AI Didn't Fail. The Way We Rolled It Out Did.

A video’s been sitting in my head all week. The title is “How AI Became More Expensive Than the Workers It Replaced,” and the facts in it are real. Microsoft banned its own engineers from Claude Code because the bill got out of hand. Meta reportedly ran through 60 trillion tokens in a single month. Companies that cut people to save money with AI are now looking at an invoice bigger than the payroll they let go.

The video’s takeaway is that AI overpromised. I don’t see it that way. I keep landing on the same thought: this wasn’t the tool’s fault. It was ours.

I’ve watched enough businesses adopt enough tools to recognize the pattern. We did the thing we always do, just faster and louder. We handed people something powerful, told them to use it, and skipped every piece of the boring work that makes a new tool actually pay off.

Nobody got trained. The mandate was “use AI,” and that was the whole plan. No one said here’s what good looks like, here’s the workflow, here’s the task this is actually for. You wouldn’t hand someone a power tool and tell them to run it as much as possible. But that’s more or less what happened across the whole economy.

Then we measured the wrong thing. Token usage became a status symbol. There’s a clip in the video of an investor saying if his $500,000 engineer didn’t spend $250,000 in tokens, he’d be “deeply alarmed.” Read that again. We told our best people the number that matters is how much of the meter they run, so they ran it, on nothing, to look busy. Tokens are a cost. The output is the KPI. We chased the easy number and paid for the privilege of looking productive.

And underneath all of it, I think we just forgot we were running businesses. We got infatuated. For a couple of years we stopped asking the questions we’d ask of any other expense. What’s the return? Where’s the leverage? What does this replace? Is this even the right tool for this job, or am I reaching for it because it’s cool? We treated AI like a genie instead of like a line item. Businesses don’t run on wishes.

Here’s what I actually believe after sitting with it. The fix isn’t less AI. It’s better systems around it. Onboard it like a teammate, not a vending machine. Measure what ships, not what you spent. And build the thing properly. There’s a whole genre of “$7 agent” builds going around that make the point: a plain script does the heavy lifting and calls the model for the one small step it can’t do itself, so it costs almost nothing. That’s the move. Put the intelligence where it belongs and let cheap, boring infrastructure carry the rest, instead of asking a model to reinvent the wheel from scratch every time it runs.

The bill came due because we deployed a genuinely useful tool carelessly and forgot to think. That’s a very human problem, and it’s a very fixable one.

That’s the whole game.