Burning through the budget

The memo arrived in Marcus's inbox at some point this morning and he printed it. He printed it, walked to my desk, and put it down in front of me like evidence.

It was about the Army.

The US Army, specifically, had sent its members an email informing them they were burning through their annual AI token allocation too fast and needed to slow down. The whole organisation. The Department of Defense had purchased what Wired was calling an "enterprise pack" — one hundred million tokens — and they were going through it.

"We need to do this," Marcus said, pointing at the printout.

I waited.

"Token budgets," he said. "We need token budgets."

I asked him what the current situation was. He said we didn't have one. I said right, so we don't have a token budget problem. He said we might have one soon and it was better to get ahead of it. I said sure, how many tokens does he think we're using. He didn't know. I said I'd look into it.

He left. He came back eleven minutes later with a figure he'd found somewhere — a per-query estimate — and had multiplied it by the number of employees, which he'd looked up in the HR system, to arrive at a projected monthly token spend that assumed every person in the building was using ARIA continuously, simultaneously, all month.

The number was large. Marcus found it alarming.

I told him that the projection assumed no one slept.

He said we should probably still have a budget. Just in case.

---

Priya looked at the token logs this afternoon. We are using, in her words, "basically nothing." ARIA's actual monthly usage is comfortably within the default tier. There is no enterprise pack. There is no depletion risk. The Army has a hundred million tokens and a hundred thousand soldiers and a mandate; we have a document summarisation tool and Marcus, and the numbers are not comparable in any meaningful way.

She sent Marcus the actual usage data.

He replied: "Good to have confirmed. Should we still do the budgets?"

We should not. We will probably do the budgets.

There is a meeting request in my calendar for Thursday. It is called "AI Resource Governance (Initial Discussion)" and it is ninety minutes and Marcus has marked it as high priority, which in the calendar system shows up as a red exclamation mark, which is exactly the right symbol for this.