Irreversibility: The Only Sane Way to Allocate Executive Attention
AI work is disposable and regenerable. What compounds are the commitments you can't take back. Route senior judgment by one-way doors, not by cost.
In his 2015 letter to Amazon shareholders, Jeff Bezos drew a distinction that most executives nod at and almost none actually organize around. Some decisions, he wrote, are consequential and nearly irreversible: one-way doors. Walk through, dislike what you find, and you cannot get back to where you were. Others are reversible: two-way doors. You can walk through, look around, and walk back out at little cost. His point was about speed: two-way doors should be decided fast, by individuals or small teams, while one-way doors deserve deliberation. But there is a second, sharper implication he left mostly unstated, and it is the one that matters now that a machine will produce most of your work. The right axis for spending scarce executive attention was never difficulty or dollar size. It is reversibility, because AI makes almost all hard work cheap to redo, and leaves the small set of things you can’t take back as the only place senior judgment still earns its keep.
What follows: why “route attention by importance” quietly means “by dollar size” and why that is wrong, why AI widens the gap between reversible and irreversible until it becomes the only distinction that matters, and how to re-sort your calendar around it.
The strongest version of “focus on the big, hard, expensive decisions”
The conventional discipline is to spend senior attention where the stakes are highest, and stakes are usually read as size: the biggest budgets, the hardest problems, the decisions with the most zeros. Concede that this is not stupid. Big, hard, expensive decisions often are consequential, and a leadership team that ignored them would be negligent. The nine-figure acquisition, the reorganization, the platform rebuild: these deserve scrutiny, and the instinct to give it to them is sound as far as it goes.
Where it fails is that size and difficulty correlate only loosely with what actually determines outcomes: whether the decision can be undone. A large, difficult, expensive project that is fully reversible (a market test you can shut down, a product line you can discontinue, an ad campaign you can pull) is far less deserving of your scarcest attention than it looks, because if it goes wrong you redo it. Meanwhile the decisions that quietly compound are often small and cheap in the moment and permanent in effect. A single sentence in a contract. A price you publish. A position you take publicly. A key person you hire, or promise equity to. A dataset you hand a partner. A commitment you make to a regulator. None of these has the dollar-size signature that draws executive attention, and every one of them is a one-way door. Route attention by size and you will lavish scrutiny on reversible bets while irreversible ones slip through on a junior sign-off.
So the honest version of the steelman is this: big decisions deserve attention, but “big” was always a proxy for “consequential,” and the truer proxy (the one AI makes decisive) is “irreversible.”
The category error
The mistake is to keep sorting decisions by difficulty when AI has changed what difficulty costs. In a world where producing a plan, a model, a design, a campaign took weeks of expensive labor, hard-and-reversible decisions genuinely warranted heavy scrutiny, because redoing the work was itself expensive. That justification is gone. When the production is cheap and regenerable, a hard-but-reversible decision is one you can simply run again: hand it to the loop, see what happens, redo it if it’s wrong. The difficulty that used to command your attention is now the machine’s to absorb. What the machine cannot absorb, and what no amount of cheap regeneration can rescue, is a commitment that has already left the building.
So the two quadrants that used to blur together now separate cleanly, and they invert the usual priority. Hard-but-reversible: delegate to the loop and redo as needed; this is most of what looks important and deserves far less senior time than it gets. Easy-but-irreversible: this is where senior judgment is the only thing standing between the company and a permanent mistake, and it is routinely handled by whoever happened to be in the room, because it didn’t look big enough to escalate.
The public wreckage of ignored irreversibility is easy to find. The merger of AOL and Time Warner in 2001 was the largest in history at around $165 billion; within roughly two years the combined company wrote down about $99 billion of goodwill, and no amount of subsequent excellence could re-separate what had been fused. It was a one-way door walked through at speed. Reversibility cuts the other way too, and treating a two-way door as a one-way door is its own error, but the costly pattern is almost always the reverse: a permanent commitment mistaken for a reversible one because it was cheap or fast to make. Netflix in 2011 published a roughly sixty-percent effective price increase and split its plans; the market reaction was severe (it shed hundreds of thousands of subscribers and a large chunk of its market value) because a published price is far stickier than the keystrokes it took to set it. And the risk lives in dependencies as much as decisions: when a single vendor’s faulty software update took millions of Windows machines offline in July 2024, grounding flights and halting hospitals worldwide, the lesson was not that the vendor erred but that its customers had walked through a one-way door of concentrated dependence without noticing it was one.
Where the tooling comes in
The strategic move against irreversibility is to convert one-way doors into two-way doors wherever you can, to buy back reversibility on the commitments that would otherwise compound against you. In the AI stack specifically, the sharpest one-way door is dependence on a single model provider whose price, policy, or availability can change without your consent, as the neighboring essays on the half-life of a prompt and on provider substitution have argued at length. A gateway is the mechanism that turns that commitment reversible: AOCore routes across many providers behind one stable interface, with model aliasing and automatic failover, so that a provider choice becomes an operational parameter you can change rather than an architectural marriage you cannot. The point is not the routing feature. It is that reversibility is something you can engineer into your commitments deliberately, and every commitment you make reversible is one fewer place your scarce attention has to stand guard.
What to do
Re-sort your calendar on one axis: for every decision on it, ask whether it is a one-way door. Take everything hard-but-reversible and push it down and out: delegate it to the loop, set the redo trigger, and reclaim the hours. Take everything easy-but-irreversible, however small it looks, and pull it up: the clause, the published price, the public position, the key hire, the data you share, the promise to a customer or a regulator. Those are where you personally belong, precisely because they will never come back for a second decision. And before you commit to any of them, ask whether you can make it a two-way door instead (pilot the price, scope the hire, structure the contract with an exit, abstract the vendor) because a commitment you can walk back is a commitment that no longer needs to consume you.
Difficulty is the machine’s problem now. Dollar size was always a bad proxy. Spend yourself on the doors that only open one way, because everything else, you can redo.