
Rizwan Z. | Nov 12, 2026 | 8 min
Ask a leadership team whether its AI agents are governed and most will say yes. Ask the same team for evidence, covering who approved each agent, what it can reach today, what it did last month and who is accountable for it, and the answer usually takes days.
That gap between belief and proof is the central finding of a new study of agentic AI security. ServiceNow and Phronesis Partners surveyed 3,766 senior security and technology leaders across 20 countries and eight industries, all from organisations already running AI agents in production.
More than half of those leaders (56%) rate their agentic security as established. Fewer than half (47%) can produce a complete evidence package on demand. And on a scale that measures what organisations have built rather than what they believe, only 13% reach a defensible posture.
The study's scale is worth understanding, because it explains why the self-ratings and the results diverge. It scores organisations on five areas:
Attack surface (25%): how new agents are discovered.
Security by design (25%): when security enters the process.
Visibility (20%): how a complete view of agents is obtained.
Accountability (15%): who is responsible.
Audit readiness (15%): how quickly evidence can be produced.
Each is judged on observable choices, not opinion. On that basis:
14% sit at Limited.
46% at Emerging.
26% at Advancing.
13% at Defensible.
The study is still a survey, so the answers are self-reported. But the questions ask what organisations do, not how confident they feel, and that is how an auditor or regulator will ask too. Nobody asks whether governance feels established. They ask to see it.
Existing control frameworks assume a person makes each decision and can be held to account for it. Agents break that assumption, and the frameworks haven't caught up. Fewer than half of organisations say their framework explicitly covers:
Agent inventory, ownership and permissions: 48%
Real-time monitoring of actions: 47%
Accountability for agent decisions: 46%
Incident response: 42%
Documentation and audit trails for agent activity: 39%, the element most often left out
A compliance leader at a German insurer named the specific gap: their process for system changes doesn't include sign-off on AI model iterations.
That one example deserves more attention than it gets. In many regulated organisations, change control is the backbone of operational resilience. In EU financial services, the Digital Operational Resilience Act has applied since 17 January 2025, covering ICT risk management, resilience testing and third-party risk. If a model update or a change to an agent's permissions can reach production without passing through the same change process as a firewall rule, the organisation has a gap in the control it relies on most.
The fix doesn't need a new framework. Agent and model changes should go through the existing change process, with the same approvals, records and rollback plans as any other production change.
One US compliance leader in retail described the problem precisely: they can't store full reasoning traces for more than 30 days, so any forensic analysis is incomplete. The study makes the same point in its recommendations. Traces may be trimmed at 30 days, but auditors ask about years.
Platform defaults show why this needs a deliberate decision. ServiceNow's Brazil release, for example, lets AI Control Tower keep session and trace data for scoring for up to 30 days, or discard it to save storage. That's a sensible setting for monitoring quality. It isn't a records retention policy.
Organisations need to decide separately which agent records count as evidence, how long each must be kept, and where they will live. At minimum, that means the approval, the scope granted and the action taken. Those decisions belong to risk, legal and data protection, not to whatever default a tool ships with.
Only 20% of organisations have a single named owner accountable for agentic AI governance. In the rest, responsibility is spread across security, IT, data and the business. Without one person to bring it together, nobody sees how agent behaviour, access and risk connect across the organisation.
The respondents describe the same problem in their own words:
A compliance leader at a UK automaker said liability frameworks are unclear, with no formal documentation splitting risk ownership across teams.
A chief AI officer in UK telecoms described trying to set guardrails while regulations, priorities and data privacy concerns keep shifting.
A CIO in the Netherlands recognised the governance gap the day an internal agent started pulling customer data from five tools and emailing summaries.
The exposure also extends beyond IT:
82% of organisations run agents in, or interacting with, operational technology or physical environments: industrial controls, building systems, medical devices.
76% have had an incident in those environments involving an agent.
Only 32% have a separate governance framework for them.
When accountability is divided by environment, the boundaries between environments become the exposure.
The study's answer is simple: give one leader responsibility for the framework, the gates, the response plan and the evidence, while each function still owns the agents it runs. Shared responsibility still needs someone accountable for the whole.
When an organisation can't produce evidence on demand, it builds it after the request, from systems never designed to answer the question together. That takes time, and it leaves gaps that a regulator will find.
The study points to where the difference comes from. Among organisations that manage agentic security on a unified platform:
51% capture evidence continuously and can produce it on demand, compared with 42% of others.
60% capture automated real-time alerts on anomalous agent behaviour, compared with 43%.
44% have fixed an agent's exposure or policy violation automatically, compared with 25%.
These are correlations from a survey, not proof of cause. But the logic holds. When the system that runs security is also the system of record, the audit trail is created as the work happens instead of being reconstructed afterwards.
Pick one AI agent running in production. Within an hour, try to answer four questions from your systems, not from memory:
Who approved this agent?
What can it reach today?
What has changed about it since it went live?
Who is accountable for it now?
If any answer means chasing colleagues or searching email, your posture isn't yet defensible, however your team would rate it. That isn't a failure; 87% of organisations in the study are in the same position. It is a clear place to start, and it matters more as agents take on work that regulators already examine closely.
EnableNow works with organisations that run regulated operations on ServiceNow. Bringing agents into the same ownership, change and evidence disciplines as the rest of the estate is becoming a regular part of that work. If you're working through it, we're happy to share what we're seeing.
Source: Deployed but Ungoverned: The State of Agentic Security, ServiceNow with Phronesis Partners, 2026. Survey of 3,766 VP-level and above security and technology decision-makers across 20 countries and eight industries.

Rizwan Z. • Nov 18, 2026 • 8 min
Read full article→

Rizwan Z. • Nov 05, 2026 • 10 min
Read full article→

Rizwan Z. • Oct 21, 2026 • 10 min
Read full article→
Let EnableNow help you modernize workflows, strengthen performance, and unlock measurable value across your ServiceNow platform.