A hotel doorman opens a door. That is the part of the job you can measure, so it is the part that looks easy to replace. Install an automatic door and you have covered the function, cut a salary, and improved the line on the spreadsheet that someone asked you to improve.

You have also stopped hailing cabs for guests, lost the person who remembers the regulars by name, removed a quiet deterrent to anyone who should not be in the lobby, and erased a small signal that this is the kind of hotel that staffs a door. None of those showed up in the job title. All of them were part of the job.

The advertising executive Rory Sutherland named this pattern in his 2019 book Alchemy, and it has aged into one of the more useful ideas for anyone deciding what to hand to a machine. The trap, as a recent academic write-up on The Conversation puts it, is defining a role by its narrowest measurable function and then automating that function while the unmeasured value quietly walks out the door. The task was never the point. The task was the visible tip of the job.

This matters right now because AI has made the visible tip cheap to copy for almost any role in your business. Customer support replies. First-draft contracts. Appointment scheduling. Invoice follow-ups. The narrow, describable version of each of those is now a weekend project. The question is no longer whether the machine can do the task. It almost always can. The question is whether the task was the whole job, and that is a much harder thing to answer honestly.

The most cited example is already a year old

Klarna spent 2024 as the poster child for replacing people with AI. The company said its OpenAI-built assistant was doing the work of about 700 full-time agents and had handled roughly 2.3 million conversations in its first month, and it froze hiring on the strength of that. For a while it was the cleanest automation story in tech.

Then the value started leaking. By 2025 Klarna was publicly walking it back. CEO Sebastian Siemiatkowski told Bloomberg the company had leaned too hard on AI, that quality had suffered, and that Klarna would start hiring human agents again so customers always had a person to reach. Fortune reported the reversal alongside survey data showing most corporate AI projects were failing to deliver the returns leaders expected. The new model is not AI or humans. It is AI for the repetitive volume and humans for the moments that need empathy, discretion, or a judgment call.

Read that as a doorman story. Klarna defined support as "answer the incoming message," automated exactly that, and discovered the role had also been doing trust, de-escalation, edge-case judgment, and the basic reassurance that a frustrated customer was talking to someone who could actually fix the problem. The assistant answered messages. It did not do the job.

Why smart teams keep falling for it

The fallacy is sticky because efficiency thinking forces it. To justify automating a role, you have to write the role down, and writing it down means reducing it to its describable parts. The taxi-hailing and the name-remembering do not survive that translation, because nobody put them in the job description in the first place. So the business case looks airtight precisely because it is incomplete. You are not comparing the machine to the job. You are comparing the machine to a flattened summary of the job, and the machine wins that comparison every time.

There is a second pull. The measurable part is usually the cheap part, and the unmeasurable part is usually where the value lives. That is not a coincidence. Anything easy to count is easy to commoditize, which is why it is cheap. The hard-to-count work, the judgment and the relationship, is what you were actually being paid for. Automate the countable layer and you can end up having removed the cost while keeping none of the worth.

A test you can run before you cut anything

You do not need a framework. You need one uncomfortable exercise: before automating a role, write down everything it does when nobody is measuring it.

For a support rep, that list includes spotting the angry customer who is actually about to churn, knowing which "small" bug reports are really a five-alarm fire, and feeding the product team the patterns no dashboard surfaces. For an office manager, it includes noticing the things that are quietly breaking before they become emergencies. For a salesperson, it is the read on whether a deal is real. If your automation plan does not account for those, you are not automating the role. You are deleting it and keeping a fraction.

Then sort the list into two columns. Volume and repetition go to the machine, where it is genuinely excellent and where keeping a human is just expensive. Judgment, exception handling, and relationship stay with a person, augmented by the machine rather than replaced by it. We walked through where that line tends to fall for small businesses in what your business should actually automate with AI, and the same logic governs how we staff and scope software teams now that AI has changed the math. The pattern is consistent across both: AI raises the ceiling on what a person can do, and it punishes anyone who uses it to lower the floor on who is there at all.

Where this lands for the work we do

Most of the AI projects worth building are not "replace the doorman." They are "give the doorman a radio." When we scope AI development for a client, the first real deliverable is not a model. It is an honest map of a workflow that separates the part a machine should own from the part that only looks automatable on a slide. Sometimes that means we build less than the client originally asked for, because the most valuable thing we can do is talk them out of automating the one piece that was holding everything else together.

The businesses that get durable wins out of AI are the ones that treat it as leverage on their best people instead of a replacement for them. That is also the version that survives contact with real customers, which is the only test that counts. If you are staring at a role and a tempting efficiency case and you are not sure which parts are the door and which parts are everything around the door, that is exactly the conversation worth having before you build. We are happy to have it: tell us what you are thinking about automating and we will help you find the value that is not on the spreadsheet, ideally before it walks out the door instead of after.