%), frustrated (46%), and confused (38%). About half saw the sender as less creative, capable, and reliable, and 42% as less trustworthy. Nearly one in three said they'd be less likely to work with that person again.
So unchecked AI output doesn't just waste time. It quietly spends the sender's reputation — and 53% of workers in the study admitted some of their own output might qualify.
Outside the office, the bill gets larger and more public. Last year Deloitte agreed to partially refund the Australian government for a A$440,000 report that contained references to academic papers that don't exist and a fabricated quote attributed to a federal court judgment.
In the courts, researcher Damien Charlotin's tracker had logged roughly 1,490 decisions worldwide involving AI-hallucinated material by May 2026, more than 1,000 of them in the U.S. Sanctions have climbed from $5,000 in the landmark Mata v. Avianca case to $15,000 per attorney in a 2026 Sixth Circuit ruling.
Those are firms with professional review processes. The errors still got through, because the process was built for a world where producing a document took longer than checking it.
Why "Just Review It" Doesn't Work
The instinct is to tell people to double-check AI output. That fails for a structural reason: review takes the same scarce expertise the AI was supposed to save, and it doesn't scale with output volume.
If a junior analyst can now produce five reports in the time it used to take to write one, the senior reviewer doesn't get five times more hours. Either quality drops, or the reviewer becomes the bottleneck, and the productivity gain evaporates. Usually it's some of both.
That's why verification has to be designed into the work rather than bolted on at the end. Catalini's "factory" metaphor is apt: factories don't inspect quality in at the loading dock. They build it into every station.
What This Means
Four practices turn verification from a drag into a capability.
- Make the sender own the check. The single biggest lever in the workslop data is upstream. Set the norm that anything AI-assisted is reviewed by the person who generated it before it goes out — and that "the AI wrote it" is never an excuse. Most workslop is a handoff problem, not a technology problem.
- Tier your review by stakes. Not everything needs the same scrutiny. Internal notes can go out with a light read. Anything client-facing, legal, financial, or published gets a named reviewer and a source check. Spending senior attention evenly across all output is how the high-risk mistakes slip through.
- Verify claims, not prose. AI is fluent; fluency isn't the risk. The failures in the Deloitte and courtroom cases were invented facts: citations, quotes, numbers. Build a habit of checking every figure, name, and reference against its source, and let the writing quality take care of itself.
- Measure the checking time. If review hours are rising faster than output is improving, your AI rollout has a hidden cost. Track it the way you'd track any other operating expense.
The teams that get the most from AI won't be the ones that generate the most. They'll be the ones whose work can be trusted the first time it's read. In a world where anyone can produce a polished draft in seconds, being right is the differentiator.
Discussion