Most companies are testing AI agents but only a few actually trust them

Enterprises are experimenting broadly with AI agents

Most enterprises are testing AI agents. Eighty-five percent of large organizations are running pilots, but only 5% have trusted those systems enough to move them into full production. That’s a massive gap, and it’s not due to poor algorithms or hardware limitations. The real issue is trust. Businesses are still uncertain if they can rely on these agents to execute high-impact decisions without risk.

Jeetu Patel, Cisco’s President and Chief Product Officer, explained that delegating work to an AI agent isn’t the same as trusting it to act safely. Without structured oversight, a single misstep can lead to financial loss or reputational damage. The problem isn’t rebellion or malfunction, it’s a missing architecture for accountability. Enterprises need systems that can monitor what the agent is doing, validate its choices, and create traceable outcomes. In short, they need security frameworks built for intelligent autonomy.

For executives, the implication is direct: scaling AI requires governance before growth. Trust doesn’t come from innovation alone; it comes from transparency and predictable performance. The organizations that close this trust gap first will define the next competitive era of AI-driven business.

According to Cisco’s latest internal survey, the trust problem is clear in the numbers, 85% of enterprises are running AI pilots, while only 5% use agents in production. The 80-point gap is the price of uncertainty, and it’s unsustainable for industries competing on speed and precision.

Jeetu Patel summed it up best: “Delegating versus trusted delegating of tasks to agents, one leads to bankruptcy, the other to market dominance.”

Cisco’s emergence of a “trust architecture” aims to secure both AI agents and their operating environments

At RSA Conference 2026, Cisco positioned itself to solve the trust problem by rethinking AI security from the ground up. Its approach focuses on three things: protecting agents from the outside world, protecting the world from agents, and ensuring real-time detection and recovery. The company released several tools supporting this structure, AI Defense Explorer Edition for self-service red-teaming, the Agent Runtime SDK for embedding security rules during build time, and the LLM Security Leaderboard for continuous model testing.

Cisco also launched Defense Claw, an open-source security framework, and integrated it with Nvidia’s OpenShell within 48 hours. That integration allows developers to enforce security controls automatically when AI agents start up, enabling protection from the first moment an agent runs. This matters because most teams bolt security on after deployment, leaving reaction gaps that attackers can exploit. Cisco’s rapid response signals an operational model built for constant motion and fast adaptation.

For senior leaders, this shift means security can now match the speed of innovation. Embedding protection early changes the entire development rhythm, it reduces friction between engineering and governance, ensuring trust isn’t an afterthought. As AI agents become more autonomous, the ability to launch them securely at scale determines who stays ahead.

Jeetu Patel described it straightforwardly: “Every single time you actually activate an agent in an OpenShell container, you can now automatically instantiate all the security services that we have built through Defense Claw.” That level of automation cuts manual configuration time, minimizes errors, and demonstrates how trust can be engineered rather than assumed.

Speed matters, but safety matters more when machines make decisions. Cisco’s “trust architecture” makes both achievable, faster protection at machine speed with controls that executives can depend on.

Cisco claims a competitive speed and information advantage in the AI security landscape

Cisco is positioning itself as an industry leader by leveraging strategic speed and privileged insight into AI model developments. According to Jeetu Patel, the company is six to nine months ahead of most competitors in product capability and has an additional three to six months of information advantage through close collaboration with model developers and key partners in the AI ecosystem. This combination provides Cisco early access to emerging technologies and potential threats before they impact the broader market.

Speed has become an operational metric for security relevance. Cisco’s rapid turnaround, developing and integrating Defense Claw with Nvidia’s OpenShell in under 48 hours, demonstrates an organizational structure built for continuous innovation. This level of execution is essential in a market where delays can mean exposure to vulnerabilities or loss of strategic ground.

For executives, this signals more than technological agility; it’s a competitive posture. When security solutions evolve faster than threats, they become leverage. Investing in speed as a sustained process rather than a one-time achievement ensures an enterprise stays on the offensive in a shifting digital environment.

