Is the Door Just a Door? AI, Jobs, and the Doorman Fallacy

There is a new genre of headline, and you have read a version of it this month. AI can already replace 11.7% of the U.S. workforce. 40% of jobs exposed. Half of all tasks automatable by 2030. The numbers vary, the energy does not: a study lands, a percentage detaches from it, and the percentage goes for a walk around the internet without its caveats.

I want to take these studies seriously — they are not nonsense — and then point at the specific assumption almost all of them share. It is an old assumption with a good name: the doorman fallacy. Once you see it, you cannot unsee it in the AI jobs debate.

What the “X% of jobs” studies actually measure

Start with the most-cited recent example. MIT’s Iceberg Index made news in late 2025 for finding that 11.7% of U.S. wage value — about $1.2 trillion — is “technically exposed” to current AI. The method is genuinely careful: take the government’s O*NET database, decompose 923 occupations into some 32,000 distinct skills, map which of those skills today’s AI tools can perform, and weight each by the wages attached to it.

Read the method and the trick reveals itself — not a dishonest trick, just a structural one. The study measures skills, then sums them up into a number about jobs. To do that, it has to assume a job is its list of catalogued skills. Anything a worker does that never made it into O*NET as a discrete, nameable skill is, by construction, invisible to the model.

To their credit, the Iceberg authors are explicit that this is a ceiling, not a forecast: the index “measures technical exposure — where AI capabilities and human skills overlap — not displacement outcomes,” and “does not predict job losses, adoption timelines, or net employment effects.” That honesty mostly evaporates by the time the number reaches a headline. (I pulled this one apart in more detail separately — the short version is that “exposure” and “replacement” are very different words.)

So hold the structure in your head: these studies break a job into its measurable parts, score the parts, and add them back up. What could go wrong with that? Exactly one thing — and it has a name.

The doorman fallacy

The term comes from Rory Sutherland, the Ogilvy vice chairman, in his 2019 book Alchemy. Picture a five-star hotel that hires a consultant to find savings. The consultant watches the doorman and reaches the obvious conclusion: this man opens a door. A door can be opened by a $5,000 automatic mechanism. Fire the doorman, install the door, save a salary every year forever.

Two years later the hotel is worse in ways the spreadsheet never predicted. Because the doorman was never really opening the door. He was recognising regulars by name, hailing taxis in the rain, sizing up who should and shouldn’t walk in, defusing the drunk guest before he became a scene, signalling — by simply existing in a good coat — that this is the kind of place that has a doorman. The open door was the cheapest and least important thing he did. It was just the only part that was easy to see.

That is the doorman fallacy: grounding your understanding of a role in its most visible function, while being blind to the tangible and intangible value sitting just under the surface. Sutherland’s deeper point is about what he’d call measurability bias — we privilege what is easy to count (salary saved, doors opened, tickets closed) and quietly assign a value of zero to what resists counting (trust, judgement, institutional memory, the feeling of being looked after). Zero is rarely the right number. It is just the convenient one.

Why AI is the perfect doorman-fallacy machine

Here is the uncomfortable fit. The task-exposure studies are, structurally, doorman-fallacy engines. They can only see a job as the sum of its listed skills, because that is the only form in which a job is available to a model. The tacit surround — the part that made the doorman worth keeping — isn’t in O*NET, so it contributes nothing to the percentage. The studies are not lying. They are measuring the open door very precisely and calling it the doorman.

And AI makes the trap far easier to fall into than the automatic door ever did, for one reason: AI is superb at reproducing the visible output and silent about everything it isn’t doing. An automatic door obviously cannot hail a taxi; nobody is fooled. But an AI support agent produces fluent, confident, on-brand replies that look exactly like the visible output of a great human agent — right up until the moment a customer needs judgement, escalation, discretion, or someone who actually cares. The failure is invisible in the demo and shows up later, in the data, as churn.

