Every few weeks these days I have an existential crisis about my career. You get used to it, but it stays annoying. This week it was triggered by an episode of AI Report about AI companies, and this piece about ATMs.
Here's the thing. Back in the '70s everyone predicted ATMs would make all bank tellers obsolete. That didn't happen. We actually got more bank tellers. Automation made branches cheaper to run. So banks opened more of them. And more branches meant more staff. They shifted: they weren't behind the counter anymore, they were giving advice.
But that's only part of the story. Starting in 2010 the number of bank tellers in America did collapse: from 332,000 to 164,000 in twelve years. Not because of the ATM. Because of the iPhone.
The ATM automated tasks within the existing system of bank branches. As long as people still came into the branch, there was work for people. The iPhone made the branch itself obsolete. Why drive to a building when you've got a branch in your pocket? Branches didn't close because the tasks got automated, but because the whole paradigm changed.
The ATM asked: how do we do the same thing, but cheaper? The iPhone asked: why are we even doing this this way at all?
How media companies deal with AI right now
Most media companies are stuck in the ATM phase right now. They look at their existing processes and ask: where can AI take over a piece? Can AI write better headlines? Can AI support journalists? This is "ATM thinking." They look at the current process, and where AI can replace a human step.
What you're really doing is shoving AI into human-shaped holes. Existing workflows were designed around human wants and limits. As long as that paradigm sticks around, there are bottlenecks that need a human in the loop. The edge cases. The exceptions. And every time that happens, you hear the skeptics: see, told you, it's just hype.
Another reason results don't show up is a very human one: people aren't motivated to automate away their own job.
This sounds obvious, but it goes deeper than conscious resistance. Because I catch myself doing it constantly too. As a programmer I work with AI every day. I build skills, I automate processes with agents, and I talk to a lot of people about everything AI can do. And yet I hold back on what I ask AI to do, because I think I can do certain things better. These models get better every month, and in more and more areas they're probably just as good as me (or better). But I don't give them the chance.
I'd bet it's even worse with journalists. I picture a whole newsroom full of journalists who each have their own voice and their own style. They all think their own special sauce is special. Let AI generate a piece on their topic from the same sources, and they'll tear it apart. Then they write the same piece with their own special sauce and suddenly it's quality. How much of that is actually a better piece, and how much is just the IKEA effect?
And I know how this feels. I've already been through this whole (mental) process over the past few years with how I write code.
Flip the question around
So the question shouldn't be: what is our current process, and where can AI take over a piece?
The question should be: what does the process look like if we start with AI, and add a bit of human oversight here and there?
Because I can crank out skills forever and teach agents my way of working. But one morning I wake up and I've been lapped by a vibe-coding teenager who built something over a weekend that makes all my beautiful workflows obsolete. Like standing on a baseball field, looking up to see where the ball lands, and then it smacks you in the back of the head.
A nice example of this philosophy is something I read this week in Every. The piece describes a workflow where you really only need 2 people to build a software product: a pirate and an architect. The pirate moves as fast as possible to build valuable features by vibe-coding. The architect takes what the pirate discovers and turns it into a reliable, structured application. Also with AI, but at a slower, more deliberate pace. Every product needs a pirate, but most products only need an architect once they've landed their first customers.
This isn't ATM thinking. This is iPhone thinking. Reshaping your entire workflow.
My blunt advice
When people ask me for AI advice, here's what I'd most love to say: fire everyone and see how far you get with AI alone. You'll be amazed how far you get.
For media companies: stop thinking about where AI fits in. Start over with the question of what the reader needs, and go build that with AI at the core. Then add journalistic oversight on top.
Or don't. Then Big Tech will do it for us soon enough. At least this week I put a few agents out on the street. Luckily they all handled it just fine.