A Splunk report released by Cisco argues that AI adoption is moving faster than organisations' ability to measure agent behaviour and costs. The October 20 report surveyed 2,050 IT operations and engineering respondents across ten countries. Its results describe respondents' experiences, rather than an independent audit of every AI deployment.
Monitoring has a measurement gap
The report says 87% of respondents regard monitoring AI as a priority. Yet 54% do not have clear business outcomes attached to that monitoring, and 56% say they cannot measure workload metrics such as language-model response quality.
This separates collecting telemetry from knowing whether an agent is useful. A system can produce many logs while still failing to explain whether the work was completed correctly or whether the result justified the resources consumed.
Unexpected bills create operational strain
Cisco reports that 55% of respondents cannot control AI costs. Sixty percent say billing spikes create more strain than a high baseline bill. The survey also highlights the cost of telemetry itself, so adding more monitoring data is not automatically a complete answer.
These findings support examining both sides of the measurement system: what the agent consumes and what it costs to observe that consumption. The report promotes observability as part of the response; Cisco and Splunk also sell products in that market.
A useful way to apply the findings
For an individual workflow, define an accepted result before selecting additional dashboards. Then connect model usage, retries, external tool activity and review effort to that result. This is an editorial evaluation approach, not a new performance test reported by Franklin.
The most useful signal is whether a team can explain a change in cost alongside a change in successful work. Survey percentages can identify common problems, but a local measurement is still needed before choosing a platform or claiming savings.