We now have receipts:

• Klarna announced in 2024 that AI had done the work of roughly 700 customer-service agents, and the CEO praised it as human-equivalent. By 2025 the company was reversing course and hiring people back; CEO Sebastian Siemiatkowski admitted the relentless focus on cost had “lower[ed] the quality” and eroded customer trust. The AI scored well on the easy, high-volume tickets — the open door — and quietly degraded on exactly the complex interactions where a human agent earns their salary.
• The Commonwealth Bank of Australia cut 45 customer-service roles for an AI voice bot, then rehired them, conceding it hadn’t properly understood the work.
• Taco Bell rolled out drive-through voice AI and walked it back after customer complaints and failures at peak times — the moments when a flustered human is, it turns out, better than a confident machine.

In each case the company priced the door and forgot the doorman.

The counter-fallacy: sometimes the door really is just a door

Now the part that keeps this honest, because LazySmart is not in the business of swapping one comforting story for another.

The doorman fallacy can curdle into its own opposite — call it the doorman-fallacy fallacy: invoking “but there’s invisible human value!” to defend any role from any automation, forever. That is just measurability bias wearing the other team’s jersey. Sentimentality about the “human touch” can protect work that genuinely is mostly the visible task, the same way a tidy spreadsheet can erase work that isn’t.

Some doors really should be automatic. Plenty of office building entrances were once staffed and are now glass and sensors, and nobody mourns. A great deal of data entry, transcription, first-pass document review, and routine image processing is just the door — rule-bound, measurable, no tacit surround worth preserving — and automating it frees people for work that has one. The fallacy is not “automating things.” The fallacy is automating things without first asking how much of the role is the door.

So the useful question is never the one the studies answer (“can AI perform this task?”). It is the one they structurally cannot ask: is the task the job? For a transcriptionist, largely yes. For a hotel doorman, a senior nurse, a relationship manager, a good manager of people — emphatically no, and the gap between “yes” and “no” is exactly the value that doesn’t show up in O*NET.

So — is it the same with AI?

Yes and no, and the distinction is the whole point.

It is the same in that AI invites the identical mistake the consultant made about the doorman — only more seductively, because it imitates the visible output so well that the missing parts don’t announce themselves until later. Every “X% of jobs exposed” study is, whether it means to be or not, a measurement of doors. Used as a forecast of replacement, it commits the fallacy at population scale.

It is different in that AI’s reach into genuinely codifiable work is real and large, and pretending otherwise is its own fallacy. The technology is not a toy. Where a role is mostly door, AI will take the door, and that is fine.

The skill — the lazy-smart move — is refusing both reflexes. Don’t read “11.7% exposed” as “12% of people are toast,” and don’t read it as “nothing will change.” Read it as: here is where the pressure is; now go find out, role by role, how much of each one was ever really the open door. That is slower than reacting to a headline. It is also the only version of the analysis that doesn’t end with you quietly rehiring the doorman two years later.

Conclusion

The studies aren’t wrong; they are narrow, and narrowness becomes error the moment a percentage about tasks is read as a verdict about people. The doorman fallacy is what happens when you mistake the part of a job you can see for the job itself — and AI, brilliant at the visible and mute about the rest, is the most convincing version of that mistake ever built.

Before you replace anyone — or believe anyone who tells you a number of someones is about to be replaced — ask the only question that survives contact with reality: what would the automatic door actually cost you? If the honest answer is “nothing much,” automate it and move on. If you can’t answer, you haven’t found the doorman yet. Keep looking. He’s the reason the place feels the way it does.

Sources

1. Rory Sutherland, Alchemy: The Surprising Power of Ideas That Don’t Make Sense (2019) — origin of the “doorman fallacy.”
2. The Conversation (Gediminas Lipnickas, University of South Australia), “The ‘doorman fallacy’: why careless adoption of AI backfires so easily.” — the fallacy applied to AI, plus the Commonwealth Bank and Taco Bell examples.
3. MIT, “The Iceberg Index: Measuring Skills-centered Exposure in the AI Economy” (arXiv) — 11.7% of U.S. wage value technically exposed; “exposure, not displacement.”
4. Fortune, “Klarna plans to hire humans again.”
5. CX Dive, “Klarna reinvests in human talent for customer service.”