In the winter of 1924, locomotive engineers on the long-haul freight lines began carrying a very specific kind of luggage: a heavy, lead-lined brick.

The trains were being fitted with “dead man’s switches,” a safety innovation designed to cut the engine if the driver became incapacitated. To keep the train moving, the engineer had to keep a foot firmly pressed on a stiff, spring-loaded pedal for the duration of the journey.

It was a perfect safety policy on paper. In practice, it was an ergonomic nightmare that caused permanent nerve damage in the legs of the men responsible for thousands of tons of moving steel. So, they used the bricks. They bypassed the safety mechanism to do the job they were actually hired to do: get the freight to the station on time.

The policy protected the company from the liability of a fainted driver; the brick allowed the driver to survive the shift.

The Tuesday Morning Shriek

Tuesday morning, 9:14 am. Hugo is standing by the office coffee machine, the kind that grinds the beans with a mechanical shriek that makes conversation impossible. It’s a mercy, really. He’s staring at his phone, specifically an all-staff email with a subject line that feels like a cold damp cloth: Updated guidance on generative AI.

There are three bullet points. They are masterpieces of legal vacuum-sealing. They forbid the input of confidential data into “unauthorized third-party linguistic models.” They mention “disciplinary action.”

They conclude with a sentence about how employees are “solely responsible” for ensuring their use of external tools adheres to the company’s non-disclosure agreements.

Hugo watches the espresso drip. He has a 42-page market analysis deck due at noon. He knows, with the weary soul-deep certainty of a man who has done this for nine years, that the manual synthesis of the three conflicting datasets required for that deck will take six hours. He has two hours and forty-six minutes.

He forwards the email to exactly nobody. He doesn’t reply asking for a list of “authorized” tools because he already knows that list is a blank sheet of paper.

Eleven minutes later, back at his desk, he opens a browser tab on his personal account, his phone’s hotspot providing a jittery but private umbilical cord to the internet. He pastes the sanitized-but-still-sensitive data into the prompt box. The deck will be finished by 11:30 am.

The policy has been followed in the only way the organization actually cares about: the paper trail is clean. If a data leak ever occurs, the company can point to the 9:14 am email and say, “We told him not to.”

Cognitive Vertigo

I’ll admit, I yawned during a similar briefing last month. Not because I’m lazy, but because the gap between the speaker’s mouth and the reality of the cubicle has become so wide it creates a sort of cognitive vertigo. We are being told to fight a fire with a set of instructions that forbid the use of water.

Most people read these AI policies as “risk reduction.” That is the corporate marketing. But if you look closer, they are actually a sophisticated form of “risk relocation.”

When an organization issues a blanket prohibition without providing a functional alternative, they aren’t stopping the behavior. They are simply moving the “event” of the behavior from the corporate ledger to the individual’s conscience.

The person who drafted that email is measured on whether a policy exists, not on whether any human being can actually work under it.

Carlos and the Safety Slant

I think about Carlos K.L. sometimes. He’s a precision welder I met years ago, the kind of guy who works on high-pressure pipelines where a single pinhole leak is a catastrophe. Carlos once explained to me the “Safety Slant.”

There’s an official way to hold the torch, dictated by a safety manual written by someone in an air-conditioned office in Omaha. If you hold it that way, you can’t actually see the puddle of molten metal because the handle blocks your line of sight. To get a perfect, x-ray-quality bead, you have to tilt your head and angle your hand in a way that technically violates the ergonomic safety code.

“If I follow the rule, the weld might fail. If I break the rule, the weld is perfect but my career is at risk if an inspector sees my grip.”

– Carlos K.L., Precision Welder

He chose the perfect weld every time. He took the personal risk to ensure the structural integrity of the pipe. Hugo is doing the same thing with his market analysis deck. He is choosing the “perfect weld”-the finished, high-quality work product-at the expense of his own compliance record.

It isn’t built by rebels or corporate spies. It is built by your most productive employees. The people who actually care about the deadline, the quality of the output, and the survival of the department are the ones currently violating the 9:14 am memo.

They are the ones using personal accounts, personal devices, and home Wi-Fi to access the tools they need to stay competitive in a world that has already moved past the “manual synthesis” era.

The organization gets the paperwork it wanted. The legal department sleeps soundly, wrapped in the warm blanket of “we warned them.” But the actual data-the precious, confidential, proprietary lifeblood of the firm-is still leaving the building.

In fact, it’s leaving through a door that nobody is watching now. By pushing the behavior underground, the policy has stripped away every trace of oversight. There is no logging, no encryption standard, no corporate governance. There is only Hugo and a hotspot.

This is the quiet mechanism by which governance and reality separate. When the rules are impossible to follow, they merely become a weapon to be used after the crash.

The irony is that the technology to bridge this gap already exists. The problem isn’t that AI is inherently “leaky.” The problem is that the interface between the professional and the model is often built for the consumer, not the fiduciary.

If the organization won’t provide the tool, the professional finds their own way. This is why platforms like

tunnel AI

exist-not as a rebellion, but as a bridge between the duty to perform and the duty to protect.

These systems act as a privacy-first gateway, ensuring that the work gets done through an encrypted, anonymous layer that satisfies the “perfect weld” and the “safety manual” simultaneously. It removes the need for the lead-lined brick.

We have to stop pretending that “Don’t” is a strategy. In a professional environment, “Don’t” is just a way of saying “Do it where I can’t see it.”

I’ve seen this play out in every industry. In healthcare, it was doctors using unapproved messaging apps to share patient photos for quick consultations because the internal paging system was a relic of the 1980s.

In law, it was associates using personal cloud storage because the firm’s server had a file-size limit that made modern discovery impossible. In each case, the policy was a shield for the organization and a trap for the employee.

The path forward isn’t more memos. It’s the realization that productivity and security are not on opposite ends of a seesaw. They are the two rails of a single track. If you pull them apart, the train doesn’t just slow down-it derails.

Systemic Dishonesty

The most dangerous person in your company isn’t the one who ignores the AI policy. It’s the person who wrote a policy that makes it impossible to do the job.

That person has created a systemic incentive for dishonesty. They have forced the “Hugos” of the world to choose between being a good employee and being a compliant one.

When we look back at this transition period, we won’t be surprised by the data leaks. We will be surprised by how long we expected people to hold down the spring-loaded pedal with their bare feet while the freight was already an hour late.

We will wonder why we didn’t just fix the switch. The deck reaches the finish line on the strength of a secret, while the policy stays behind to guard a vacant office.

Effective governance requires more than a signature on a PDF. It requires an acknowledgment of the physics of the modern workplace. If the workload assumes the presence of AI, then the policy must provide a safe path for that AI to exist.

Anything else is just risk relocation-a game of hot potato where the most dedicated employees are the ones most likely to get burned when the music stops.