Usage Deep Dive

why it matters high ai spend rarely comes from a single obvious change it usually builds up through specific behaviors a model being overused, a retry loop generating excess calls, a misconfigured integration burning tokens, or a team experimenting without visibility usage details exist to help you identify exactly what is driving usage and whether it is intentional this is where you move from noticing that usage increased to understanding precisely why it increased without this level of detail, cost discussions become reactive and speculative with it, they become factual and actionable how to use it digging deeper use usage details when you already know that usage has changed and you need to determine what caused it investigate spikes and anomalies start by narrowing the time range to when usage changed adjust the aggregation interval to zoom in on short lived spikes separate transient bursts from sustained growth correlate changes with deployments, launches, or experiments short intervals surface bursts longer intervals confirm trends compare teams, applications, or projects switch from a combined view to a per workload view you can compare a small set of workloads directly exclude unrelated units to reduce noise identify which team, application, or project is responsible this quickly establishes ownership and scope identify model driven usage filter usage by model to understand which models account for the majority of usage whether higher cost models are being used more than expected if usage shifted from one model to another model level visibility is often where inefficiencies first appear examine usage patterns, not just totals totals hide behavior use different value calculations to reveal patterns maximum values highlight peaks and bursts average values show steady state behavior minimum values help confirm baseline usage these perspectives help distinguish normal growth from problematic usage focus on top consumers review the top consumers for the selected metric and time range to understand which workloads are responsible for most usage whether usage is concentrated or distributed where follow up conversations should start clear ownership speeds resolution save and revisit investigations once you have a useful configuration, save it saved views allow you to revisit the same investigation over time monitor known problem areas share a consistent view with other stakeholders this turns ad hoc analysis into a repeatable process when to use this page use usage details when a budget threshold or alert is triggered you see unexpected spikes in usage costs are increasing faster than planned you need concrete evidence to explain what changed this is where ai usage becomes understandable, not just measurable