There’s a video going around right now, “How AI Became More Expensive Than the Workers It Replaced.” It’s a good watch, and the headline facts are real. Microsoft quietly banned its own engineers from using Claude Code because the bill got absurd. Meta reportedly burned through 60 trillion tokens in a single month, hundreds of millions of dollars in API fees. Token prices have more than doubled, from about a dollar per million to over two. Per the video, Uber, Nvidia, and Amazon are all reporting the same sticker shock. Companies that laid people off to “save money with AI” are now staring at an invoice bigger than the payroll they cut.

The story the video tells is: AI overpromised, it was supposed to be cheap, it got expensive, the magic is fading. I don’t buy the conclusion. The tech didn’t fail. The rollout did.

Every single “AI got too expensive” headline is actually a story about bad change management, wrong incentives, and no strategy. It’s the same mistake companies have made with every new tool for fifty years. We just made it faster and louder this time, and handed everyone a corporate card to do it. Here’s what actually went wrong.

1. We handed people a power tool and skipped the training.

Somewhere in 2023 the mandate went out: use AI. So companies dropped these tools on their employees and, in a lot of cases, forced usage, without any of the change management we already know how to do. No enablement. No “here’s what good looks like.” No workflows, no examples, no guardrails, no one asking which of your tasks this is actually for.

We know better than this. You don’t hand someone a power tool, tell them to use it as much as possible, and walk away. But that’s exactly what happened, at scale, across the whole economy. The waste wasn’t a surprise. It was baked in the moment we skipped the boring part.

2. We measured the wrong thing.

This is the one that gets me. Token consumption became a status symbol. They called it “tokenmaxxing.” There’s a clip in the video of an investor saying that if his $500,000 engineer didn’t spend at least $250,000 in tokens, he’d be “deeply alarmed.” Nvidia’s CEO out there encouraging every developer to burn more tokens.

Think about what that does. You’ve told your best people that the number that matters is how much of the meter you run. So they ran it. On trivial tasks. To look busy at the end of the month. The demand wasn’t real demand, it was a leaderboard.

Tokens are not a KPI. They’re a cost. The KPI is the output: the thing that shipped, the problem that got solved, the quality of the work. Nobody measured that, because it’s harder to measure than a token counter. So we chased the easy number and got exactly what we paid for: a very expensive way to look productive.

3. We had no strategy. We threw it at everything.

Don’t point AI at everything on day one. Start small. Start simple. Point it at what it’s genuinely good at and build out from there.

And the biggest strategic miss is architectural. The durable, cheap, reliable way to use AI isn’t “hand the model an ambiguous task and let it regenerate a solution from scratch every single time it runs.” That’s slow, expensive, and non-deterministic. The same job costs a different amount and produces a different result every run.

The pattern that actually works is the opposite. You build the workflow in code, deterministic and repeatable and cheap, and you drop AI in only at the seams that need judgment. Logic. Classification. Pulling meaning out of messy data. The parts a normal program is bad at. There’s a whole genre of “$7 AI agent” builds floating around that make this concrete: a plain script doing the heavy lifting, calling the model for one small step it can’t do itself, like reading a paragraph and pulling out the one number that matters. The thing barely costs anything because it’s 95% plumbing and 5% intelligence, placed exactly where intelligence is needed.

Treat the model as a genie you re-summon for the whole task every time, and you’ll pay genie prices. Treat it as one smart component inside a system you actually designed, and it’s almost free.

4. We forgot we were running businesses.

This is the root of all of it. We got infatuated with the toy.

For a couple of years we stopped asking the questions we’d ask of literally any other expense. What’s the ROI here? Where’s the leverage? What does this actually replace? Is this the right tool for this specific job, or am I using it because it’s cool? Those are table-stakes operator questions, and we suspended them the moment AI got impressive, because it was impressive, and impressive makes you sloppy.

We treated AI like a genie. Businesses don’t run on wishes. They run on systems, on unit economics, on the unglamorous work of matching the right tool to the right problem. The companies now getting the giant invoice didn’t get outsmarted by AI. They just forgot they were the operator.

So what do you actually do?

A few things I keep coming back to:

  • Onboard AI like a new teammate, not a vending machine. You’d never give a new hire the whole company and say “go.” You’d give them a role, context, and a way to check their work. Same here.
  • Measure outcomes, never activity. If your dashboard shows token spend and not what got shipped, you’re measuring the wrong end of the pipe. Kill the vanity metric before it trains your people to game it.
  • Buy back time on the right tasks. AI is genuinely cheaper than people for some things (drafting, code scaffolds, pulling structure out of chaos) and genuinely more expensive for others. The video notes call centers are still cheaper with people. Know which is which for your business. That’s the whole strategy.
  • Design the system. Don’t summon the genie. Put the intelligence where it belongs and let cheap, deterministic infrastructure carry the rest.

The bill came due, not because AI is a bad deal, but because we deployed a genuinely useful tool carelessly, measured the wrong thing, and forgot to think. The fix isn’t less AI. It’s better systems around it.

That’s the whole game.