A motion designer at Cash App apologized to me once, mid-session, for how much he'd spent. He'd been chasing a single lighting treatment across maybe two hundred generations and he wanted me to know he wasn't being wasteful, that he'd get it on fewer next time. I asked what he thought the run had cost. He guessed high. The real figure was under twenty dollars, and he had been sitting in that chair for three hours, which made him by far the most expensive input in the experiment.
That's the moment I now watch for in every room. Not whether people can use the tools. Whether they flinch before they explore.
That team eventually cut production time to roughly 10% of what it had been, with an estimated $3.5 million in year-one savings and about 30% more output. The flinch was the thing standing between them and all of it, and it took longer to remove than any technique I taught.
The reflex is fifteen years old and it used to be correct
Creative people built the rationing habit honestly. Render time was finite. Studio days cost real money. Photography was a line item with a number next to it and a producer whose job was to protect that number. You learned, correctly, that exploration is expensive and you should arrive with a direction already chosen.
Then the cost of making a version fell through the floor and nobody told the reflex. So you get senior people treating a few dollars of compute with the same caution they'd apply to a shoot day, while burning an afternoon of their own salaried time deciding which single prompt to commit to. The math is not close. The generation is the cheapest thing in the room. The person weighing whether to run it is the most expensive.
Every's Laura Entis wrote about the enterprise version of this on September 17, and I'd point creative leaders at it. Their team's argument is that you should burn more tokens, not fewer, because the spending is how you find out what the tools can actually do. She notes their CEO runs more than three times the token spend of anyone else at the company. One engineer put a day and a half and billions of tokens into a 3D modelling experiment and got, in his words, "very janky" results — and they kept the run anyway, because it gave them a baseline to test future models against.
I want to be careful about how that translates. Every is a company of people who write and build software. A creative team's version of this isn't token count. It's versions, territories, and dead ends. But the underlying error is identical: teams are optimizing a line item that rounds to nothing and starving the thing that actually produces the work.
What the rationing costs you, specifically
Three things, in the order they show up.
You lose the range. The first output is not the good one. It is the median of everything the model has seen, which by definition is the most obvious version of the idea. Teams that stop at version three ship the obvious idea and then complain that AI work looks generic. It looks generic because they quit before it stopped being generic.
You lose the learning. The janky run is not a waste, it's a measurement. Every failed direction tells you something about where the model breaks on your brand, your product, your category. Teams that never fail expensively have no map. They're guessing at every new job, forever.
You lose the people. Your most curious person is the one most likely to run the big experiment, and also the one most likely to read a nervous glance from a producer as a policy. Punish that twice and they'll stop volunteering. I've watched the best operator in a department go quiet for exactly this reason.
Install a floor, not a cap
The instinct when you hear all this is to remove the budget. Don't. A team with no constraints doesn't explore more, it just stops thinking about value at all. What I put in place instead is a floor and a review.
The floor. Every person gets a standing exploration allowance they do not have to ask permission to use. Small. The point isn't the amount, it's that spending it is the expected behaviour rather than a thing you justify afterwards. If someone has to send a message before they can try something, they won't try it.
The review. Borrow the three questions Every's head of operations puts to anyone who runs up a big bill: what did it cost, what did it buy us, what did we learn. Run them on anything unusually large, in the open, with no penalty attached to a bad answer. The review is what separates this from permission to waste money. You are not approving spend. You are making sure the spend produced a finding somebody else on the team can use.
The owner. Someone holds the log. One page, one line per significant run: what was tried, what came back, whether it's worth repeating. This is the artifact that turns individual burn into team knowledge, and it's the one nobody ever volunteers to keep.
The failure mode, and it is real, is burn as a vanity metric. A team that reports token spend as evidence of adoption has learned the wrong lesson precisely. The output of exploration is a finding, not a number. If the run ends and nobody can say what it settled, that's the case to review — not the invoice.
The ratio to say out loud
Here's the sentence I'd put in front of a creative department this week. The cost of one more version is a rounding error against an hour of your time, and the cost of shipping the obvious version is your reputation.
That's it. That's the whole reframe. It isn't a tool, it isn't a platform decision, and it doesn't require anyone's budget approval. It requires a leader to say, on the record, that exploring is the job and that nobody is in trouble for a run that didn't work. Until someone with authority says that out loud, your team will keep behaving like compute costs what a shoot day costs, because nobody has told them otherwise.
The teams that get real return out of this are not the ones with the best tools. They're the ones where trying the tenth idea doesn't require courage.