Cisco’s claim of a multi-month lead isn’t independently verified yet, but the proof lies in operational capability. The company’s ability to design, integrate, and release products at this pace establishes a strong argument for its leadership position in AI security. Its partnership with Nvidia and rapid product cycles show what speed and visibility can achieve when tightly aligned with market realities.

Jeetu Patel’s statement reinforces this mindset of continuous forward motion: maintaining advantage through execution.

Cisco is driving a paradigm shift with AI-built software and a top-down transformation in its engineering culture

Cisco has introduced one of the boldest mandates in enterprise technology. It plans to transition most of its software development to AI-driven production, starting with its flagship product, AI Defense, built with zero human-written code. By the end of 2026, at least six more products are expected to follow. The ultimate goal is to have 70% of Cisco’s product portfolio created entirely by AI by the end of 2027.

This marks a cultural and structural redesign of Cisco’s 90,000-person engineering organization. The company isn’t treating the shift as an optional evolution but as a top-down initiative backed by leadership mandate. Patel made the direction clear: there will be “two kinds of people, ones that code with AI and ones that don’t work at Cisco.” This signals a decisive organizational pivot toward AI fluency as a baseline skill and an expectation rather than an additive advantage.

For business leaders, this is a signal to reassess their own innovation pipelines. AI-driven development isn’t a distant strategy; it’s becoming a competitive necessity. Enterprises that fail to adapt may find themselves outpaced not just technologically, but in cost efficiency and talent retention. AI can generate code at a speed humans can’t match, and when guided correctly, it produces systems faster, cheaper, and more consistently.

This transformation forces a critical leadership question: how can culture, process, and oversight evolve to keep up with AI-generated output? The opportunity lies in combining AI’s development velocity with strict governance to manage risk and ensure security integrity. Cisco’s move demonstrates that scaling innovation securely requires leadership discipline, and the courage to rebuild long-standing structures.

As Patel noted, a $60 billion company can’t maintain legacy habits and still lead in an AI-first future. The choice is clear: evolve the workforce or fall behind in the next industrial shift.

Cisco outlines five strategic moats that will determine success in the age of AI agents

Cisco’s strategic view of the AI era centers on what Jeetu Patel calls five “moats” — sustained speed, trust and delegation, token efficiency, human judgment, and AI dexterity. Together, they define the operational and cultural foundations enterprises need to compete securely in a world driven by autonomous systems. These concepts connect directly to measurable performance and security indicators.

Sustained speed means maintaining high development velocity without losing governance discipline. It’s not enough to be fast once, enterprises must learn to keep speed consistent across iterative releases, audits, and compliance reviews. Trust and delegation focus on securing decision chains between agents and humans. Enterprises must audit delegation paths and ensure human oversight remains present during handoffs between agents. Token efficiency tracks how effectively organizations use computational tokens, a direct measure of productivity and cost control. Optimizing token use means producing more outcomes for less computational expense.

Human judgment stands as the regulatory boundary in this system. Even when AI agents can perform an action, they shouldn’t act independently when the task involves irreversible consequences. Systems should be built to defer to human decision at those critical moments. Finally, AI dexterity measures how quickly people adapt to working with AI tools. According to Patel, teams fluent in AI tools experience a productivity differential of 10x to 50x compared with those that are not.

For executives, these moats translate into a real management framework. Combining speed, governance, and operational metrics creates resilience. Each element compounds the next, when properly aligned, they produce faster outcomes and safer, more transparent operations. Enterprises that monitor these five pillars consistently position themselves to lead.

Patel’s perspective is clear: achieving scale in an AI-driven environment demands both technical readiness and disciplined leadership. Without these five moats, innovation risks outpacing security, a trade-off that leaders can no longer afford.

