Two surveys describe the two ends of the same pipe
Neither one mentions the other. One was fielded on three continents by a company that sells enterprise AI search. The other was fielded in the United States by a research collaboration with Stanford. They point the same way, from opposite ends, with no reason to agree.
At the sending end: across 6,000 full-time digital workers in the United States, United Kingdom and Australia, 69% admitted to shipping AI-assisted work they had not properly reviewed or did not fully understand. At the receiving end: across 1,150 American desk workers, 40% said they had been handed AI-generated material in the past month that looked finished and was not, and put the time spent sorting out each instance at close to two hours.
Two readings of two quantities, not two readings of one
These are not two measurements of the same thing. One is people admitting what they send. The other is people reporting what they receive. Which is exactly what makes them worth putting side by side — they were built to agree with nothing, and they agree anyway.
The cost of producing a document fell enormously in about two years. The cost of reading one did not move at all. It cannot — it is bounded by a person's attention, and no tool has changed that. So the ratio between what it costs to send and what it costs to absorb has been quietly rewritten, and every extra unit of output has to land on somebody.
A different problem from the one most teams are working on
Checking your own output is a discipline question and it is being discussed everywhere. This is an accounting question. The person who saved the twenty minutes and the person who spent the two hours are not the same person, and only the first of them has any reason to mention it.
Which is what makes it stable. Nobody is behaving badly. The sender genuinely finished faster. The receiver genuinely cannot tell, from the outside of a polished document, whether the thinking behind it happened. The organization measures output. Output went up.
The useful question, and it is not the one that gets asked
The useful question is not whether AI saves you time. You already believe it does, and you are probably right. It is whether anything you sent this month arrived somewhere needing work you were certain you had already done. That is a question your team can answer more reliably than any survey can, because the receipts sit in their inboxes and in yours.
Once you can count what is arriving at the receiving end of your own team, the "AI saves us hours" figure can be read against something. Right now it is not being read against anything, because half the ledger has never been opened.
Half the ledger has a cost you can put a number on: See what the uncounted half is costing your organization →
Keep reading: Nobody checks what AI hands back · Leaders using AI visibly can cost trust
This is the thinking behind a Trusted AI Culture.
Sources: Glean Work AI Institute, "Work AI Index 2026" (6,000 digital workers, US, UK and Australia, December 2025 to January 2026). BetterUp Labs with Stanford Social Media Lab (1,150 US desk workers, September 2025), reported in Harvard Business Review. Both are self-reported, and the two figures measure different populations answering different questions.
Written with AI in the loop: my idea, AI drafted and sharpened, my judgment on the way out. Every word is mine to stand behind. — Darren