A New Jersey gas station, a doorman, and the AI mistake everyone makes
Most companies automate the job title, not the job.
Ollie on Work is a weekly newsletter about what I’m learning from building with AI, advising leadership teams, and trying to bridge the gap between what technology can do and how businesses actually work. If someone forwarded this to you, you can subscribe here:
New Jersey is the last state in America where it’s illegal to pump your own gas.
Every station has an attendant. Someone comes to your window, fills your tank, maybe checks your oil. It’s been this way since 1949. Multiple attempts to change the law have failed - not because of some powerful gas station lobby, but because 73% of New Jersey residents prefer it this way.
In a country that self-serviced everything else decades ago - checkouts, check-ins, banking, boarding passes - the people of New Jersey are actively choosing to pay a bit more to have a human do something they could easily do themselves.
They’re not stupid. They’re not lazy. They just understand something that most businesses have forgotten.
The doorman fallacy
There’s a concept I keep coming back to, popularised by Rory Sutherland. He calls it the doorman fallacy. It goes like this.
A consultant visits a hotel. Sees a doorman. Defines the doorman’s job as “opening the door.” Replaces him with an automatic door. Claims the cost saving. Leaves.
Five years later, what the hotel can charge per night has dropped. There are people sleeping in the entrance. The most loyal guests have quietly moved on. But the consultant who made the recommendation is long gone - and nobody connects the decline to the decision.
Because the doorman was never just opening the door. He was hailing taxis. Recognising regulars. Providing security. Signalling status. Making guests feel like they’d arrived somewhere worth arriving at. None of that was in the job description. None of it appeared on a spreadsheet. All of it was real.
The saving was easy to measure. The value destruction was invisible - until it wasn’t.
Where the conversation goes wrong
I’ve been in a lot of these conversations recently. The ones where a team sits down to figure out where AI fits into what they do. And the thing that strikes me every time is how early the conversation goes wrong.
Not because people are rushing to automate things they shouldn’t. Most aren’t anywhere near that stage yet. The problem is more basic than that. People talk about “workflows” and “processes” without ever having properly defined what those actually are. The word gets thrown around in every meeting, every strategy document - but when you ask someone to describe, step by step, what actually happens in a process they own, most people can’t do it. Not because they’re incompetent. Because nobody ever asked them to.
So the first thing I tend to do isn’t talk about AI at all. It’s much simpler. I ask people to describe what they actually do.
I was in a session recently where a team was looking at how they handle incoming data from external partners. On paper, it’s straightforward: data comes in, gets checked, gets logged. The kind of thing that looks perfect for automation.
But when I asked the person who actually does this work what their day really looks like, a completely different picture emerged. They’re not just processing data. They’re spotting when something looks off before it becomes a problem. They’re maintaining a relationship with the people sending it. They know which partners are reliable, which ones cut corners, and which ones need chasing before a deadline. They’re an early warning system dressed up as an admin function.
That conversation alone - before anyone mentions a single tool - changes everything. It changes what you’d choose to automate, and how.
Automate the data processing? Absolutely. But if you remove the person without understanding what they were really doing, you don’t just lose an employee. You lose the early warning system. And you won’t notice until something blows up.
Among the companies that are actually deploying AI - and honestly, it’s still not that many - the default is to define a role by its most visible task, automate that task, and claim the saving. Without ever asking what invisible value gets destroyed in the process.
The better question isn’t “which roles can we automate?” It’s “what’s actually happening here that we haven’t bothered to articulate yet?”
And even once you’ve understood the full picture, there’s a harder question underneath it - whether the process should exist in its current form at all. But that’s for next week.
One thing you can do this week
Before you automate anything, go and sit with the person who does the work. Not their manager. Them.
Ask them what they actually do all day - not what the process document says. Listen for the things that don’t have names. The judgment calls. The relationship maintenance. The pattern recognition that only comes from doing the same thing, with the same people, for years.
That’s where the real value lives. And that’s what no spreadsheet will ever show you.