Is AI the villain or the savior?

People aren't just your biggest cost. They're your biggest revenue — and your last line of defense.

I was talking with a finance person about AI a while back, and I started doing the math out loud. Say it costs a million dollars to build an agent, and it replaces five analysts who each make a hundred grand. You've spent a million to save five hundred thousand a year. How is that better? He said: NPV. Run the net present value over enough years and the savings clear the build cost.

That's not how any of this works. The models change every few weeks. When they do, the thing you built on top can quietly break — an agent that routed correctly on Monday drops to 71% accuracy on Wednesday because the model under it shifted, and nobody catches it until the complaints roll in. So it's not a million dollars once. It's the build, plus the maintenance, plus the rebuilds, plus the engineer whose whole job is now keeping the agent from drifting. One engineering firm put it bluntly: the real question nobody asks in the planning meeting is the fully-loaded cost of keeping the agent reliable over 24 months. For most organizations, they wrote, that number would change the business case entirely.

I'm telling you this because there's a serious new paper making the rounds — "The AI Layoff Trap," by two genuinely sharp researchers — and I want to take it seriously, because most people won't. Their argument is elegant: in a competitive market, each firm that automates captures the full cost saving but only feels a sliver of the damage, because the real cost — laid-off workers who stop spending — lands on everyone's customer base, not just theirs. So every firm races to automate past the point that's good for any of them. They prove even perfect foresight doesn't stop it. Only a tax does.

It's a smart model. And reading it as someone who studies organizations rather than markets, what it's really describing is institutional behavior — everyone moving the same direction because everyone else is, no one able to afford being the holdout. That's a story organizational theorists have told for forty years. The economics dresses it in new clothes, but I recognize the body underneath.

Here's where I part ways with it, and it comes down to three assumptions a rigorous model can't survive without — and the AI era breaks all three.

One: it assumes automation works. The whole arms race depends on the AI actually doing the job cheaper. But 95% of enterprise AI pilots fail. 78% of companies have an agent in pilot; 14% have scaled one. Gartner thinks 40% of these projects get cancelled by 2027. You can't race toward an efficiency you haven't achieved — but you can lay people off chasing it, which is a different and worse thing.

Two: it assumes displaced people are only lost demand — never new supply. In the model, a laid-off worker is a customer who stops spending. Full stop. But that's exactly the person who now has the same tools and can go build something. The displaced aren't just a hole in your revenue. Some of them become your competition. The model has no room for that, because it was built before that was true.

Three: it assumes people are passive — and so do the boardrooms. This is the one that matters most. Zuckerberg told his own people the 8,000 cuts were to pay for the AI infrastructure bill. His framing: the company has two cost centers, compute and people. And there it is — the error sitting in plain sight. People aren't just a cost center. They're the revenue engine. They're the institutional memory that catches the mistake before it becomes a lawsuit. They're the judgment that the agent doesn't have and won't have in the next five years. Treating them as a line item to trim against a data-center invoice isn't strategy. It's an accounting category mistake with a human cost.

The paper's authors actually leave themselves an escape hatch — they admit that if displaced workers land in better roles fast enough, the whole trap reverses and firms automate too little instead of too much. They assume that won't happen because historically it hasn't. My entire argument is that this is the moment it can, if leaders choose to make it so — because the tools that displace people are the same tools that let the people who keep them outrun everyone else.

So no, I don't think we're all racing off a demand cliff with no brake. I think someone in the building has to be awake enough to notice that the cheapest thing you can cut is also the thing generating the revenue, holding the knowledge, and exercising the judgment. The companies that remember that — the ones who use AI to make their best people exceptional instead of redundant — are the ones who'll still have customers, and a workforce, when the model that promised to replace everyone breaks again next Tuesday.