In a YouTube video breaking down advertising executive Rory Sutherland’s “doorman fallacy,” there’s a story that should be required listening for anyone drawing up an AI implementation plan.
A hotel has doormen, working in shifts, opening and closing the front door for guests. Management brings in consultants to find efficiencies, and the consultants notice something obvious: the door can open itself. So they recommend the obvious fix, install an automatic door, cut the doorman salaries, and book the savings as pure upside.
For a while, it works. Then the hotel’s average rate starts to slide, loiterers start hanging around the entrance, and returning guests stop feeling like regulars because nobody remembers their name anymore. The door still opens exactly as advertised. It’s everything else that quietly falls apart.
The consultants’ mistake, as Sutherland tells it, is confusing a job with a task. Opening the door was only ever one task inside a much bigger job, one that also included hailing cabs, keeping undesirables at bay, recognizing returning guests, and generally making the hotel feel like the kind of place worth paying more for. The consultants automated the part of the job that happened to be visible and mechanical, and assumed the rest would look after itself. It didn’t, because the rest was never nothing. It was just easier to overlook, and a lot harder to put a dollar figure on.
This is, almost word for word, the same distinction Marc Andreessen draws about AI and work: “a job is a bundle of tasks, and AI works on the tasks, not on the person holding them.” I use that line in my book because it’s the cleanest way I’ve found to explain why so many AI rollouts overpromise and underdeliver. Executives hear “AI can do the task” and mentally file it under “AI can do the job.” Then, months later, they’re surprised to discover the job had a doorman’s worth of intangibles hiding inside it the whole time.
You can see the pattern everywhere AI meets a job title. A chatbot answers the top ten support questions perfectly, so the plan is to shrink the support team. What the chatbot doesn’t do is notice the customer who’s about to churn, calm down the one who’s genuinely furious, or flag the product bug three tickets are hinting at without quite saying so. An AI drafts a serviceable first-year associate memo, so someone floats cutting junior hires. What it doesn’t do is turn into the partner who catches the same issue five years from now, because nobody’s junior year got automated into somebody else’s future judgment. The task gets automated cleanly. The job, it turns out, was carrying a lot more weight than the task ever did.
None of this is an argument against automating the task. You should. The door ought to open itself; nobody needs a person standing there all day pulling a handle, unless the hotel wants to start charging extra for “artisanal door-opening,” which, in fairness, someone has probably tried. The real argument is against skipping the step where you sit down and write out everything else the person in that role was actually doing, especially the parts that never made it into the job description because they were too obvious to the people doing them and invisible to everyone watching from outside. That inventory is the actual work of putting AI to use well. Not deciding whose job to eliminate, but re-engineering the bundle of tasks so the machine takes the mechanical ones and a person keeps the parts that require, for lack of a better word, judgment.
Skip that step and you get an automatic door and a hotel that quietly stops feeling like anywhere you’d want to stay.
More on the ideas above: AI’s Missing Playbook: Aligning AI With the Goals That Move Your Business
It’s time to re-engineer the work, not just replace the person doing it.