One-way doors disguised as AI strategy
AI is the ultimate friction eraser. And the ultimate excuse for irreversible decisions.
TL;DR
The Paradox: Enterprise AI is sold as the ultimate friction solvent. In reality, deploying AI without friction does not create efficiency; it creates a high-speed liability engine.
The Mechanism: AI breaks Jeff Bezos’s classic “one-way vs. two-way door” rule. At scale, an AI agent’s execution speed silently mutates safe, reversible decisions (two-way doors) into catastrophic, irreversible outcomes (one-way doors).
The Action: Stop using AI to erase all friction. Strip cognitive drag from low-stakes workflows, but aggressively re-insert operational and accountability friction (rate limits, kill-switches) to survive looming EU AI Act enforcement.
Last week, the Educated Guess Substack published a brilliant piece - “Friction is A Feature”. It explains very well that a well‑placed friction is not an accident. It is a deliberate design choice that forces people to weigh tradeoffs before they act.
Earlier that week, I saw a sharp example of what happens when you ignore that rule. Oracle just executed a double-elimination of friction to fund its AI ambitions.
First, they attacked the financial friction. A 162,000-person payroll is a massive drag on margins. To fund an AI data-center build-out, Oracle needed capital fast. So they stripped the friction of up to 30,000 salaries out of their operating budget and moved that cash to infrastructure.
Second, they attacked the operational friction. They executed this massive, irreversible decision via a simultaneous 6 a.m. mass email. No manager conversations. No arguments. No time to say goodbye. They bypassed the agonizing, high-friction work of looking human beings in the eye.
In their boardroom, this looked like a frictionless masterstroke. A clean spreadsheet update to fund the future.
But instead of destroying the friction, Oracle simply relocated it. They traded a morning of discomfort for years of execution risk, moving the drag from their P&L directly into the surviving culture.
The same thing happens when teams deploy AI. They use it to wipe out friction in the wrong places, only to watch the damage explode downstream. They fail because they assume all friction is the same.
It is not. There are at least three kinds.
Three kinds of friction (and why AI keeps attacking the wrong ones)
"Friction" is a lazy word. We use it to describe anything that slows us down. But inside a company, you are managing three entirely different forces.
Cognitive friction
This is the mental drag of figuring out what a screen means. A dense dashboard. A confusing form. A button that transfers a million dollars but looks exactly like the button that refreshes the page. You actually want cognitive friction when the stakes are high to force the user to wake up and pay attention.
Generative AI removes this. It auto-fills the form, summarizes the text and pre-selects the choice. It makes thinking optional.
Operational friction
These are the checklists, the two-manager sign-offs, the manual data reconciliations. The tedious work you hate doing on a Friday afternoon. But this drag exists for a reason. It acts as a shock absorber. It prevents a bad input from instantly becoming a catastrophic output. When you deploy an AI agent, it bypasses this friction. It reads the email, changes the CRM status and triggers the client refund in three seconds. The shock absorber is gone.
Accountability friction
The audit logs. The model documentation. The human-in-the-loop review. For the last three years, product teams treated this compliance friction as annoying legal overhead. Now, the EU AI Act is making it a hard constraint. By August 2026, if you deploy high-risk AI without built-in friction - explainability, kill switches, documented oversight - you face fines up to 7% of your global revenue.
Why does AI keep attacking the wrong ones?
Because of incentives. Vendors sell speed. Executives buy efficiency. Cognitive and operational friction are easy to measure and eliminate. You can put “hours saved” on a slide. Accountability friction is invisible until you are audited or breached. So the market naturally builds high-speed machines with no brakes.
If you let your teams blindly erase friction, you end up with massive liability. To fix this, you need a rule for when to let the AI run and when to force it to stop.
By the way, this rule already exists. Jeff Bezos wrote it ten years ago.
One-way vs two-way door: what Bezos actually said
Ten years ago, Jeff Bezos laid out a decision framework for Amazon shareholders. He split decisions into two types.
