Somewhere in an agency right now, a strategist is watching a little meter tick toward fifty dollars and doing the math on whether one more round of headline variants is worth it.
That meter is real. Digiday reported this week that PMG gives each staff member a $50 daily limit on AI use through its internal portal. Roast has capped staff tokens since March, using a set allotment inside its Claude subscription and raising it when the work spikes. Dept caps usage at the team level rather than per person. Brainlabs runs soft checkpoints instead of hard limits: hit the top of your tier and you show your work to a department head, who decides whether you get more. And according to Bloomberg, as cited by Digiday, Apple has put a $200 daily cap on its engineers.
Fifty dollars a day. Let that sit for a second.
For three years the pitch around generative AI has been some version of “infinite.” Infinite variants, infinite copy, infinite concepts, infinite drafts at zero marginal cost. Whole operating models were rebuilt on the assumption that output had stopped costing anything. Decks promised clients that the bottleneck was gone.
The bottleneck was never gone. It was just being subsidized.
Digiday’s piece lays out the argument plainly: AI prices are probably as low as they’re going to get, propped up by enormous capital spending at Meta, OpenAI and Anthropic. When the subsidy thins, the bill shows up. Agencies capping usage now are building the habit of accountability before the invoice forces it. The opposing camp says the reverse, that if costs are going up anyway, you should burn cheap tokens while they’re cheap and figure it out later. Both camps agree on the part that matters. AI has a price, and the price is going up.
This isn’t a forecast. It’s already in the earnings. Back in July, Digiday reported that Publicis saw a 7 percent rise in other operating costs, driven partly by AI. CFO Loris Nold said on the call that “the productivity benefits that those tools can generate do offset the cost.” Maybe so. But nobody reports an offset unless there’s a cost to offset.
So welcome to the moment AI became a line item. Like film stock. Like studio days. Like the one good photographer you could afford twice a year.
And the second anything has a budget, the only question that matters is who decides where it goes.
Notice where the caps actually land. Not on the CMO. Not on the client. On the person at the keyboard at 4 p.m., with a deadline and a meter. Roast frames its allotment as something each individual is trusted to spend well. That’s a cultural shift dressed up as a finance policy. For the first time in years, a junior creative has to defend why an idea deserved the compute, which is just a new way of asking the oldest question in this business: why this one?
Here’s what the infinite era quietly erased: the discipline of choosing. When output is free, nobody has to decide which idea deserves to be made, because you can just make all of them and let the dashboard sort it out. Fifty headlines. Two hundred thumbnails. A “test and learn” plan that is mostly a plan to avoid having an opinion. The machine didn’t replace judgment. It let a lot of people skip it and call that efficiency.
Rationing brings judgment back, whether anyone asked for it or not. A strategist with fifty dollars a day can’t generate everything. She has to know which three directions are worth exploring before she opens the tool. She has to kill ideas in her head instead of in a spreadsheet. She has to be, for lack of a better word, a creative director.
Dept’s Jonathan Whiteside put the upside well: “Constraints are sometimes useful for us to find new ways of working.” That’s the most understated sentence in the article and the most important. Constraints are where taste lives. Every great creative format in history, from the thirty-second spot to the six-second bumper to the vertical frame, was a constraint someone figured out how to love.
Now the counterpoint, and it deserves a fair hearing. Caps can absolutely backfire. Put a meter on a junior designer and you might teach her to stop experimenting, to stick with the safe prompt that worked last week, to treat curiosity as a cost center. The agencies that skipped hard caps in favor of model routing and training have a point: you can protect the budget without making people afraid of the tool.
But notice what that argument concedes. Even the people against caps are building systems to decide who gets the expensive model and who gets the cheap one, which tasks deserve heavy compute and which can run on something lighter. That’s allocation. That’s judgment, with an org chart. Nobody in this debate is arguing that AI should stay unlimited. They’re arguing about how to ration it.
And here’s the uncomfortable implication most of the industry hasn’t said out loud. If AI is rationed, then the person who knows what’s worth making is worth more than the person who knows how to make it. The prompt jockey who can generate a hundred options in an hour just became a cost. The strategist who can look at a brief and say “it’s these two, not the other ninety-eight” just became the most valuable person in the building.
That’s a reversal of everything the last three years of AI hype promised. The hype said execution was the scarce skill and the machine would democratize it. Execution did get cheaper. What got scarce was the thing you need before execution: a point of view.
The pricing models are already bending around this. Digiday reported this summer that Cheil’s CEO Joe Maglio was moving clients to output-based pricing, with half of existing clients transitioned. When you price on outcomes instead of hours, nobody pays you for the hundred variants you generated. They pay for the one that worked. The variants are your cost. The choice is your product.
We feel this in our own work. The fastest way to waste an afternoon is to open an AI tool before you know what you’re looking for. The fastest way to save one is to spend twenty minutes arguing about the idea first, out loud, with people who’ll tell you it’s bad. Then the machine is useful. Then it’s a power tool in the hands of someone who knows what they’re building, instead of a slot machine for people hoping the next pull is the one.
Think about how LA’s best kitchens run. Nobody in a serious restaurant on Melrose orders unlimited saffron and figures out the menu later. The chef decides what the plate is, then buys exactly what the plate needs. Ingredients are expensive, so the menu is a series of decisions. That’s not austerity. That’s cooking.
Marketing forgot how to cook. It got handed an all-you-can-eat buffet and started serving everything on one plate. Now the buffet is closing, the bill is coming, and the teams that never learned to choose are about to find out what their output was worth without the subsidy.
So stop asking how much AI your team uses. Ask what your team decided not to make this week, and why.
If nobody can answer, the fifty dollars was never the problem.
