The enterprise AI industry has a serious problem, and it sits in the infrastructure that governs what models can do, not in the models themselves.
A new study from OutSystems, KPMG, and CIO Dive, surveying 858 executives, exposes a remarkable disconnect:
• 82% believe their organizations can keep pace with governing AI agents.
• Only 5% have implemented automated controls, audit trails, and formal human review points.
• Just 26%believe they are well prepared to resolve conflicts between AI agents.
Enterprises are deploying AI that can access sensitive systems, call external services, and execute transactions without continuous human supervision. Most still govern these agents with security architectures built for human employees and traditional applications. We spent years making AI smarter and little effort building the infrastructure to contain it.
The harder question is whether enterprises can enforce controls when their trust in AI fails.
At Graphiant, we believe the answer begins with the network.
Agentic AI is all about action.
An AI agent queries customer databases, initiates financial transactions and coordinates with other agents.
Every interaction introduces a potential security,compliance, and operational risk.
Most AI governance strategies attempt to address these risks through model guardrails, application permissions, identity systems, and administrative policies.
Those controls are necessary but not sufficient.
A compromised or misconfigured agent can misuse an application it is authorized to access allowing it to reach an unapproved service, move sensitive data outside the enterprise and open communications with another autonomous system. Once agents work together, one mistake can cascade across organizations.
Relying only on controls that run inside the agent's own application environment is a structural weakness.
An autonomous agent should never have unrestricted control over the infrastructure responsible for containing it.
Graphiant puts that infrastructure outside the agent's control.
Graphiant's Private AI Fabric provides an independent,policy-driven networking layer that governs which systems, applications, and destinations AI workloads can communicate with.
Graphiant enforces human-defined network policy rather than trusting the agent's instructions. An agent can only use paths that policy permits, and if a destination is not authorized, the network blocks the connection.
Graphiant delivers this through its stateless core architecture which provides encrypted connectivity with granular segmentation,distributed security enforcement and continuous Data Assurance.
The result is an independent boundary around what AI systems can reach. Unlike application-level guardrails, it does not depend on the agent choosing to follow instructions.
The first question every board should ask: Which AI systems are operating in our enterprise, and where is our data going? Many organizations cannot fully answer it.
Employees use approved AI platforms, experiment with unsanctioned tools, and increasingly interact with AI functionality embedded in SaaS applications.
Traditional network monitoring sees only fragments of this. Graphiant's AI Usage Dashboard and Data Assurance show network-observable AI usage, including:
• Which AI platforms are communicating with enterprise systems.
• Which sites, departments, and network segments generate AI traffic.
• Which AI applications are approved, unapproved,or blocked.
• How much network traffic is associated with AI usage.
• Where application traffic travels across the network.
That exposes shadow AI, gives teams the network evidence to investigate it, and feeds the AI inventory enterprises should keep alongside their application and agent registries.
An agent that processes customer invoices needs a financial application and perhaps a specific set of customer records. It does not need payroll systems, HR databases or arbitrary internet destinations.
Conventional networks often give workloads far broader connectivity than they need. Graphiant uses zero-trust segmentation to isolate AI workloads; it defines permitted destinations and applications, enforce zone-based firewall policies and restrict traffic between environments. These boundaries are enforced independently of the agent, so if the invoice agent tries to reach an unauthorized system, Graphiant can block the connection.
Application-level authorization still governs permitted destinations, but Graphiant limits unauthorized access and lateral movement. An agent should never inherit broad network access just because it authenticated.
The next question is where the data actually went. Sensitive information can cross many different environments; public clouds, data centers, SaaS platforms and AI providers. The information crosses many different jurisdictions before it arrives, which raises sovereignty, confidentiality and regulatory risk. Graphiant's Data Assurance builds data-in-motion governance into the network. Enterprises can:
• Define approved network paths and geographic transit restrictions.
• Encrypt traffic end-to-end across the Graphiant fabric.
• Isolate traffic between enterprise tenants and partner environments.
• Observe application traffic and its network path.
• Investigate deviations from expected network behavior.
• Apply policies that block unauthorized applications and destinations.
