Abdul Samad SiddiquiSenior AI Engineer · Karachi

The doorman fallacy in AI strategy

Rory Sutherland's doorman fallacy explains why AI rollouts that only automate the measurable part of a job keep backfiring, and what CBA's reversed layoffs show about it.

ai strategyorganizational designcase study

Line drawing of a faceless hotel doorman under a canopy, with suitcase, umbrella, clock and key icons orbiting his head on a dashed line.

A hotel doorman’s job, on paper, is to open a door. That is the part you can time, measure, and, if you are building a business case for automation replace. It is also almost none of what the job actually is.

The ad man and author Rory Sutherland calls this the doorman fallacy: the mistake of defining a role by its most visible, most measurable task, then assuming that if you automate that task, you’ve automated the role. A real doorman hails cabs, remembers regulars’ names, spots trouble before it starts, makes a tired guest feel like someone noticed them walk in. None of that shows up in a job description that says “opens doors.” All of it disappears the moment you replace him with a motion sensor.

Why this keeps happening with AI

The doorman fallacy is not a historical curiosity. It is the default failure mode of AI-driven restructuring, because AI is extremely good at automating the legible slice of a job, the part that can be written down as a script, a macro, or a prompt, and organizations are structurally inclined to mistake that slice for the whole.

A clean example: in 2025, Commonwealth Bank of Australia (CBA) announced that 45 customer-service roles would be made redundant, citing an AI voice bot it said had reduced call volume. Within weeks, the bank reversed the planned cuts, apologised to affected staff, and kept the roles. The measurable task, answering routine calls, was real. What got missed was everything the role absorbed around that task: de-escalation, judgment calls on edge cases, the slack capacity that keeps a service desk from falling over the moment volume spikes or the bot gets confused.

This is not an isolated incident. A 2025 survey by workforce analytics firm Orgvue polled 1,163 senior leaders. Among the subset who said they had made AI-related redundancies, 39% of respondents, 55% admitted those decisions turned out to be wrong. That is not “55% of companies regret using AI.” It is a narrower, still uncomfortable number: of the leaders who actually pulled the trigger on AI-driven layoffs, more than half now think they got it wrong. Worth stating precisely, because the headline version of this stat tends to get rounded up into something scarier and less accurate than what the data says.

What the measurable slice hides

The doorman fallacy shows up any time someone frames a role as a single function instead of a bundle of responsibilities with fuzzy edges. A few patterns I’ve seen repeat:

  • Support agents don’t just answer tickets; they triage severity, spot patterns across tickets that no single ticket reveals, and build the institutional memory that turns a new policy change into something the team actually understands.
  • Recruiters don’t just screen resumes; they read between the lines of a candidate’s hesitations, manage a hiring manager’s unrealistic expectations, and keep a pipeline warm during the weeks nobody is responding.
  • Junior engineers don’t just write code to spec; they ask the “wait, why are we doing it this way” questions that catch bad requirements before they ship.

An AI system can take a real, meaningful bite out of each of these jobs. It cannot yet absorb the whole bundle, and treating “can do the legible 60%” as “can replace the role” is exactly the gap that burned CBA.

Decompose the task, not the role

The useful move is to stop asking “can AI do this job?” and start asking “which tasks inside this job are legible enough to hand to AI, and which ones depend on judgment that doesn’t compress into a script?” That is a task-level question, not a role-level one, and it tends to produce a different answer: augmentation of a role, not replacement of it.

In practice that looks like:

  1. Map the role into tasks, not responsibilities. “Handles billing disputes” is a responsibility; “looks up the transaction, checks it against three policy exceptions, decides whether to refund” is a task you can actually evaluate for automatability.
  2. Separate the measurable tasks from the judgment calls. AI handles volume and consistency on the former. Humans handle the exceptions, escalations, and the 20% of cases that don’t fit the pattern, which is usually where the job’s real value sits anyway.
  3. Watch what the role absorbs informally. The doorman hails cabs because cabs need hailing and he’s standing there, not because it’s in his contract. Those informal absorptions are often the first thing to break when you remove the person and keep only the narrow function.
  4. Pilot before you cut. CBA’s reversal created avoidable uncertainty for staff and cost public goodwill. Running the voice bot alongside the team first would have made the missing work easier to see.

None of this is an argument against using AI to change how work gets done. It is an argument against believing you’ve automated a role because you’ve automated its most countable task. The doorman still opens doors. He also does the rest of the job nobody wrote down, and that’s usually the part a sensor can’t do.

Sources and further reading