This is the first in a series we’re calling AI ROI Dispatches, where we share stories from CloudZero and our customers on tying AI spend to real business outcomes.
The challenge
Last week, our go-to-market team consumed 32% of all the AI tokens used across CloudZero. We know this, because we have real-time visibility into that spend per team and even per colleague and use case.
Sales engineers and marketing led the charge, but account executives (AEs) and sales development reps (SDRs) weren’t far behind. Every one of them has unlimited access to Claude Code (and Fable, for a short while).
Many companies and leaders see that kind of consumption curve and get nervous. The common reaction we’ve seen in the news right now is to clamp down with hard token limits, restricted models, paused tool rollouts.
Most of the leaders doing it will tell you privately that it’s the wrong move. They’re not capping AI because they want to. They’re capping it because they can’t connect the spend to any result, and controlling the number is the only move they have when they can’t account for it.
The problem underneath all of this is attribution, not spend. When all you can see is a rising number, the instinct is to cut what you see as a loss. But a dollar that closes a deal and a dollar that produces nothing look identical on a bill. That billing data you get from providers doesn’t make a distinction.
playbook
The AI Cost Optimization Playbook
Traditional cloud cost management is broken. Here’s why — and how to make the switch to cloud cost intelligence.
What we did
Because we’re Customer Zero for our own AI spend monitoring capabilities, we don’t look at the costs as a single rising number. We automatically map every dollar to the activity it drives. For the GTM team, that breaks down into categories like pipeline analysis, meeting prep, and content creation.
This means that, at CloudZero, we’ve stopped asking only “how much are we spending on AI?” and started asking “what is each kind of AI work returning?” And we now get clear answers.
That framing changed what one of our top sales reps did next. He went off and built a demo-prep skill in Claude Cowork. That is now a repeatable workflow that researches an account and assembles a tailored CloudZero demo before a call. He then rebuilt it, and in testing noticed the output was meaningfully sharper on Opus than on Sonnet. The tradeoff: the upgraded skill costs an average of $5.21 per run.
In isolation, $5.21 a run is the kind of line item an arbitrary token cap kills without a second thought. But we weren’t looking at it in isolation. We could see the cost sitting right next to the outcome it produced. And it’s a significant outcome, too.
The benefits
The results were hard to argue with. Across the meetings he prepped for with the upgraded skill, the rep went five for five on conversion, and every one of those meetings scored as “high engagement” in our pipeline scoring.
So, with those outcome metrics in front of us, we decided to roll the skill out to the entire sales team.
Yes, our AI spend went up. Deliberately. We chose the more expensive model and pushed the workflow to more people. A team running on instinct and a token cap would have shut it down. We did it because we could see this spend was one of the highest-returning dollars in the GTM budget. AI spend went up, but our AI ROI also went up, faster.
That’s the whole argument for cost-per-anything, and it’s battle-tested. The companies that win the AI era won’t be the ones that spent the least. They’ll be the ones who could see the economics clearly enough to spend more on what works. And only on what works.
If you’re trying to govern AI spend and understand ROI, that’s the problem we built CloudZero to solve. We use it ourselves and the results speak for themselves. And you should, too.
