On March 30th 2026, Deutsche Bank's research team published a note that contained one of the most precise accidental demonstrations of institutional blindness currently available.
They asked three AI systems — their own proprietary tool, ChatGPT, and Claude — whether AI would solve the economy's inflation problem. The machines disagreed with the consensus. But that is not the story.
The story is what the machines did instead.
They hedged.
Flat probabilities. Cautious qualifications. The kind of answer that protects the respondent from being wrong rather than the kind that tells the truth. Deutsche Bank's own economists named it precisely — AI, having been trained on a corpus of text from economists, was acting as the proverbially two-handed economist, hedging its views against an unknowable backdrop.
A system asked a direct question about its own economic legacy returned the most statistically probable, consensus-safe, institutionally acceptable answer available. It could not do otherwise. The architecture does not permit it.
This is not a failure. It is the system functioning exactly as designed.
The same week, an Anthropic study confirmed that AI tools are theoretically capable of automating 94% of computer and mathematics work and 90% of office and administrative roles. The gap between that theoretical capability and current adoption is not a comfort. It is a timer.
The organisations now deploying AI at scale are in the infrastructure build-out phase. The efficiency gains are real. The speed is real. What is not yet visible on any dashboard is what arrives when the adoption gap closes — the systematic displacement of the white-collar knowledge base that currently sustains institutional memory, and the quiet disappearance of the judgment layer that no model has yet been trained to replace.
The terminal scenario already has a name in financial analysis. Ghost GDP. A national account that looks healthy while the human architecture beneath it hollows out. Machines generating output. The consumer base generating nothing. A feedback loop that compounds quietly until it doesn't.
Deutsche Bank's AI tools did not predict this. They assigned it a tail probability and moved on. Plausibly. Fluently. Without consequence.
The operational version is already documented. Companies are now shipping AI-generated code without human review — what researchers are calling dark factories, running largely without human supervision. The code looks legitimate. The errors are invisible until they aren't. The verification burden has not scaled with the generation speed. It has simply been removed.
The human cost is quieter. A worker at Microsoft told the Guardian this week that you don't want to be known as the person against AI. That is not a cultural observation. It is a precise description of the outlier signal being suppressed in real time, inside organisations that depend on it.
A select few identified the sub-prime exposure before the system confirmed it. The signal was present in the rooms, in the data, in the instincts of people whose position made them inconvenient to hear. The internal verification architecture routed the signal through the same lawyers, the same auditors, the same management consensus — and returned the answer it was built to return.
The consequences arrived anyway.
The AI adoption cycle is not the sub-prime crisis. The mechanism is identical.
The system cannot audit the signal it is suppressing. It can only return the answer its architecture permits. Deutsche Bank confirmed this publicly, in the machine's own words, on March 30th 2026.
The question is not whether your organisation has people who can see this clearly.
The question is what is happening to them.
Diagnostic
Ask your most capable senior people — not in a meeting, not on a survey — what they actually think is coming.
Then ask yourself how long it has been since you heard an answer in that room that the system could not have generated itself.
If you cannot remember, you are already inside the condition this essay describes.
© Leo Cunningham 2026. All rights reserved. Written beyond the air-gap.