Telcos race to rein in agentic AI with zero-trust governance

Agentic AI represents a fundamental shift from traditional AI applications to autonomous systems embedded within business workflows, demanding a governance-first infrastructure approach rather than treating it as a mere application challenge. Steve Shirkey emphasizes this transition as an infrastructure and operating model transformation, requiring organizations to 'agent-proof' APIs, authentication, and data layers to safely enable autonomous execution, much like the move from monolithic applications to microservices.

Unlike conventional AI models that generate outputs, agentic AI systems act with Non-Human Identities (NHIs) possessing their own credentials, permissions, and audit trails, which necessitates continuous authentication and zero-trust security frameworks. As Shirkey notes, traditional governance validates employees once, but agentic AI demands ongoing oversight, with companies like Red Hat assigning cryptographic identities to each agent session to enforce strict access controls and prevent broad permissions that could lead to misuse.

The governance imperative for agentic AI extends beyond output quality to include rigorous evaluation of intent interpretation, permission adherence, explainability, and human oversight. Autonomous agents pursue multi-step goals, increasing complexity and risk, which Chip McClelland highlights by contrasting static automation for repetitive tasks with AI agents orchestrating complex workflows. This necessitates governance models that treat AI agents like human workers, assigning explicit identities, revocable permissions, and accountability to ensure safe, observable, and controllable autonomous actions.

Despite nearly 75% of organizations planning to deploy agentic AI within two years, Deloitte’s 2026 research reveals a stark governance maturity gap, with only 21% having robust models to manage autonomous agents. This gap underscores the urgent need for integrated infrastructure layers encompassing identity management, tool integration with strict permissions, orchestration for task sequencing, monitoring, and human oversight to ensure reliable, secure, and accountable execution within real business workflows, as Shirkey and multiple analyses stress.

By mid-2026, thought leaders like Steve Shirkey emphasized that agentic AI represents not just an application challenge but a fundamental infrastructure transformation requiring continuous oversight, secure identities, and governed execution. This shift mandates treating AI agents as Non-Human Identities with cryptographic session identities under zero-trust principles, as Red Hat exemplifies, to enable dynamic governance and prevent novel failure modes such as prompt injection attacks. Enterprises must therefore redesign business processes and 'agent-proof' APIs and authentication layers, akin to the industry's historic move from monolithic applications to microservices, ensuring autonomous agents operate safely within complex environments.

The core infrastructure enabling agentic AI hinges on a four-layer stack—model, framework, observability, and infrastructure—that collectively ensure continuous, reliable agent operation while preventing silent failures. While the model layer provides reasoning capabilities, and frameworks manage orchestration with careful error handling to avoid token waste, it is the observability and infrastructure layers that are critical yet often overlooked. Observability delivers real-time telemetry on token spend, tool call failures, and latency drifts, enabling early detection of issues before costly runaway API usage or error loops occur. Meanwhile, infrastructure failures tend to be silent and insidious, as a flawed network or process supervision can cause agents to stop running unnoticed for weeks, underscoring the need for robust process supervision and monitoring.

Preparing infrastructure for agentic AI demands a paradigm shift from hardware-centric to workflow-first planning, focusing on mapping processes, dependencies, and data flows to identify where agents add genuine value and to preempt bottlenecks, especially in hybrid cloud and on-premises environments. This approach recognizes that agentic AI introduces substantial CPU and networking demands beyond traditional compute accelerators due to multi-agent coordination and complex multistep workflows. As Chip McClelland and Shirkey note, while deterministic automation remains essential for repetitive tasks, agentic AI orchestrates complex workflows requiring multi-mode automation platforms supported by dynamic resource prioritization driven by observability telemetry. Importantly, modernization efforts should prioritize reusing existing infrastructure assets and invest selectively to address genuine capability gaps rather than pursuing wholesale rip-and-replace strategies.

Governance-first frameworks for agentic AI emphasize zero-trust principles by treating AI agents as distinct Non-Human Identities with explicit, revocable permissions and cryptographic identities, as demonstrated by Red Hat's Model Context Protocol Gateway. This approach enforces least-privilege access, risk-based permissions, and tiered autonomy aligned to the risk profile of tasks, ensuring that agents only receive narrowly scoped access necessary for their functions. Steve Shirkey highlights that breaking out of chatbot loops to enable real-world task execution introduces new failure modes, necessitating defence-in-depth security measures such as isolated execution environments and continuous policy enforcement to mitigate risks like prompt injection or rogue agents.

Operational controls are critical to safe and accountable agentic AI deployment, incorporating human-in-the-loop approvals for high-risk actions, comprehensive audit trails capturing decision rationale and tool usage, and continuous real-time monitoring across enterprise environments. The TM Forum-aligned Catalyst framework exemplifies this by embedding immutable audit records, policy evaluation engines, and structural human gates before consequential writes, enabling organizations like Deutsche Telekom and Vodafone to reduce policy update times by over 90% through AI-driven automation while maintaining strict compliance. These controls not only ensure traceability and explainability but also facilitate rapid incident response and rollback capabilities essential for regulated sectors such as telecom and public services.

Embedding governance throughout the AI lifecycle—from design and deployment to runtime monitoring and incident response—is paramount to managing the elevated operational risks of agentic AI. This requires early classification of AI systems by risk and autonomy levels, redesigning business processes, and 'agent-proofing' APIs, authentication, and data layers to align with industry standards such as NIST AI Risk Management Framework and ISO/IEC 42001. Cross-industry collaboration among MNOs, standards bodies, vendors, and open-source communities is essential to develop interoperable architectures, trust mechanisms, and shared testing environments that prevent fragmentation and ensure compliance, as emphasized by NGMN and TM Forum initiatives.

