The team that can't audit the cloud bill isn't watching the AI bill either
Cloud waste is a nobody’s-job problem: one overworked engineer owns infra, security, on-call, and cost for five teams, and cost is the thing that quietly loses. Now watch what happens when that same team ships an AI feature.
A third meter, and no one to read it
They didn’t hire anyone. They didn’t free anyone up. They added an LLM API — a third metered thing to not-watch, stacked on the two they were already drowning under. So the outcome is predictable: if nobody had time to audit the AWS bill, nobody is auditing the OpenAI bill.
The caps don't get built because they're nobody's job.
Except the AI meter is worse in two specific ways (the whole reason this series exists):
- It moves faster. Cloud creeps up over weeks. A leaked key, an agent loop, or a retry storm 10בs an AI bill in hours.
- It can be turned against you. Cloud waste costs money. An AI feature can also be used as a free LLM, or jailbroken into a liability. That’s not a finance problem or a security problem — it’s both, which is exactly why neither team picks it up.
Why the caps never get built
Here’s the uncomfortable part for anyone who read the earlier lessons and nodded along: the gateway, the spend caps, the guardrails — none of them are hard. They stay undone for the same reason the cloud bill goes unaudited: they’re nobody’s job. A control with no owner is a control that lives on the backlog forever, right up until the incident that finally makes it someone’s job — expensively, and at 2 a.m.
The technical playbook in this series is the easy half. The hard half is that a stretched team won’t run it, because running it isn’t anyone’s actual role.
What “owning it” actually looks like
Owning the meter isn’t a full-time hire — it’s a small, boring set of responsibilities that currently belong to no one:
- A 15-minute weekly read of the AI and cloud spend, so a spike is caught in days, not on the invoice.
- The caps and alerts wired once — a hard, fail-closed spend cap and a real-time alert to a channel someone actually watches.
- A standing rule that every new AI feature ships with the controls — a gateway route, a per-user
budget,
max_tokens— before it goes live, not after the first bad night. - A one-page runbook for the only two calls this ever generates: “the bill spiked” and “a key leaked.”
None of it is hard. All of it is unassigned — and unassigned work doesn’t happen, it waits for an incident to assign it. The dashboards were never the missing piece; the owner was.
The fix is the same fix
An owner. Either resource it for real — give someone the mandate and the protected time — or rent one from the outside whose entire job is capping and guarding the meters so your firefighters don’t have to. The dashboards were never the missing piece. The owner was.
If your AI feature shipped fast and nobody’s job is watching what it spends or how it’s abused, that’s the gap I fill. Every message comes straight to me — I read and reply to each one myself, usually within a day. Send me your setup and I’ll tell you where it’s fail-open — free, within a business day.