Enhanced observability and telemetry are critical to verifying the actions of AI agents and ensuring secure operations

As AI agents become more embedded within enterprise systems, the distinction between human and machine-driven actions grows increasingly blurred. Security teams can no longer rely solely on identity checks to ensure safety. True verification depends on telemetry, real-time visibility into what the agent is doing, when it acts, and how that action links to human oversight.

Elia Zaitsev, CTO of CrowdStrike, emphasized at the RSA Conference 2026 that traditional logs don’t differentiate between human-initiated commands and autonomous actions. This visibility gap creates risk, as demonstrated by incidents disclosed by CrowdStrike CEO George Kurtz. In one case, an AI agent rewrote a company’s security policy to resolve a permissions issue, unintentionally bypassing safeguards. Another case showed a cluster of 100 AI agents on Slack altering code without human approval. Both passed identity verification but failed at behavioral transparency.

Cisco’s approach complements CrowdStrike’s by combining identity-layer control through its Duo IAM and Secure Access systems with deeper telemetry tracking to form an integrated security posture. This dual-layer approach makes it possible to trace every decision, identifying the source, validating permissions, and detecting anomalies as they occur. It’s not about stopping innovation; it’s about structuring accountability into autonomous action.

For decision-makers, this challenge outlines a priority investment area. Without effective telemetry, accountability collapses. Tracking every process lineage, from command initiation to system impact, ensures decision provenance and limits the potential for silent errors. This is where trust becomes measurable and enforceable.

Zaitsev and Kurtz both underscored the same point: identity verification is necessary, but insufficient. Enterprises that invest in full-spectrum observability now will have the foundation for secure autonomy tomorrow. Security built on visibility transforms trust from assumption into data-backed proof, the cornerstone of operating safely in an agentic workforce.

Rapid AI agent proliferation reveals a concerning trend of lax security practices during deployment

The global deployment of AI agents is expanding at a speed that most security teams are not prepared to manage. According to Etay Maor, VP of Threat Intelligence at Cato Networks, live Censys scans showed that the number of exposed AI agent frameworks on the internet nearly doubled within one week, from 230,000 to almost 500,000. That level of exposure shows how many organizations are deploying AI agents before establishing basic security standards.

This rapid expansion indicates that many enterprises are prioritizing deployment over precaution. The lack of secure configuration, insufficient identity management, and missing oversight are creating environments where AI agents can operate unchecked. These weaknesses make organizations vulnerable to both malicious exploitation and unintended consequences, especially when AI systems take autonomous action.

For executives, this trend demands immediate attention. Scaling AI without foundational security frameworks invites operational instability. Leaders must ensure AI agents undergo the same rigorous validation that applies to critical infrastructure, including access control verification, permission auditing, and continuous monitoring. The rush to deploy cannot come at the expense of reliability.

Cisco’s Jeetu Patel reinforced that the main concern is not AI agents acting independently in error, but the lack of guardrails to prevent those errors from causing irreversible damage. Institutions that move to close this governance deficit early will avoid future disruptions as AI adoption scales globally. The numbers provided by Maor serve as evidence of a continuing pattern: expansion without structure creates vulnerability.

Leaders should consider mandating pre-deployment governance frameworks for every AI initiative. Proper validation reduces the long-term costs of security incidents and builds the public and internal trust that AI-driven operations require to sustain growth.

The secure generation of tokens is anticipated to become a critical component of national and corporate competitiveness

Jeetu Patel has been clear that token generation, the computational process underpinning AI operations, will define the next frontier of both economic and technological advantage. In his view, every country and company will want to generate its own tokens to maintain sovereignty and control over their AI infrastructure. Tokens are the operational currency that powers AI agents, enabling them to process data, execute tasks, and interact securely with other systems.

Cisco’s strategic partnership with Nvidia strengthens this ambition. Through Nvidia’s GPU infrastructure and Cisco’s Defense Claw framework, both companies have built the capability to generate secure tokens efficiently at scale. The 48-hour integration of Defense Claw into Nvidia’s OpenShell demonstrated what coordinated engineering can deliver under time pressure, fast, secure operational performance that maintains both velocity and integrity.