Type 1 decisions are one-way doors. They are highly consequential and nearly impossible to reverse. Acquisitions, mass layoffs, shutting down a core product. You walk through, the door locks behind you. Bezos said these decisions demand heavy friction. You make them slowly, with deep consultation and deliberate process.
Type 2 decisions are two-way doors. They are reversible. UI tweaks, pricing experiments, testing a new vendor. If you walk through and it fails, you just walk back. Bezos argued these decisions require zero friction. Small teams should make them instantly.
When humans do the work, treating a two-way door like a one-way door is a fatal error. When companies wrap reversible experiments in heavy operational friction - bureaucracy, committees, consensus - they paralyze themselves.
But AI breaks the physics of this rule entirely.
If you map the three types of friction onto Bezos’s doors, the rule for AI deployment becomes obvious. You use AI to aggressively strip cognitive friction out of two-way doors. You let agents summarize, pre-fill, and draft options so humans can make reversible choices faster.
But you never let AI remove friction from a one-way door. For decisions that change legal exposure, safety, or irreversible system states, you use AI to analyze scenarios. You do not let it execute the action.
Here is the trap. Instead of simply walking through doors, AI actually mutates them.
If you let an agent apply a theoretically “reversible” two-way decision across ten thousand customer accounts in two seconds, you just created a one-way door. The blast radius, the reputational cost and the sheer mess of cleaning up the data make rollback practically impossible.
This is the paradox of operational friction. What paralyzes a human is exactly what tames an agent. To keep a two-way door reversible when using AI, you actually have to reintroduce operational friction in code. You need rate limits, batch approvals, and phased rollouts. Without that friction, the speed and scale of AI will permanently mutate your reversible experiments into systemic failures.
This is why regulators are forcing accountability friction back into the enterprise. They realize that AI at scale turns two-way doors into one-way disasters.
If you map this structural shift, the resulting matrix is the only AI governance chart that actually matters.
State 1: Pre-AI (Human Physics)
The problem: Everything is covered in friction. Humans treat two-way doors like one-way doors.
State 2: Post-AI (Agent Physics)
The shift: Cognitive friction drops to zero. But operational and accountability friction must spike to prevent agents from mutating two-way doors into one-way disasters.
The simple test for 2026 AI decisions
If you apply this matrix, your execution plan becomes simple.
You do not need a 50-page governance charter. For every AI-driven decision or agent deployment in your organization, force your product and risk teams to answer three questions:
What kind of friction are we removing—and where will it reappear?
If you are erasing cognitive drag from a low-stakes workflow, proceed. If you are erasing operational friction from a high-stakes workflow without adding blast-radius controls, you are building a liability engine.
Is this actually a two-way door, or does AI turn it into a one-way failure at scale?
Any AI deployment that pushes configuration changes, communications, or decisions to thousands of entities must be treated as a one-way door until proven otherwise. The blast radius dictates the door type, not the initial intention.
If a regulator or auditor shows up in 2027, where will they expect to see friction?
If you only “automate friction away,” regulators and reality will put it back in uglier places: sudden enforcement actions, public incidents, or reputational collapses. Can you point to the exact line of code where you deliberately re-inserted friction to satisfy the EU AI Act?
Here’s the most important takeaway:
AI is not here to magically erase friction. It is here to move friction away from busywork and toward responsibility.
We want fewer human cycles wasted on mindless tasks, and more friction around the moments where someone, somewhere, will have to sign their name under the outcome.
My final ask
These Signals reflect conversations I am having with executives right now, just written down.
If this helped you see your organization's blind spots more clearly, do two things.
Key Sources
Friction is a Feature, Oracle cuts up to 30,000 jobs globally, European Artificial Intelligence Act comes into force, Jeff Bezos’s 2016 Letter to Shareholders, Cognitive Friction Explained, Reducing Organizational Friction, Compliance Creates Friction Before it Creates Value