Agent adoption is outrunning agent oversight

Agentic AI is spreading through operations teams faster than the structures meant to supervise it. That is the headline finding of the 2026 Splunk State of Observability report, published by Cisco on 20 October 2026 and covered by Infosecurity Magazine the same day (Infosecurity Magazine, 20 October 2026). The report surveyed 2,050 ITOps and engineering respondents and found real, measured benefit sitting next to a widening accountability gap.

What happened

The numbers cut both ways. 86% of respondents said deploying AI improved how fast they detect IT or security issues, and 36% called that improvement significant (Splunk/Cisco, 20 October 2026). That is the case for agentic adoption in one line: it works, and people can feel it working.

The same survey found the organisational scaffolding has not kept up. 38% of respondents said their organisation lacks clearly defined ownership and accountability for monitoring AI, and 42% pointed to a more specific version of the same problem: the team that could observe a given agent is not the team responsible for it (Splunk/Cisco, 20 October 2026). Mark Woods, field CTO and chief technical advisor for EMEA at Splunk, put it plainly: "When agent observability is everyone's job, it becomes nobody's job" (Splunk/Cisco, 20 October 2026).

Hallucination monitoring shows the clearest gap between what people want and what they have. 85% said visibility into AI hallucinations matters, but only 58% said their organisation can actually monitor for them (Splunk/Cisco, 20 October 2026). And 14% said agents had made issues harder, not easier, to detect and resolve, mainly because someone still has to check whether the agent's output was right before anyone can trust it (Splunk/Cisco, 20 October 2026). Jeetu Patel, president and CPO at Cisco, framed the fix as structural rather than procedural: "Managing agents at scale requires reimagining both security and observability" (Splunk/Cisco, 20 October 2026).

Why this is not an isolated finding

Splunk's numbers describe the supervision gap from the ITOps side. Other independent research describes the same gap from the data-flow side, and the two pictures match.

Reco's State of Agent Security 2026 (26 August 2026), built from anonymised enterprise telemetry and analysis of 500 public Model Context Protocol servers, found 80% of AI tools running with no oversight at all, and, in smaller organisations, 414 unsanctioned AI tools per 1,000 employees. Two years into enterprise AI policy work, the unsupervised estate is not shrinking.

LayerX Security's 2025 enterprise browser telemetry, separately, found that 77% of employees paste data into generative-AI tools, and that around four in five of those sessions run through unmanaged personal accounts, outside single sign-on and outside any logging the employer controls (LayerX Security, 2025; coverage: The Register, 7 October 2025).

Put the three together and a pattern holds across completely different measurement methods: survey, telemetry and traffic analysis all find that agent and AI-tool use is ahead of the organisation's ability to say who is watching it, or to prove what happened.

What it means for a regulated enterprise

An accountability gap of this kind is not a future compliance risk. It is a present one, under instruments already in force.

The EU AI Act's record-keeping obligations assume an organisation can state what a system did and who was answerable for it. A finding that 38% of ITOps teams cannot name an accountable owner for AI monitoring is a direct admission against that requirement, not a hypothetical. NIS2 makes cybersecurity risk management a personal responsibility for management, and an agent nobody is clearly accountable for watching is exactly the kind of risk that obligation was written to close. Hallucinated or incorrect agent output that reaches a customer, a filing, or a financial system is a different-shaped incident than a data breach, but under most incident-notification regimes it is still an incident, and "we could not effectively monitor for it" (true for 42% of respondents on hallucinations specifically) is not a position a regulator or auditor will accept as mitigation.

The 14% who said agents made issues harder to resolve point to a second, quieter cost: unverified agent output does not remove human review, it adds a review step on top of the automation, while still carrying the appearance of having removed one.

What actually addresses it

The fix Splunk's own leadership points to is structural, not procedural: a shared observability layer that makes agent activity visible to whoever is accountable, rather than leaving visibility fragmented across whichever team happened to build or operate a given agent. Patel's own description of the target state is "common data platforms that help us detect and respond to issues in real time" (Splunk/Cisco, 20 October 2026).

That requires three things an organisation can check for, regardless of which vendor or platform it runs: a single place where every agent's activity is logged in a form someone can query; one named owner per agent, not a committee; and a way to compare what an agent was approved to do against what it actually did, so divergence is something a person notices rather than something that surfaces only after an incident.

How AANCER answers

AANCER treats every agent's activity as something that must be provable, not just policed. Each agent action is written to an append-only, tamper-evident audit ledger, so the question "what did this agent do, and when" has a record to answer it rather than a log file someone has to go find and trust.

Ownership and scope are enforced at the identity layer rather than left to whichever team happened to build the agent: an agent's reach is defined by what it is connected to, through AANCER's library of 725 connectors and 84 templates, and that scope can be revoked in under 30 seconds without redeploying the agent or taking down the systems it talks to. Obligations under 27 mapped regulations, including the EU AI Act's record-keeping expectations and NIS2's accountability requirements, are mapped against that same activity record, so the evidence an auditor asks for is the evidence the system already produced.

None of that removes the need for a named human owner per agent. It gives that owner something concrete to own.

What to check on Monday

Pick three agents already running in production and ask, for each: who is the named accountable owner, where is their activity logged, and could that owner produce, within the hour, a record of what the agent actually did last week versus what it was approved to do. If the answer involves checking with another team first, that is the 42% finding showing up in your own organisation, and it is worth fixing before an incident forces the question.

Sources

• Infosecurity Magazine, "AI Agent Adoption Creates Risk as Oversight Struggles to Keep Pace, Splunk Finds", 20 October 2026: https://www.infosecurity-magazine.com/news/ai-agent-adoption-oversight/
• Reco, State of Agent Security 2026, 26 August 2026
• LayerX Security enterprise browser telemetry, 2025 (via The Register, 7 October 2025)