contributors
Terry Peters
Managing Director, Brand Experience
July 16, 2026 • 5 Min Read
An agent is right 95% of the time. That number sounds reassuring until you think about what happens to the person reviewing its output over the following months.
In the first week, the reviewer reads carefully. They check the logic, verify the inputs, flag anything that looks off. By month two, they’re moving faster. The agent has been right every time they looked closely, and looking closely takes time. By month four, they’re approving output they haven’t fully read. By month eight, they can’t quite remember what a wrong answer looks like because they haven’t seen one in a long time. At that point, the 5% failures start landing on customers, regulators, or balance sheets, unfiltered. The org chart still says there’s a human in the loop. But to what end?
Human-in-the-loop degrades predictably and for structural reasons.
Supervisor drift isn’t a character flaw. It’s what happens when you put a person in a review role without designing the role around the act of finding things wrong. In a model that centers around approval, a reviewer who clears a hundred items a day with no flags looks like a high performer. And if the incentive is speed and volume, over time, that incentive will win.
Adversarial review is a different posture, not a different policy.
Ultimately, human-in-the-loop degrades because the model was designed for approval, not for scrutiny. One way to fix it is by flipping the model on it’s head, making it adversarial. Reviewers are scored, not on what they clear, but on what they catch. This changes the question they ask themselves from “Does this look right?” to “Where is this wrong?”
Injected failures keep the skill sharp when the agent keeps performing well.
Organizations already understand this principle in cybersecurity. We’ve all received simulated phishing emails made to look real. These aren’t designed to catch people making mistakes. They’re designed to keep people from losing the ability to recognize one.
Injected failures are a structural solution to a structural problem. But the instinct behind them — that human capability degrades without deliberate design — applies far beyond the review queue.
The same logic holds for how junior employees develop judgment when agents are doing the visible work. It holds for how organizations discern which decisions actually require a human, and which roles exist to produce output versus develop the people who will one day catch what output misses. In each case, the question is the same: did someone design this, or did it just emerge from whatever the technology left behind?
The right answer lies not just in AI, but in having the people who know what to do with it.
Redesigning workflows so that human oversight is real rather than documented is a process problem before it’s a governance one. That’s where the work starts.