Briefing #37: The Pattern I Keep Seeing in Organizations That Think AI is Free
Most leaders treat AI like a subscription — a flat monthly line item. It isn't. It's compute at scale, and the bill arrives whether or not you've built the case for it.
A client of mine was, in their words, “dabbling” with AI.
And I want to be clear that they were doing it for the right reasons. They’d set their teams loose to explore — find use cases, build some enthusiasm, hunt for the places where AI would actually pay off. That’s a sensible way to start and I’ve recommended versions of that approach myself.
There was one thing they hadn’t done. They hadn’t set up any governance. There were no guardrails on spend, no caps to be found. Everyone had unlimited access, by design, because limiting it felt like it would defeat the purpose of exploring.
At the end of the first month, the bill arrived. And they were stunned.
Not just by the total, but by what was underneath it. When they looked, they found that a handful of people in the organization were capable of spending thousands of dollars in tokens in a single day. Not over a quarter. In a day. That’s what multi-agent workflows running on high-reasoning frontier models will do when nobody’s watching the meter.
“I had no idea AI could be that expensive,” they told me.
It wasn’t a budgeting mistake. It was something more fundamental.
Surprise: AI isn’t a flat-fee subscription
When that leader said they had no idea AI could be that expensive, what they were really telling me is that they’d never internalized what AI actually is.
They were thinking of AI the way most leaders do: as software. A tool you buy a seat for. A ChatGPT subscription — twenty or thirty dollars a head, flat, predictable, forgettable. The same mental slot as your email or your CRM.
But the AI that’s now doing real work in organizations isn’t that. It’s compute at scale, metered and scaled with ambition. The more capable the workflow you point it at, the more it consumes. A leader who asks an agent to reason its way through a complex task, calling other agents, holding huge amounts of context, running in parallel, is using AI as more than just software. They’re renting a small data center by the minute.
Treating that like a SaaS line item is the category error sitting underneath every “I had no idea” moment. And once you see AI as infrastructure that meters, the cost stops being a shock and starts being something you can actually design around.
The trouble is that almost nobody is being told to see it that way.
Where “AI is free” comes from
I don’t think leaders invented the idea that AI is essentially free. I think they were sold it.
For the last couple of years, the dominant story about AI — the one coming out of the venture and private-equity world — has been intoxicating: AI lets you produce massive output without massive input. Write the code, run the function, ship the product, with a fraction of the people. Output decoupled from headcount. Value decoupled from cost.
It’s a seductive story. It’s also, increasingly, not holding up.
Here’s why. That story made sense in an earlier era, the era of subsidized, market-building growth ventures, where investors would happily fund years of losses to win a market, betting that profits would come later once the base was locked in. In that world, the price you paid for a service rarely reflected what it actually cost to deliver. Someone else was covering the difference to buy your loyalty.
That era is over. The frontier AI companies have enormous capital to recoup — the data centers alone are staggering — and the patience of the market has shortened. Investors today want return, and they want it with less risk and on a shorter horizon. The “win the market first, monetize later” bet is out of fashion. Which means the subsidy is thinning, and the real cost of running AI is starting to surface in the bill.
You can see it surfacing already, and not just in my clients’ invoices. Uber, by its own account, burned through its entire 2026 AI budget in four months, driven by how fast its engineers adopted agentic coding tools. Its own COO has been openly skeptical about whether the spend is producing results, saying the link between AI usage and value “is not there yet” (Fortune). Another company reportedly spent something on the order of half a billion dollars in a single month after rolling AI out across the organization with no usage limits (Tom’s Hardware). Goldman Sachs has estimated that agentic workloads could push token demand up by as much as 24 times (Tom’s Hardware).
These aren’t reckless companies. They’re some of the most sophisticated operators in the world. And they got surprised by the same thing my client did.
The thing nobody planned for
Notice what’s missing from almost every one of these stories: the economics.
There’s a detail from the Uber situation that deserves a closer look. They reportedly ran an internal leaderboard ranking teams by how much AI they used. Adoption as a competition. The most consumption wins.
I understand the instinct — you want momentum, you want people leaning in. But look at what it actually rewards. It rewards consumption as a proxy for progress. It turns “we’re getting value from AI” into “we’re spending a lot on AI,” as if those were the same sentence. They are not.
This is the part most organizations genuinely have not thought about: the total cost of making AI core to how the business runs. There is a cost. It’s real, it’s recurring, and it’s larger than the subscription line suggests. And the actual business question, the one that’s been skipped, is whether running AI as infrastructure makes economic sense for a given job, or whether the people you already have are the more sensible investment. CNBC recently framed it exactly this way: ”tokens or humans” is becoming a real corporate trade-off. You can’t answer that question if you’ve never costed it.
I’m not going to lay out the full investment case here — that’s got to be highly tailored to your business, and that’s a topic that deserves its own space. For now I want to leave you with the order of operations, because getting the order right is crucial:
First, the cost and the investment case: what does AI actually cost to run at the scale you’re encouraging? Is it worth it for this particular outcome?
Second, the guardrails: caps, alerts, and visibility — so that exploration can’t stealthily become a half-million-dollar month.
Third, and only third, the hunt for use cases: pointed at the most urgent problems the business actually needs to solve, not at whomever can light up the leaderboard.
Most organizations are doing these in reverse. They turn people loose first, discover the cost last, and never get to the investment case at all.
One question before you turn anyone loose
If you take one thing from this, let it be a question you can ask yourself this week:
Do I know what AI will cost to run at the scale I’m encouraging?
If the honest answer is no, then you’re not exploring AI. You’re running an uncapped meter and hoping the number at the end is one you can live with. My client was a smart, capable operator, and that’s exactly the position they ended up in — not through carelessness, but because nobody had told them AI was anything other than free.
It isn’t. It never was. AI leaders who get this right will know what AI costs before the bill teaches them.
AI, Upfront publishes every Monday. If this was useful, subscribe to get it in your inbox. And if there’s a topic you’d like me to tackle — reply and let me know.