Ungoverned Agentic AI: What ServiceNow's New Study Found

Ungoverned agentic AI is no longer a prediction about where enterprise risk is heading. It is a measured condition inside most large organizations right now. On October 20, ServiceNow published The State of AI Control, a study run with independent research firm ThoughtLab, and its core warning is blunt: ungoverned agentic AI exposure is compounding inside enterprises, and the frameworks meant to contain it were built for a pre-AI world. ServiceNow newsroom The State of AI Control report

This post walks through what the study actually measured, the five control failures it names, how it lines up with IBM's governance research from the same week, and what the findings mean if you are shipping agents into production.

What the study measured

ServiceNow partnered with ThoughtLab to survey 550 senior executives across eight industries and 13 countries in the Americas, EMEA, and APAC. The study asked what organizations can see (whether every AI system, agent, and workflow sits in a current inventory), what they can enforce, and whether they can account for AI spend. The State of AI Control report

The headline finding is what ServiceNow calls the accountability gap: the distance between how much AI an organization is running and how much of it it can see, explain, or stop. The longer that gap stays open, the more it costs in operational risk, audit complexity, and spend. Amanda Grady, GVP and GM of AI Platform Security at ServiceNow, frames the stakes plainly: unaccounted AI is unmanaged risk. The State of AI Control report

Gap 1: Deployment has outrun visibility

Eighty-one percent of leaders in the study say more AI is running on their systems than they can account for. Only 20% maintain a live inventory that updates automatically as AI is added or changed. The rest depend on manual updates, which means the inventory is stale the moment a team spins up a new agent or workflow. The State of AI Control report

This is the same shadow-agent problem we covered when Microsoft launched Guardian Agents this week: you cannot govern what you have not found. The ServiceNow data puts a number on how widespread that blindness is. Discovery is not a nice-to-have first step. It is the precondition for every other control.

Gap 2: Incidents are common, real-time alerts are not

Ninety percent of surveyed organizations had an agentic AI security incident in the past year. Only 36% contained it within 24 hours. In that window, the report notes, an agent can reach multiple systems, move data, and change records. The State of AI Control report

The reason containment lags is how monitoring works today. Fifty-two percent rely on automated periodic monitoring and 20% use manual log review: checking on a schedule rather than watching as it happens. Only 27% get a real-time alert when an agent acts outside its limits, and only 29% can trace an agent's actions end-to-end across all connected systems. An agent that starts misbehaving at 2 a.m. may not surface until hours later. The State of AI Control report

If your team is working on that traceability problem, our AI agent observability checklist covers the practical telemetry side of it.

Gap 3: Governance exists on paper, not in enforcement

Three out of four organizations define AI policy centrally. But enforcement fragments where AI actually runs. Only 27% enforce central policy consistently, while 49% enforce it inconsistently, 20% define policy separately per tool with some coordination, and 3% define it separately with no coordination at all. Just 19% say formal policy covers nearly all their AI agents and AI-enabled workflows. The State of AI Control report

Nearly half of leaders describe AI deployment as running ahead of their controls. The report's warning is that inconsistent enforcement is where shadow systems take root: a rule that holds in one tool becomes irrelevant in the next, and what slips past an inconsistent gate may not be found until something breaks. The State of AI Control report

Gap 4: Accountability is divided, authority does not match

Fifty-nine percent of organizations split AI accountability across functions with a coordinating body, which often lacks the authority to act. In the 29% where a single leader or team owns the full AI lifecycle, only 37% say that owner's authority fully matches the accountability. They own the problem but may lack the budget, policy, or system access to solve it. The State of AI Control report

Where a single owner exists, it usually sits in technology: nearly half report to the CIO (27%) or CTO (21%), and a third (33%) name a Chief AI Officer. The report's point is that AI governance is still treated as an IT problem when it is actually a business decision about how AI fits into the enterprise. The State of AI Control report

