- In Splunk's 2026 State of Observability, 38% of 2,050 IT and engineering respondents said their organization has not yet set clear ownership for monitoring AI.
- Splunk found 85% of respondents rate visibility into AI hallucinations as important, but only 58% say their organization can monitor for them effectively.
- Name one accountable owner for each AI agent in production, and check that someone can see what it does and stop it.
Who is watching the AI?
An AI agent that answers customers or acts on company data behaves like a new hire. Someone should know what it did today. In some organizations, that person has not been named.
Splunk's 2026 State of Observability was published by Cisco on October 20. It surveyed 2,050 IT operations and engineering respondents. Of them, 38% said their organization has not yet set clear ownership and accountability for monitoring AI.
That means most respondents did not report this gap. It is still a large minority.
Infosecurity Magazine, which covered the report, says businesses are rolling out agents faster than they can build controls. That is the outlet's summary, not a measured result. The survey also asks about AI monitoring in general, not only agents.
The idea here is simple. Running AI well is not only a tooling problem. It is a question of whose job it is to watch it.
Why ownership gets lost
Infosecurity Magazine's account of the report says agents cut across several IT groups. These include software engineering, infrastructure, web development and information security. Each group holds a piece of the picture.
In the survey, 42% of respondents said their main difficulty in observing AI is that another team is responsible for it.
Mark Woods, field CTO for EMEA at Splunk, made the point in plain terms. When everyone shares the job of watching agents, it can end up being no one's job.
This resembles the bystander effect. Each team sees one slice of an agent and can assume another team sees the rest. That is WebPulse's interpretation, but the survey figures are consistent with it.
What watching an agent means
Observability means understanding what a system is doing from the data it produces. For an agent, knowing it is switched on is not enough. You need to know whether it is doing the right thing.
The survey shows a gap on this point. Splunk found 85% of respondents say visibility into AI hallucinations is important. A hallucination is a confident statement from a model that is false.
Only 58% said their organization can effectively monitor for them. By WebPulse's arithmetic, that is a 27-point gap between what matters and what is covered.
Infosecurity Magazine adds that spotting an agent's mistakes is hard. Some organizations may also have no way to stop the action. It gives one example: an attacker could trick an agent into exposing sensitive information. The outlet says proper monitoring tools could prevent that.
Jeetu Patel, president and CPO at Cisco, links the two worries. He asks whether agents have been compromised and whether they are doing their job well. He calls these questions inseparable.
This shows why a split between operations and security is awkward. In WebPulse's reading, an agent that is wrong and an agent that is manipulated can look alike from the outside.
The gains are real, and so is the checking time
The report is not a case against AI. Of those surveyed, 86% said AI deployment improved how fast they detect IT or security issues. Some 36% called the improvement significant.
But 14% said agents made issues harder to detect and resolve. One main reason was the extra time spent checking whether the agent was right.
That time is easy to miss. If staff must verify answers by hand, part of the promised saving disappears.
The evidence has limits. This is a survey, so the figures are self-reported. The coverage does not say how respondents were chosen, and it names no specific incidents. The report also comes from Splunk and Cisco. Patel's conclusion is that common data platforms are key to trusted agents. Readers should weigh that.
Questions to put to your team
First, can you name one person accountable for each agent in production? A committee does not count.
Second, can your team see each action an agent takes, and stop it? Finding an error afterward is a weaker control than halting it.
Third, how do you find hallucinations today? In the survey, 85% rated this visibility important, yet only 58% said they can monitor for it effectively. Where do you sit?
Fourth, do operations and security look at the same data about agents? Patel says observability and security data are converging.
Fifth, how much staff time goes into double-checking agent output? Count it before you count the savings.
Naming an owner is a small step. In this survey, 38% of respondents said their organization had not yet taken it.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Infosecurity Magazine.
