Which Doors Should Machines Walk Through?

Jeff Bezos gave managers a simple test for how much care a decision deserves. Some decisions are two-way doors: walk through, dislike what you find, and walk back out at little cost. These should be made quickly, by individuals or small teams, because caution buys nothing. Others are one-way doors: irreversible, or nearly so, and worth slow deliberation and senior sign-off. His warning was that growing organisations tend to apply the heavyweight, one-way process to everything, and grind to a halt as a result.

The Clean Mapping To Automation

The framework transfers naturally to a question every organisation now faces: where should we let AI act on its own? The mapping is clean. Reversibility lowers the cost of error; AI errs at some rate; so route automation to wherever mistakes are cheap to undo. Fully automate the two-way doors, and keep humans stationed at the one-way ones. As a default posture this is sound, and it captures the real logic of Bezos’s idea rather than merely borrowing its vocabulary. Both frameworks sort choices by the cost of being wrong and match the amount of caution to that cost.

Automation Changes The Door

But the analogy does something Bezos never had to consider, and the interesting part lives in that gap. In his version, the door type is a fixed property of the decision. Introduce automation and the actor starts changing the door. A single automated pricing change, refund, or moderation call is trivially reversible. Ten thousand of them, executed before anyone looks, can move a market, drain a budget, or feed corrupted data into a downstream model. A door that is two-way at human scale can become one-way at machine scale, because automation changes the volume and speed at which the decision is made. Reversibility must therefore be judged at the throughput the system will actually run at, not one action at a time.

A second effect compounds the first. A two-way door is only useful if someone notices they should walk back through it. Manual processes carry natural friction that surfaces problems; full automation strips out the human who would have spotted the drift. The door stays technically reversible while becoming practically one-way, because detection lag lets damage accumulate before anyone reacts.

Reversibility Is Engineered, Not Given

This is where a fair objection sharpens the point rather than defeating it. One might say these failures describe negligent automation, not automation as such. A competent system has rollback, logging, rate limits, and alerting built in, so the aggregate stays reversible by design; ten thousand bad calls are undone by reverting a configuration and issuing refunds, which is precisely what keeps the door two-way. This is right, and it yields the essay’s central refinement. Reversibility is not a fixed feature of the decision but a property you engineer. The door stays two-way for as long as the build keeps it two-way, and no longer.

The Real Axis Is Reversibility Plus Detection

That reframing changes the deployment rule. The axis for confident automation is not reversibility alone but reversibility paired with fast, cheap detection. Where both hold, automate fully. Where a decision is reversible but errors are slow or costly to notice, let AI act behind a human checkpoint or a circuit breaker rather than hands-off. The engineering, rollback, monitoring, and rate limiting is not overhead bolted onto the framework; it is the thing that determines which door you are standing in front of.

Using AI To Build New Doors

There is also a prize hiding on the other side. Because AI is cheap to run, it does not only get routed to two-way doors; it can manufacture them. Cheap simulation, reversible staging environments, quick experiments, and drafts a human approves all convert what would have been an irreversible commitment into a reversible test. Used this way, AI turns one-way doors into two-way ones, which is arguably the larger opportunity than simply automating the reversible decisions that already exist.

The Practical Rule

The practical guidance, then, is narrower and more honest than “automate aggressively wherever things are reversible.” Automate confidently where reversibility is engineered and errors are quick to detect. Add a human where reversibility holds but detection is slow. Deliberate, as Bezos did, where the door is genuinely one-way. And wherever possible, use the technology to build more two-way doors rather than merely to speed through the ones already open. The framework survives the translation to AI, but only once we accept that the machine is not just choosing a door. It is reshaping which doors exist.

If your team is deciding where to let AI act on its own, a fixed-fee advisory session is built around exactly these calls: which decisions to automate, which to put a human behind, and how to engineer the reversibility and detection that make the difference. Get in touch to discuss.