This is the part of the AI boom that’s going to annoy a lot of executives: the tech isn’t the scariest thing. The bill is.
Most companies can tell you their cloud costs down to the dollar. They’ve spent years getting religion about it. Dashboards, alerts, budget owners, monthly reviews, the whole grown-up routine. But when it comes to AI usage—especially token-based spending—a lot of them are basically guessing. And guessing is fine for hobbies. It’s not fine for a company.
From what’s been shared publicly, one CIO (at AMD) put it in a way that should make anyone with a finance job sit up: a single trained employee can spend around $200 a week on AI tokens. Scale that across a big organization and you’re suddenly talking about potential annual costs north of $400 million. Maybe that exact number won’t map to every company. But the shape of the problem absolutely will.
Here’s what’s really going on: we’re letting people “swipe the card” for AI all day long, and we’re acting surprised when the statement shows up.
Token pricing is sneaky because it doesn’t feel like spending. Nobody feels the pain in the moment. You’re just pasting text, asking for a summary, spinning up an agent to “help,” hitting run again because the first answer wasn’t right. It’s tiny decisions stacked into a mountain. Cloud costs used to work like this too—until finance teams and platform teams built muscle around tracking it. AI is just cloud cost drama with better marketing.
And yes, I’m going to say the quiet part: the “AI adoption” conversation is often a distraction. It’s easy to talk about what AI can do. It’s harder to talk about what it costs when everyone uses it constantly, badly, and with zero discipline.
Imagine a customer support team that starts using AI to draft replies. Sounds harmless. But then people start using it for internal notes, then for translation, then for rewriting the same message three times because they want it “more friendly,” and suddenly the tool is part of every click. Or picture a sales org where everyone is generating custom outreach. Great. Now multiply it by every draft, every tweak, every “make it shorter,” every “now write it in my tone,” every day of the year.
The really sharp edge is agents. If you let systems run on their own—pulling data, writing outputs, checking things, trying again when they fail—you’ve created spending that can grow while nobody is watching. A human at least gets tired and goes to lunch. Software doesn’t.
That’s why the most important word in this whole story isn’t “AI.” It’s “visibility.” Public reporting says Cisco is trying to handle this with a tool that tracks token spend in real time and can stop “rogue agents” quickly. That sounds like the right category of response. Not sexy, not a demo you show at a conference, but the kind of control you need if you don’t want AI to become a surprise tax.
Still, I don’t love where this could go.
Because the moment you make costs visible, you create a new fight. Who gets the budget? Who gets restricted? Who gets blamed? And how fast does “cost control” turn into “we’re going to lock this down so hard that only two teams can use it”?
There’s also a real risk that companies do what they always do: punish the users instead of fixing the system. They’ll roll out AI broadly, then freak out at the bill, then send a memo telling everyone to “be responsible.” That’s not a plan. That’s just guilt with a logo on it.
The fair counterpoint is that this is normal. New tech always starts messy. People explore, costs spike, then the tooling catches up. Cloud went through this. Mobile app spend went through this. Even printing went through this, back when someone had to beg the office manager for more toner. So maybe the right move is: let the chaos happen for a bit, learn what matters, then put guardrails in place.
I get that. But the difference is speed. AI usage spreads faster because it’s personal. People don’t need a ticket to start. They don’t need a long project. They just start using it because it makes today’s work easier. And that’s exactly why costs can blow up before the company even agrees on what “good use” looks like.
What’s at stake is not just money. It’s trust. If leadership gets burned by runaway spend, they’ll overcorrect. They’ll treat AI like a risky substance instead of a tool. That slows down teams that actually know what they’re doing. And it rewards the teams that can argue best in meetings, not the ones delivering real value.
On the other hand, if companies ignore the bill and keep celebrating “AI everywhere,” they’ll end up cutting something else to pay for it. Hiring freezes. Smaller raises. Budget fights. And the people paying the price won’t be the ones who ran 50 prompts to get the perfect slide headline.
So yeah, I’m pro-AI, but I’m not pro-blank-check. If you can’t measure it, you don’t control it. And if you don’t control it, someone else will—usually finance, usually late, usually with a blunt instrument.
Should companies treat AI token spending like a shared utility everyone can freely use, or like a metered resource that teams have to justify every time they turn the knob?