The Downshift
Marcus built the company's support system in the years when nobody counted the tokens, and he is proud of it. Then a FinOps lead arrives with a mandate to cut the AI bill seventy percent in a week — and a quiet question Marcus cannot answer about how much of his elegant machine was ever needed at all. A near-future corporate story about the end of spending like it was free.
The email arrived at 9:02 on a Monday, which Marcus Vale would later think of as the exact minute the good years ended.
Subject: AI cost review — let's talk. From a name he did not recognize, Priya Anand, with a title that had not existed at the company two months earlier: Director of Financial Operations. There was a calendar invite stapled to it for eleven o'clock, one hour, a conference room booked under the placeholder Token Diet.
Marcus read it twice and felt the specific dread of a man whose work was about to be discovered. Not for being bad. For being expensive.
He had built the support system three years ago, in the era everyone now spoke of in the past tense, the way you might talk about a city before a flood. Back then the rule had been simple and exhilarating: use the best. The best model, the biggest context window, the most reasoning the menu offered. His director at the time had said it like a benediction — don't make the customer feel the cost, make them feel the magic — and Marcus had taken it as a mandate to build something beautiful.
And it was beautiful. Every incoming ticket went to the flagship, the most capable model money could rent. The whole conversation history rode along with it, the entire relevant knowledge-base article, a system prompt he had tuned over hundreds of late nights until it sang. The thing answered a furious customer at two in the morning with more patience than Marcus had ever managed himself. He had named the internal service Atlas, because it held up the whole sky.
Nobody had ever asked what Atlas cost. That had been the point.
Priya Anand turned out to be younger than her title and entirely unbothered by it. She had a laptop, a single sheet of paper, and the calm of someone who had had this conversation at four other companies and knew exactly how it went.
"I want to start by saying the system is good," she said. "Genuinely. The satisfaction scores are the best in the category. I'm not here to break it."
"But," Marcus said.
"But it costs us more per resolved ticket than a human agent did in 2021." She slid the paper across. He did not want to look at it and looked at it. A single number, annualized, sat at the top with a comma in it he had never had occasion to picture. "We burned this year's entire AI budget by April," she said. "All of it. The whole line. Atlas is most of why."
"It's answering forty thousand tickets a day," he said, and heard the defensiveness in his own voice. "At a quality a human floor could never hit. That number is the cost of magic."
"It's the cost of magic on every ticket," Priya said. "Including the ones that just need someone to read them their order status."
Marcus opened his mouth and found he had nothing to put in it.
She let the silence sit for exactly as long as it needed to, which was how he knew she was good. "Here's the mandate," she said. "Seventy percent off the AI bill. One week. Same satisfaction scores or I lose the argument upstairs and they do something worse than ask you nicely." She closed the laptop. "I think you can get there without anyone noticing. I think you've always known you could. I think you didn't, because nobody made you."
He wanted to be angry at her. The maddening thing was that she had said the quiet part — you've always known — and she was right.
He spent the first two days proving her wrong, which is to say proving her right.
He pulled a week of tickets and read them the way he had not read them in years — not as a stream his system swallowed, but one at a time, like a person. And the shape of it was undeniable and a little humiliating. Most of them were nothing. Password resets. Where is my order. How do I change my address. The kind of thing the cheapest model on the market could answer in its sleep, and which Atlas was answering with the most expensive reasoning engine ever built, like using a particle accelerator to open a jar.
The hard ones — the genuinely ambiguous, the angry-and-also-confused, the edge cases where a wrong answer cost real money — those were maybe one in six. For those, the flagship earned every cent. For the rest, Atlas had been paying concert pianist rates to play chopsticks, forty thousand times a day, for three years.
On Wednesday night, alone in the office because he could not do this in front of anyone, he built the thing he had spent three years not building. A small, cheap classifier model in front of Atlas — a doorman. It read each ticket and made one decision: easy or hard. Easy went to a budget model that cost a fraction. Hard escalated to the flagship, untouched. He took the system prompt and the knowledge-base content, the parts that were identical on every single call, and cached them, so the system stopped paying full price to re-read its own instructions every time.
It took him four hours. He had told himself for three years it would take a quarter and a team. That had also, he understood now, been a way of not doing it.
He ran it against the week of real tickets in shadow mode, the cheap path and the old path side by side, and he sat there at one in the morning waiting for the budget model to embarrass itself so he could feel better.
It did not embarrass itself.
He read every downshifted answer, hunting for the seam, the place where cheap showed. On the routine tickets there was no seam. The budget model told the customer where their order was in clean, friendly, correct sentences. It reset the password. It changed the address. On the handful of genuinely hard tickets the doorman had — correctly, every time he checked — sent them up to Atlas, where they belonged.
The projected bill came in at sixty-eight percent below the old one. Satisfaction, modeled against the historical scores, did not move.
Marcus sat with that for a long time. He had expected to feel caught. What he felt instead was stranger and harder to hold: the quiet vertigo of learning that the cathedral you built could have been a chapel, and that everyone would have prayed just as well.
It was not that Atlas had been wrong. In 2023, with the cheap models as bad as they were, routing everything to the best had been the only safe call, and he had made it, and it had been right. The models underneath him had simply gotten good while he was not looking — good enough that the safe call quietly became the wasteful one, and nobody sent a memo to mark the day it changed. He had kept building the cathedral because the cathedral was the part of the job that felt like art. The doorman did not feel like art. The doorman felt like accounting.
He was beginning to suspect that the difference between those two things had been costing the company a comma-sized number for a year.
Priya came by his desk on Thursday. She did not ask how it was going; she looked at the dashboard over his shoulder, the two cost lines diverging, and she nodded once like a person confirming a sunrise.
"Sixty-eight," she said. "I'll take sixty-eight."
"I can probably get more," Marcus said, and was surprised to find he meant it. "There's retrieval bloat. We're sending the whole article when three paragraphs would do. And there's a long tail of mid-difficulty stuff I could route smarter."
She smiled for the first time. "There it is."
"There what is."
"The thing every good engineer does about ninety seconds after they stop being insulted." She straightened up. "You stop defending the bill and you start hunting it. You're going to be dangerous now. The expensive ones never become the cheap ones. The cheap ones were always the best engineers who finally got told to count."
After she left he sat looking at Atlas — at the name, mostly, glowing in the service registry where he had typed it three years ago full of a pride he could still remember and no longer entirely trust. It held up the whole sky. It had held up a sky that, it turned out, mostly needed a porch roof.
He did not rename it. He had earned the name, in the year it was true. But he opened a ticket to himself, the first one, the start of the long unglamorous work that the next two years were going to be made of, and he gave it a title that was half a joke and half a confession, the way the best engineering names always are.
fix: stop paying concert rates for chopsticks.
Then he went home, in the dark, in the first hour of the era where somebody was finally counting, and he was a little surprised to find he was looking forward to it.