It is a Friday afternoon and I am approving things. A budget, a brief, a client note, each one landing in my inbox already finished, already formatted, already sure of itself. I read, I nod, I send. It feels like competence. It feels, in the moment, like precisely the job. The uncomfortable question arrives later, in the lull after the last email: when did I last actually disagree with one of these before it went out?
In March 2026, researchers published a study of 758 Boston Consulting Group consultants that had been running since 2023, and one half of it went straight onto LinkedIn.1 Consultants using GPT-4 finished 12% more tasks, worked 25% faster, and produced work rated 32% higher in quality. The other half travelled less well. The researchers had slipped in tasks they knew the model would get wrong, and on those the consultants with AI did not merely lose their edge, they underperformed the colleagues working without it. The machine handed them a confident answer. They took it. They moved on. Some of the sharpest consultants in the world stopped at the surface of the work and called it finished.
The researchers named the pattern cognitive surrender, and it is the workplace crisis nobody is putting on a slide.
The oversight that isn’t
For two years the reassuring line has held: the machine does more of the work, but the human stays in the loop. We review. We verify. We own it. A Wharton study in May 2026, with the lovely title Thinking: Fast, Slow, and Artificial, found people followed wrong AI answers 80% of the time when the answer arrived with confidence.2 When the AI was right, human accuracy leapt 25 points above its own baseline. When the AI was wrong, accuracy dropped 15 points below it. The presence of an answer was steering the human more than the human’s own experience was. The loop is there. The people are there. The oversight, often, is not.
The reviewer is tired
There is a plain mechanical reason for this. A BCG survey of 1,488 knowledge workers early in 2026 found 14% in a state they are calling AI brain fry, the particular exhaustion of monitoring and approving the machine’s output all day.3 Those workers carried 33% more decision fatigue than their colleagues, and their major error rate had climbed 39%. Productivity rose with one to three AI tools, then fell off a cliff at four or more. What that describes is a workforce generating more, overseeing more and thinking less, trained by hundreds of fine approvals to assume the next one is fine too.
What is actually being lost
The pillar under all of this is judgement, and judgement is an active practice: holding a question open when the pressure says close it, weighing accumulated experience against the evidence in front of you, refusing the answer that merely looks certain. Ethan Mollick named the cost in May, in an essay called Choosing to Stay Human.4 He calls badly-prompted AI writing meaning-shaped attention vampires, text that looks substantial and hands the reader nothing back. The same goes for AI analysis and recommendations. Work can look complete without being evaluated. A sign-off can look like judgement without being it. When you approve a strategy or a model or a client note, you are making a claim: I understand this, I stand behind it, this is right enough to act on. Cognitive surrender hollows the claim while leaving it standing. And when the thing goes wrong, nobody asks the AI.
The three-question test
The most useful thing a leader can do this year is decide what “reviewed” actually means, and then mean it. Three questions, asked with the honest intention of answering them:
- Can I explain the reasoning without opening the document? Not the summary. The actual logic: why this and not the alternative, what the load-bearing assumption is, what would have to be true for it to be wrong. If you cannot, you reviewed the format.
- If someone challenged a claim in this, live, in front of people, could I defend it? From your own understanding of the subject, not by reading the document back. The machine can generate the claim. Defending it is yours.
- Do I know what this got wrong, or where it is thin? Good output usually carries real substance and real gaps in the same breath. If you cannot find the gap, you have not done the work.
What to do this week
Find one piece of AI-generated work you approved in the last five days. Open it again, slowly, without the pressure that was on you the first time, and run the three questions. Mark where more presence would have caught something. Treat it as calibration, not a flogging. Then write one sentence for your team, in plain words, defining what a person has to be able to do before AI-generated work leaves the building. One sentence is the whole boundary between oversight and rubber-stamping.
Agentic AI is doing more unattended work this year than last, and by December the machines will be finishing hours of it before a human ever looks. Cognitive surrender in leadership is a design outcome, not a character flaw, and every system that rewards speed over scrutiny accelerates it. Mollick’s worry is the one I share: these defaults are being set now, without much planning, and they will be hard to reverse once a generation has built its habits around them. The efficiency is real. So is the gap it opens between the work and the person whose name is on it, the gap that stays invisible until the morning something breaks and someone asks who, exactly, stood behind this.
Sources:- Dell’Acqua et al., field experiment with 758 BCG consultants, Harvard Business School / Organization Science, March 2026[↩]
- Wharton, “Thinking: Fast, Slow, and Artificial”, May 2026[↩]
- BCG survey of 1,488 knowledge workers, 2026[↩]
- Ethan Mollick, “Choosing to Stay Human”, One Useful Thing, 26 May 2026[↩]
