More than half of organizations (55%) say they can't effectively control their AI costs, according to Splunk's State of Observability 2026 report.
Observability is the practice of monitoring the health of software systems, and telemetry is the logs, metrics and traces that monitoring produces. Splunk said AI costs are becoming less predictable as vendors move to consumption-based pricing, where customers pay per use rather than per user. Agents add to the problem. They work independently and can burn very different numbers of tokens, the chunks of text that models bill by, depending on the task, its complexity and the model chosen. Another 57% of respondents worry their telemetry costs now exceed the infrastructure being monitored.
Finance teams feel it most. Three in five said unexpected spikes in AI bills cause more strain than a consistently high baseline cost. That points toward a preference for flat, seat-based pricing, though it is an inference from that answer, not something respondents were reported to have said outright.
There is a bright spot. Some 81% said monitoring AI workloads has increased collaboration between observability and security teams, and 73% can trace problems to their underlying security causes, up from 65% last year. "This report shows that observability and security data are converging," said Cisco Chief Product Officer Jeetu Patel, who added that teams need to collaborate more than ever.
Cost attribution splits along maturity lines. Among respondents Splunk classed as observability leaders, 91% can accurately attribute AI costs to particular agents, models and workloads, versus 69% of non-leaders. Splunk sells monitoring tools, so it has a stake in that finding looking important.
Cloud computing went through a similar round of bill shock, and the fix was a dedicated discipline of cost tracking. The 22-point gap between leaders and non-leaders suggests AI spending is early on that same path, and unevenly so.