For a multinational running AI agents across clouds and regions, connectivity follows business and compliance rules instead of default routing.
It is not enough to know that sensitive data arrived securely. Enterprises must also be able to control and verify the path it took.
Only 5% of surveyed organizations have automated controls,audit trails, and formal human review points in place. That should concern every audit committee. Manual oversight cannot keep pace with systems that act at machine speed, so organizations need independent evidence of network activity.
Data Assurance provides application classification,flow-level visibility, and historical path data. This allows security and compliance teams to see which destinations were contacted, where traffic originated, and how it moved through the fabric. Graphiant also documents integration with existing security monitoring and response systems.
An AI agent may report that it complied with policy. Graphiant provides independent evidence of the network communications that actually occurred.
That evidence does not replace transaction logs, approval records, or application audit trails, but it adds an independent record that can be essential when reconstructing an incident.
Can the organization stop an AI agent before it causes widespread damage?
Say an agent starts contacting unauthorized infrastructure,unapproved AI services, or systems outside its assigned environment. Telling it to stop is not a security control.
Data Assurance can flag the suspicious traffic, and operator scan block specific applications, domains, or traffic categories across selected sites, segments, or the whole enterprise. They can also revoke compromised network Edge identities. All of this is enforced at the network layer, without the agent's cooperation.
Graphiant has documented an environment in which Data Assurance detected communications between AI infrastructure and the Tor anonymization network, classified the activity as high risk, and provided the operator with evidence and blocking controls.
A rogue agent may ignore its instructions. It cannot simply ignore a network boundary that prevents its traffic from reaching an unauthorized destination.
The survey's finding that only 26% of executives feel prepared to resolve conflicts between AI agents points to another gap.
Picture one agent updating a customer record while another acts on an earlier version and a third sends the result to an external partner.Resolving that conflict takes application-level orchestration, transactional safeguards, and human escalation. Limiting its impact takes control over communications between agents, systems, and organizations.
Graphiant provides policy-governed connectivity across sites,clouds, and partners, with segmentation that limits which systems can communicate. Its stateless core supports large-scale, dynamic connectivity without placing customer routing state in every backbone router. That helps keep an orchestration failure from becoming an unrestricted connectivity or data exposure event.
As AI becomes more autonomous, the system that enforces an agent's operational boundaries should not depend entirely on the agent's own judgment or behavior.
Application security controls what actions software is authorized to execute.
Identity management establishes which machine identities and credentials are trusted.
AI governance platforms manage agent registration, ownership,testing, approvals, and workflow oversight.
Graphiant complements these systems by enforcing network reachability, protecting data in transit, and providing independent visibility into network communications.
Together, these controls give enterprises defense in depth. Graphiant's private, policy-driven fabric runs across locations so enforcement stays consistent as AI scales beyond individual applications and enterprise boundaries.
Boards should demand more than assurances that AI governance policies exist.
They should ask management to demonstrate:
Inventory: Which AI applications, agents, and services are operating across the business? Can management reconcile its agent registry with the AI services and network traffic actually observed?
Accountability: Does every production agent have a named business owner, documented purpose, and approved operating boundary?
Access: Can management demonstrate that AI workloads have only the network connectivity and application permissions necessary to perform their assigned functions?
Oversight: Can the organization reconstruct an agent-related incident using independent network telemetry, application audit records, and human approval evidence?
Testing and intervention:Has management verified that prohibited network connections are blocked,approved workflows remain functional, and operators can contain a compromised workload?
Graphiant strengthens the technical evidence and enforcement behind several of these questions.
Organizations are racing to deploy agents that improve productivity, reduce operational costs, accelerate decision-making, and create new business models.
But deploying more intelligent software does not automatically create a more resilient enterprise.
The organizations that succeed will be those that combine AI innovation with enforceable governance.
They will know which AI systems are operating, where information is moving, which communications are authorized, and how to intervene when something goes wrong.
That requires infrastructure that turns policy into enforcement.
The future of AI governance depends on controlling what AI agents can technically reach.
At Graphiant, we believe that control belongs in the network.
.png)
.png)