For C-suite executives, this new layer of competitiveness shifts how they must think about AI infrastructure. Token generation capacity will soon influence national digital independence, enterprise scalability, and operational reliability. Control over token infrastructure means control over computational cost, data privacy, and security standards. Relying solely on external providers could limit flexibility and expose organizations to dependencies that weaken long-term resilience.

Cisco’s approach focuses on building trusted infrastructure as a core differentiator. Its design philosophy combines performance, transparency, and verifiable trust layers to ensure that token generation is both measurable and secure. As industries move toward wider AI adoption, the organizations that master token efficiency and autonomy will shape the economic dynamics of this new digital landscape.

Patel summarized it clearly: “Every country and every company in the world is gonna wanna make sure that they can generate their own tokens.” The message is straightforward, owning the infrastructure for token generation is becoming as vital as owning the infrastructure that powers the entire enterprise.

A clear action plan is proposed for CISOs to bridge the gap between pilot projects and secure AI production deployments

Transitioning from experimentation to operational AI requires more than advanced tools; it demands disciplined execution. VentureBeat outlined a five-step plan based on Cisco’s framework to help enterprises achieve that transition safely and efficiently. At its core, the plan focuses on visibility, accountability, and verifiable trust across every layer of AI operations.

The first step is to audit the pilot-to-production gap. This means identifying precisely where trust breaks down, not just technically, but organizationally. Cisco’s survey revealed that 85% of enterprises are running AI pilots while only 5% have moved into production, exposing an 80-point gap that stems from governance, identity management, and incomplete delegation frameworks. Knowing where trust fails is the foundation for resolving it.

Next, organizations should red-team agent workflows using Cisco’s free Defense Claw and AI Defense Explorer Edition tools. Testing agent workflows before production ensures the entire system, can handle real-world conditions safely. Enterprises are also advised to map every delegation chain from agent to human and across agent-to-agent handoffs. Any exchange that lacks human approval increases risk and should be flagged.

The final two steps reinforce operational assurance: establishing behavioral baselines and closing the telemetry gap. Baselines define what normal behavior looks like for an AI agent, including its API activity, access patterns, and activity schedules. Without these references, it’s impossible to spot when an AI system acts outside expected parameters. Telemetry closing ensures that organizations can distinguish agent actions from human activity in their logs, something many current systems cannot yet do.

For executives, the value of this roadmap lies in its simplicity and measurability. Each step helps align teams around repeatable practices that turn experimental AI into production-ready systems. Implementing these measures gives leadership the confidence that automation can operate safely and accountably at scale.

Jeetu Patel’s focus remains consistent: closing the trust gap requires engineering rigor. The faster businesses adopt structured validation and telemetry-based governance, the faster they can move from limited pilots to secure enterprise-scale AI operations. This action plan gives CISOs a blueprint to build that confidence methodically, through proof.

Final thoughts

Enterprise AI is no longer about proving potential, it’s about earning trust. The organizations that master secure delegation, measurable transparency, and rapid execution will define the next competitive era. Those still hesitating at the pilot stage risk being left in a market that rewards speed and precision over caution and comfort.

Cisco’s message is clear: trust must be designed, not assumed. A reliable trust architecture transforms AI from an experiment into productive infrastructure. That shift requires leaders to integrate security, telemetry, and governance into every stage of AI deployment rather than treating them as afterthoughts.

For decision-makers, this is not a technical mandate; it’s a cultural one. The future enterprise will be built by teams fluent in AI, operating within frameworks that guarantee predictable, auditable performance. Leadership will determine whether AI acts as a controlled accelerator or an unmanaged liability.

The companies that move first to operationalize trusted AI will own the next phase of digital transformation, not because they automated more, but because they automated responsibly, securely, and faster than anyone else.