Gap 5: AI spend runs without limits

Only 28% of organizations can attribute nearly all AI spend to a team or use case, and a quarter can attribute only about half of their spend, or less. Seventy-eight percent enforce AI budgets manually or not at all; only 22% enforce automatic hard limits. And only 19% of organizations actively match routine AI tasks to appropriately sized models, which means many suspect routine work is running on expensive large models but cannot prove it, because the allocation between tasks and models is invisible. The State of AI Control report

The causal chain matters here: you cannot attribute spend for AI you have not inventoried, and you cannot govern AI you cannot see. The five gaps are not a menu. They compound. The State of AI Control report

Executives cannot agree on the first fix

Asked which gap is most urgent, executives split: 43% say governing AI consistently across systems, 23% say accounting for spend tied to value, 15% say complete visibility into where AI runs, 10% say catching and remediating problems, and 9% say clear accountability and authority. The State of AI Control report

Our read: the split itself is evidence of the problem. If governance feels most urgent to nearly half, it is because the visibility and ownership layers underneath it were never finished. Governance rules enforced against an incomplete inventory produce confident-looking compliance on top of unknown exposure.

IBM's EMEA study says the same thing from another angle

The ServiceNow findings did not land in a vacuum. A week earlier, on October 14, IBM published research on more than 3,200 senior executives across Europe and the Middle East. Eight in ten said an AI-related incident (agent failure, data exposure, decision error, or security breach) reached their board or C-suite in the past year. Fewer than half, 40%, changed AI governance policy or frameworks in response. IBM study via Newswire Canada Source

IBM's forward-looking numbers rhyme with ServiceNow's current-state ones: 35% of IBM's respondents run agentic AI in production today, 75% expect agents to hold some decision autonomy by 2027, yet only 37% say their governance is designed to scale with agentic deployment. Two independent studies, two weeks apart, pointing at the same gap between deployment speed and control readiness. IBM study via Newswire Canada

What ServiceNow recommends

The report lays out five strategies, which we are summarizing here as ServiceNow's own recommendations, not ours. The State of AI Control report

1. Make visibility operational. Automate discovery of AI systems across native, third-party, and cloud platforms, extend it to shadow AI, and connect the inventory to governance workflows so every new asset gets an owner with no manual handoff.

2. Turn oversight into discipline. Alert in real time when an agent acts outside its limits, route findings to the accountable owner with remediation SLAs, contain in graduated steps (pause, restrict, terminate), and track time-to-containment as a standing metric.

3. Embed governance gates in the deployment pipeline. Require automated security and compliance review before any AI reaches production, enforce one policy set at the gate across every platform, and make the gate the decision point.

4. Establish accountability with authority. Name a leader explicitly accountable for AI governance and give them the power to set policy, allocate budget, approve or reject initiatives, and enforce controls.

5. Attribute spend, then enforce limits. Attribute cost by team, use case, and model; audit whether routine tasks run on appropriately sized models; set and enforce budget limits automatically.

The report closes on a line worth quoting: the organizations that control AI do not build taller walls around it. They govern it where it actually lives, in the work. The State of AI Control report

What this means if you ship agents

This section is our synthesis, not the study's. Three practical takeaways follow from the data.

First, treat inventory and tracing as one system. The 81% visibility gap and the 29% end-to-end tracing number are the same failure at different layers. An agent registry that knows an agent exists but cannot follow its actions across systems is half a control. When we wrote about rogue AI agents in enterprise AgentOps, the containment path started from exactly this: know the agent, watch the agent, then gate the agent.

Second, measure time-to-containment, not incident counts. Only 36% contained within 24 hours, and most monitoring is periodic. If you cannot say how long it would take to pause a misbehaving agent tonight, that is your first metric to build, before any governance framework on paper.

Third, expect the bar to keep rising. Between ServiceNow's data, IBM's board-level incident numbers, and Microsoft's Project Perception launch, the market is converging on the same position: agents are insiders, and they get governed like insiders, with identity, monitoring, and the ability to revoke. Teams that build those muscles now will scale agents faster, because they can prove they can stop one.