More than half of the ITOps and engineering professionals surveyed say they have no control over AI costs. Unexpected spikes in the bill cause more operational pressure than high baseline costs. That is according to the sixth edition of the State of Observability study, conducted by Oxford Economics on behalf of Splunk, part of Cisco.
The report Splunk State of Observability 2026: Restoring Predictability in the AI Era is based on a survey of 2,050 ITOps and engineering professionals in practitioner and management roles. Oxford Economics carried out the research in February and March 2026 in ten countries: Australia, France, Germany, India, Japan, New Zealand, Singapore, Saudi Arabia, the United Kingdom and the United States. The Netherlands was not part of the sample.
Spikes in the bill create pressure
According to the report, organisations run anywhere from dozens to thousands of agents across multiple models. More than half of respondents say they have no control over AI costs, as usage is rising and becoming less predictable. Of those surveyed, 60% say unexpected spikes in the bill create more operational pressure than high baseline costs. For 57%, telemetry costs exceed the cost of the infrastructure being monitored.
Sander ten Hoedt, Cloud & AI Lead for Cisco Benelux and Switzerland, says that enthusiasm for AI at many organisations, including in the Netherlands, is clashing with the invoice. Ten Hoedt sees observability as part of a strategy for AI tokenomics, which he says makes AI spending measurable and easier to plan.
Agent monitoring is lagging
Of the respondents, 83% are concerned about the unintended consequences of AI agents. According to the report, monitoring agents requires different metrics from those used for traditional applications, linked to business outcomes and to the way agents make decisions.
In practice, this is still only partly successful. For example, 87% call AI monitoring a priority, while 54% have not linked it to clear business objectives. More than half (56%) cannot measure AI workload metrics such as the quality of LLM responses. Insight into hallucinations is considered important by 85%, but 58% say they are effective at it.
Observability and security are working more closely together
Now that agents are gaining more access to operational processes, 81% of respondents believe ITOps, engineering and security teams should collaborate more. According to the report, data sharing between these teams has risen from 74% to 84%.
Jeetu Patel, President and Chief Product Officer at Cisco, expects most companies will soon have more agents than people at work. According to Patel, organisations will then need to be able to determine quickly whether agents are functioning properly or have been compromised. To that end, he advocates shared data platforms on which security and observability teams jointly detect and respond to problems in real time.
Leaders allocate costs more often
The report identifies a group of leaders. Of that group, 91% can accurately allocate AI costs to specific agents, models and workloads, compared with 69% of other respondents. Among leaders, 94% say observability and security are working more closely together thanks to AI monitoring. In addition, 86% of them can trace performance problems back to a security cause, compared with 71% of the rest.
Splunk links these results to three practices. Leaders regularly deploy agents for remediation tasks, closely track model quality, costs, usage, security and privacy, and link telemetry to business impact and strategic planning.
The full report is available via Splunk State of Observability 2026: Restoring Predictability in the AI Era.
This article was originally published in Dutch on MSP Business. Read the original →
