Splunk survey exposes widening gaps in AI agent oversight

Businesses are deploying artificial intelligence agents faster than they can monitor their behaviour, leaving gaps in security, accountability and cost control, according to a global Splunk study released by Cisco on Tuesday.

The 2026 State of Observability report found that 87% of technology professionals consider AI monitoring a priority, yet 56% cannot measure important workload indicators such as the quality of responses generated by large language models. Another 54% said monitoring efforts were not linked to clearly defined business outcomes.

The findings, based on a survey of 2,050 information technology operations and engineering professionals across 10 countries, highlight difficulties facing companies as autonomous software moves beyond answering questions to carrying out tasks and interacting with business systems.

Cisco said 83% of respondents were concerned about unintended consequences of AI agents. Although 85% considered visibility into fabricated or inaccurate AI outputs important, only 58% believed their organisations could monitor such hallucinations effectively.

The weaknesses matter because agents can make decisions, execute software commands and access company information with varying degrees of human supervision. An incorrect response may therefore trigger an operational action rather than remain an isolated error in a chatbot conversation.

Responsibility for oversight also remains fragmented. The study identified growing pressure on engineering, operations and cybersecurity teams to share information about how agents function, what they access and whether their actions comply with organisational safeguards.

Cisco president and chief product officer Jeetu Patel said managing agents at scale required companies to reconsider security and observability together. He said questions about whether agents had been compromised and whether they were performing their intended tasks could no longer be treated separately.

Financial exposure emerged as another major concern. Some 55% of respondents said they could not control AI costs effectively, while 60% considered unexpected billing spikes more disruptive than consistently high expenditure. More than half, 57%, said the cost of collecting and processing monitoring data exceeded the infrastructure costs of the systems being monitored.

These expenses can become difficult to allocate when businesses operate multiple agents across different models and applications. The report said 67% of respondents used a single observability platform, intended to provide a consolidated view of systems and their associated costs.

The research also recorded closer cooperation between technology operations and security specialists. Eighty-one per cent said monitoring AI workloads had increased collaboration between those teams, while 73% could trace performance problems to security causes, compared with 65% a year earlier.

The proportion troubleshooting problems jointly with security teams rose to 77%, from 62% in 2025. Sharing of operational data between the teams increased from 74% to 84%, according to the survey.

Separate research has identified similar governance shortcomings. A September survey by EY of 202 senior AI executives at large companies found that 49% of respondents using agentic AI said their governance frameworks had not been updated specifically to address its risks. Twenty-six per cent said their organisations could not detect unauthorised agents operating internally.

Those figures describe respondents’ assessments rather than independently verified breaches. They nevertheless illustrate why access permissions, monitoring records and intervention procedures are becoming important considerations alongside the productivity gains promised by autonomous systems.

Splunk’s survey also found differences between national markets. Among UK respondents, 32% said they always used AI agents to diagnose the root causes of technical problems, compared with 20% globally. Fifty-five per cent in the UK expressed strong concern about unintended consequences.

Saudi Arabia recorded particularly positive assessments of productivity, with 98% saying AI allowed them to focus more on innovation, against a global average of 84%. In Germany, 74% considered visibility into AI privacy very important.

Cisco separated organisations with advanced monitoring practices from other respondents. Among those classified as leaders, 91% said they could attribute AI expenditure to agents, models and workloads, compared with 69% of other organisations. The company said 94% of the leading group reported collaboration with security teams, against 80% elsewhere. These comparisons reflect the survey’s classification of monitoring maturity and do not establish that adopting a platform caused better results. The report was published as companies assess the operating costs of autonomous applications.

Oxford Economics conducted the research between February and March 2026, covering professionals from software development and infrastructure operations to chief technology and information officers. Respondents worked across 16 industries, including financial services, healthcare, manufacturing, government and telecommunications.