The Arrow of Time: The AI Decisions You Can Never Undo

Part III of III. Governance ranks AI decisions by risk, and lately by value. It almost never asks which ones can be undone.

I. Time’s arrow

In 1928 the astronomer Arthur Eddington gave physics a phrase it has used ever since: “time’s arrow.” He meant the one-way character of time, which has no equivalent in space. You can walk back across a room; you cannot walk back into yesterday. A dropped glass shatters and never reassembles itself, because entropy never decreases.

Organisations obey the same law, though we rarely notice. Most decisions a company takes can be reversed: a price changed, a supplier swapped, a feature rolled back. A few travel only one way.

Part I of this series, The Risk of Everything, asked whether AI governance can say what a system is worth. Part II, Special Relativity, asked whether it can act at an agent’s speed. This final part asks the question boards ask least: which of our AI decisions can never be taken back?

II. One-way doors

Jeff Bezos gave business its most useful version of Eddington’s idea. In his 2015 letter to shareholders he described Type 1 decisions as “consequential and irreversible or nearly irreversible,” one-way doors that deserve slow and careful deliberation. Type 2 decisions are two-way doors: if you dislike what you find on the other side, you walk back.

He warned against mistakes in both directions. Organisations that apply the heavy Type 1 process to every decision become slow and stop experimenting, which is the argument of Part II. Organisations that rush Type 1 decisions with the lightweight Type 2 process, he suggested, eventually go extinct.

Regulators already understand this. When the EU AI Act sets out how to decide whether an AI system is high-risk, one of its criteria is whether the harm it could cause is easily corrected or reversed. Yet boards rarely apply the same test to their own AI decisions. A risk register records likelihood and severity. A value register, where one exists, records size and timing. Neither records permanence, so a decision that can never be undone travels through the same approval path as one that could be reversed next quarter.

Reversibility is the missing dimension. Risk, value and speed describe a decision at the moment it is made; reversibility describes what you are left with afterwards. And it is rarely binary. It is a cost curve: some doors swing freely, some open only for a fee, and some close behind you.

An office door at the end of a dark corridor opens onto open sky and a sheer drop — Some doors swing both ways. Some open onto nothing.

III. Four doors that close

Here are the AI decisions I see enterprises take most casually, and regret most expensively.

1. Closing the junior pipeline. Earlier this week, in Paid to Disappear, I wrote about what workers lose beyond pay when AI takes their work. There is a governance cost too. Stanford’s Digital Economy Lab finds that employment for 22 to 25 year olds in the most AI-exposed occupations is now about 19% below where it would be had it kept pace with less exposed peers, while experienced workers show no comparable gap. The authors are careful to call this descriptive rather than causal. But the direction is clear, and a company that stops hiring graduates because an agent can now do their tasks is taking a decision whose consequences arrive a decade later. The people who will supervise, audit and overrule tomorrow’s agents learn that judgement by doing junior work today.

This is also a collective-action problem. Any single firm can stop training juniors and hope to poach trained people later. If every firm does it, there is nobody left to poach.

2. Dismantling a capability. On a spreadsheet, rehiring looks like a two-way door. In practice it is a door with a toll. Klarna, which moved aggressively to AI customer service, began bringing humans back in 2025 after its chief executive conceded that cost had been “a too predominant evaluation factor.” It rebalanced rather than reversed; headcount has still fallen from about 7,000 to 3,000. In August 2025, Commonwealth Bank of Australia reversed 45 job cuts tied to an AI voice bot after a challenge from the Finance Sector Union, admitting the roles had not in fact been redundant. Gartner now predicts that by 2027, half of the companies that cut staff because of AI will rehire for similar functions, under different job titles.

So the role can be refilled. What is much harder to recover is the tacit knowledge: the exceptions an experienced agent knew to escalate, the customers who trusted a particular voice, the informal workarounds nobody documented. Those leave with the people, and the people rarely come back. The cost does not only fall on the firm. Steven Davis and Till von Wachter found that men displaced in mass layoffs lose the equivalent of 1.4 years of earnings in good times, and 2.8 years when unemployment is high.

3. Data trained into a model. Here the board needs one distinction. Data held in a retrieval system, a database the model consults, can be removed. Data trained into a model’s weights largely cannot. A paper by researchers from Stanford, Cornell, Microsoft Research, Google DeepMind and others concluded that machine unlearning “is not a general-purpose solution” for controlling what a generative model does. A January 2026 survey from Georgia Tech notes that supposedly forgotten information may re-emerge through fine-tuning or adversarial prompting. Approving a fine-tune on customer or proprietary data is a one-way door that often looks like a configuration setting.

