Operations | Monitoring | ITSM | DevOps | Cloud

People run Codex on their own laptops. When Codex is signed in with a ChatGPT subscription, OpenAI’s own admin console shows who used it and how much: messages and credits. What it doesn’t show is what any of that usage was for, or how it compares to what your team spent on other AI tools. The CloudZero desktop agent for macOS installs on a Mac, sees the traffic from AI coding tools, and prices what those tools use.

AI is expensive because the model bill is only part of the cost. Three components set the floor: model subscriptions, per-token API pricing, and infrastructure. Three more make it move: adapting models to your business, catching and fixing errors, and rising energy and datacenter costs. Efficiency doesn't fix it, because cheaper AI gets used more, not less. Businesses are willing to spend on AI. Research from Deloitte found that in 2025, 85% of organizations increased their AI investments.

AI budgeting is the process of planning, allocating, and forecasting an organization's AI spend: model and API costs, AI infrastructure, tooling, and the people running it all. It differs from traditional budgeting because AI spend is usage-based, scales with product success rather than headcount, and often spans multiple providers.

The instinct when adopting AI coding agents is to optimize the agent. Compare models, tune prompts, argue about which editor has the better completion, and treat the agent as the thing that determines how fast the team moves. Then the commits go up and the product does not. The team building Upsun Dispatch took a different route, and the result is worth copying. They did not find a better agent.

Months have different numbers of days, and a monthly cost chart built on raw totals mixes that calendar difference into the trend. A 28-day February next to a 31-day March shows a 10.7% increase even when daily spend never moved. The same math works in reverse: real growth in a short month can look flat, hiding an increase worth investigating. That costs you time in two places. The first is triage.

If your engineering teams run on ServiceNow, incidents are where they get work done. Putting cost work into an incident gives it the same path to resolution as any other work item your team handles. When a cost anomaly arrives as an incident, your teams route it, assign it, and resolve it on their usual SLAs. When a savings recommendation arrives as an incident, an engineer owns it and acts on it. Now you can send either straight into ServiceNow.

A strong AI business case ties a specific goal to a measured outcome and a fully-loaded cost. Most fail because they skip one of the three: no clear mandate, an over-broad "AI fixes everything" scope, or a cost estimate that ignores adaptation and error-correction. Build it in six steps: define goals, identify uses, break work into tasks, evaluate models, assess total cost, then launch and refine. Most companies are now spending on AI. Far fewer can show what they got back.

You shipped a skill. It worked. You closed the tab. That’s the whole problem. Model choice is a decision you make once, at the moment you’re least equipped to make it: before the skill is even authored. Then you never revisit it, because the skill stopped being interesting the day you got it working. So go back and check. Here’s how.

Teams running containerized workloads on Google Cloud have long relied on JFrog as their single source of truth for container images. The missing piece has been getting Google Cloud’s own runtime services — like Cloud Run and Google Kubernetes Engine (GKE) — to pull directly from JFrog for every container image pull. I’m happy to say that the gap is now closed. Artifact Registry in Google Cloud has introduced a new repository mode called Connector that addresses this requirement.