The governance challenge in scaling agentic AI is less about AI capability and more about architectural design, ownership clarity, and operational discipline. Surveys reveal that 69% of organizations find agentic AI governance extremely challenging, with many failing governance reviews due to lack of operational accountability rather than model performance. Effective frameworks mandate clear business and technical ownership, risk-based permissions, continuous discovery to combat shadow AI, and new performance metrics tailored to autonomous operations, such as task success rates under autonomy and escalation frequency. This governance-first mindset transforms AI from a reactive security concern into a CEO-level resilience imperative, demanding leadership alignment and organizational readiness.

Lineation.ai’s 2026 launch of a Zero Trust runtime security platform marks a pivotal advancement in protecting autonomous AI agents by assigning each a Non-Human Identity with default-deny access and real-time validation of API calls. CEO Cameron Manavian highlights that traditional perimeter defenses fall short for agentic AI, and their distributed runtime approach slashes incident triage from hours to minutes while embedding continuous policy-as-code evaluation and immutable forensic audit trails to meet stringent compliance standards like SOC 2 and the EU AI Act.

The TM Forum Catalyst initiatives, exemplified by collaborations among telecom operators and vendors, have established a governance-first execution framework that integrates multi-agent orchestration, semantic context layers, and human-in-the-loop gates within existing Open Digital Architecture landscapes. This reusable platform approach tackles fragmented data and siloed APIs, enabling AI agents to safely access business context and execute actions under strict policy control, auditability, and oversight, which can reduce customer service costs by up to 20%, boost productivity by 40%, and compress deployment cycles from months to weeks.

The Telco AEGIS Catalyst project demonstrates how agentic AI-driven security lifecycles can transform telecom operations from reactive to proactive threat management by leveraging AI orchestration, automated playbooks, and shared threat intelligence. Grounded in TM Forum standards, this governance-first platform achieves over 30% automation and 95% pre-validated security changes, delivering faster detection and response times while ensuring every mitigation action is traceable, explainable, and auditable—strengthening trust in error-sensitive telecom environments.

Industry-wide, the NGMN Alliance underscores that scaling Agentic AI in mobile networks hinges more on robust governance, operational controls, and ecosystem collaboration than on individual AI models. Their reference frameworks call for interoperable agent communication, secure identity management, human-in-the-loop controls, and unified standards across bodies like 3GPP and TM Forum to overcome fragmentation and ensure safe, trustworthy deployments. This governance-first mindset is echoed by Fabric.AI’s Governed Vibe Ops platform, which integrates multi-vendor data and AI-assisted operational artifact generation with embedded code review and testing tools, exemplifying how enterprise IT can operationalize agentic AI with security and reliability at scale.

Deutsche Telekom and Vodafone’s TM Forum Catalyst project showcases a practical blueprint for automating complex regulatory compliance through agentic AI, reducing policy update cycles from weeks to a single day by ingesting dense legal texts and translating them into enforceable network access policies. By embedding zero-trust safeguards such as policy dry-runs, human approvals, and negative testing, this initiative leverages telcos’ unique position as trusted custodians of sensitive data to deliver sovereign-ready network compliance, illustrating how governance-first AI infrastructure can shift compliance from a downstream audit burden to an upstream automated capability.

Scaling Agentic AI in telecom networks demands a governance-first approach that integrates robust operational controls, continuous human oversight, and ecosystem-wide coordination. The Next Generation Mobile Networks (NGMN) Alliance emphasizes that commercial adoption hinges less on individual AI models and more on comprehensive support systems encompassing trust, policy adherence, observability, and explainability. Operators must implement mechanisms for human intervention and validation of agent behavior in live environments to mitigate operational risks and maintain service quality, especially given the critical nature of national infrastructure and stringent uptime SLAs.

Ecosystem fragmentation poses a significant barrier to scalable Agentic AI deployment, with diverse stakeholders—including mobile network operators (MNOs), vendors, standards development organizations (SDOs), hyperscalers, and open-source communities—often misaligned on architectures, interfaces, and responsibilities. The NGMN report calls for intensified cross-industry collaboration to develop telecom-grade interfaces, trust mechanisms, and shared testing environments, advocating for harmonized data models and semantic mappings across bodies like 3GPP, TM Forum, ETSI, and O-RAN to ensure interoperability and reduce the risk of siloed, incompatible solutions.

A phased, tiered autonomy model emerges as a strategic imperative for operationalizing Agentic AI, where autonomy is granted incrementally with human override capabilities at every stage. This approach ties the degree of agent autonomy to the risk and potential blast radius of network functions, enabling controlled closed-loop automation across complex multi-domain workflows. Complementing this, new performance metrics—such as task success rate under autonomy, false remediation rate, rollback frequency, and escalation frequency—are essential to evaluate autonomous operations effectively, as traditional reliability metrics like MTTR and MTBF fall short in capturing the nuances of agentic behavior.

Organizational readiness remains a critical yet under-addressed challenge, with many operators lacking formal guidance on the structural and cultural changes necessary to support autonomous AI-driven network operations. The NGMN highlights that most telecom operators currently occupy early maturity levels between manual and semi-autonomous systems, underscoring the need for incremental governance and autonomy enhancements to safely scale Agentic AI. This progression demands continuous evaluation, lifecycle assurance, and human-in/on-the-loop controls to adapt to evolving operational complexities and regulatory requirements.