4. Agent actions with external effects. Part II was about agents deleting things inside the firm. The harder category is what agents do outside it: a payment sent, an email delivered to a customer, a contract term accepted, a regulatory filing submitted. Once those leave the building, no rollback in your own systems can recall them. Any agent permitted to act externally is permitted to walk through one-way doors on your behalf, at machine speed.

There is a fifth door worth watching, one that closes slowly rather than all at once. Andreessen Horowitz’s 2025 survey of 100 enterprise CIOs found that agentic workflows were already making it harder to switch between models; one CIO admitted every prompt had been tuned for a single provider. From January 2027 the EU Data Act stops most switching charges between data processing providers, which lowers the contractual cost of leaving. It does nothing for the engineering cost of rebuilding thousands of prompts, evaluations and agent behaviours around a different model.

IV. A second key for one-way doors

In Part II I argued that an agent should never be able to delete a production volume without a second key it does not hold. The same principle belongs at board level.

Add a reversibility field to every entry in the risk and value registers from Part I, and let it change the approval route.

Classify before approving. Define the one-way doors explicitly. For example: any AI-attributed restructure affecting more than a set number of roles, the retirement of a human capability, a fine-tune on customer or proprietary data, any agent authorised to act externally, and any single-provider commitment above a set level of spend. These go to a different approval path.
Require a reversal cost. The approval paper states what it would cost, in money and months, to undo the decision in two years. An honest “we do not know” is useful information too.
Keep a seed team. Retain a minimum human capability in every function you automate: a small group that still does the work by hand, trains successors and could scale back up. A value register will flag this as cost without return, so price it explicitly as an insurance premium against a rebuild.
Rehearse the rollback. As with Part II’s five-minute drill, test it safely: a tabletop exercise, or one team for one day working without its AI tools. What you discover is your real reversibility.
Fund the ladder collectively. Because the junior pipeline is a shared problem, the answer is partly shared too: industry apprenticeships, graduate pools funded across firms, and public schemes such as those Singapore is expanding through SkillsFuture.

This costs something. A seed team is a line on the budget, and a separate approval path adds a few weeks to a handful of decisions each year. What it buys is speed everywhere else. Once everyone knows which few doors are one-way, the many two-way doors can be walked through without a committee.

V. Voices on the record

These are public positions taken by members of my usual panel and their critics, not responses to this essay.

Erik Brynjolfsson warned in The Turing Trap that when machines substitute for people rather than augmenting them, “workers lose economic and political bargaining power,” and an economy can settle into an equilibrium that is hard to escape. He was describing a one-way door for a whole society.

Daron Acemoglu insists that the direction is chosen, not given. “The direction of technological change is something we determine socially and economically,” he told McKinsey in 2021. A reversibility rating is one way a board can make that choice deliberately.

Yoshua Bengio chairs the International AI Safety Report, whose 2026 edition observes that open-weight models “cannot be recalled once released.” Frontier labs face the starkest one-way door in the field; enterprises face smaller versions of it every week.

The counter-argument deserves a full hearing. In February, before Stanford’s August update, Martha Gimbel of the Yale Budget Lab argued that AI was not yet showing major macroeconomic effects on jobs, and that some companies were blaming AI for cuts with other causes. David Autor and colleagues have shown that about 60% of American employment in 2018 was in types of job that did not exist in 1940. Time’s arrow points towards new work as well as away from old work.

I accept both points. New work will come. My concern is narrower: it rarely arrives for the same people, in the same places, at the same stage of their careers, and a board approving an automation decision today is not accountable for the economy of 2040. It is accountable for the door it is about to walk through.

VI. A prediction

By the end of 2028, at least one Fortune Global 500 company will disclose a material cost to rebuild a function it previously automated with AI. When it does, the line item will be called something like “capability restoration”, and the board that approved the original decision will discover it never recorded the decision as irreversible.

VII. The trilogy test

Three parts, three questions to ask of any material AI decision before it reaches the board.

Risk and value: can we say what this system is worth as precisely as we can say what it might cost us?
Speed: can we stop it in five minutes, and do our controls act at its speed rather than ours?
Direction: if we are wrong, can we come back through this door, and what will the return journey cost?

Heisenberg and Einstein overturned our intuitions about the universe. Eddington named one that is true and that organisations keep forgetting: some things only go one way. The firms that come through this transition well will not be the ones that avoided one-way doors. They will be the ones that knew which doors they were walking through.