Enterprises are forging ahead with the adoption of AI in networking. But there are tensions between the need for AI to improve efficiency and help avoid falling behind the competition, and the security gaps that organizations fear could lead to costly breaches. Add in concerns about the security practices of vendors, the rising cost of AI, and uncertainty over the direction this technology will take as it matures, and it’s clear that enterprises are facing some tough decisions about how to move forward.
Our recent survey shows that most organizations have moved beyond the pilot and experimentation stages of AI in networking into full adoption. Well over three-quarters (78.5%) have deployed AI-driven networks, and more than a quarter (27.8%) already have autonomous network operations.
AI in networking operations has clear potential to improve network performance, reduce downtime, proactively troubleshoot problems, reduce alert fatigue, increase visibility, and unify networking and security.
But our survey uncovered some fascinating contradictions in enterprise attitudes towards AI. These most likely reflect the friction that many organizations are experiencing between the perceived benefits of AI and the security and governance risks it creates, which most organizations have yet to fully address.
Agentic AI Flips the Security Script
In our survey, nearly one quarter (23.8%) said that security, privacy, and data quality are the primary barrier to AI adoption. This isn’t surprising given that AI – and particularly agentic AI – changes the shape of the enterprise attack surface.
To achieve what they’re designed to, agentic AI deployments have to plug into many different enterprise systems, like ERP and billing systems, customer databases, and communication tools. They’re also increasingly given the autonomy to make and implement decisions, write data, and send communications.
So they often have widespread system access, but compared to humans, AI agents are typically constrained by far fewer controls and are given more and longer-lasting privileges than they really need. This makes them a potential weak link in security terms, and one that’s often lacking critical human oversight.
These risks are no less significant when it comes to the use of AI in networking. On average, our survey showed that organizations believe introducing AI into networks will make them 13.4% more vulnerable in the next five years. Data privacy and AI training data exposure tops the list of AI-driven networking concerns, with almost half of respondents (47.1%) saying they were worried about it. Ineffectively secured AIs in network operations could provide attackers with access into and across the network, be manipulated into ignoring malicious activity or traffic, reveal sensitive information, or allow it to be reconstructed.
The Security Arms Race Is Faster Than Ever
More than half of organizations in our survey (51.8%) are already using AI-powered threat detection. AI’s value in cyber defense is manifold: it has the scale and breadth to analyze a sea of information to spot and mitigate threats faster, it can automate repetitive tasks, and it helps focus decision-making by prioritizing information and alerts.
But the same capabilities that make AI so helpful to enterprises are, of course, being used by attackers to find and exploit vulnerabilities. In 2025, Crowdstrike built a simulated command-and-control (C2) platform to demonstrate how this works in practice. This showed how agentic and generative AI capabilities dramatically speed up the process of compromising a host via social engineering, then enable device and system reconnaissance, data exfiltration, and the generation of custom code to dynamically evade security defenses.
The big question is who can stay ahead in the security arms race. Nearly two-thirds (65%) of enterprises in our survey believe that AI-powered defenses will outpace AI-powered attacks, and only 11.3% believe the attackers will keep ahead.
Our respondents were more conservative when it came to AI-driven networks specifically, though. Here, 37.2% said that AI-powered attacks evolving faster than defenses was a concern, possibly reflecting that AI-driven networks are potentially more vulnerable than other AI use cases.
AI Demands New Security Models
Nearly three out of five people (59.5%) in our survey believe AI-driven networks require a fundamentally different security approach compared to traditional networks.
Just like any other use case, building AI into the network expands the threat surface and opens up the risk of AI-specific attacks like prompt injection and model poisoning. As the autonomy of agentic AI increases, the risk grows too – and these kinds of attacks could have a catastrophic effect on the performance, uptime, and regulatory compliance of enterprise networks.
AI creates a dynamic environment that static, rules-based security models can’t cope with. Instead, zero trust frameworks – often deployed as part of a broader SASE implementation – continuously challenge and validate agents’ credentials and access requests, provide limited access for a limited time, and assign the bare minimum of privileges to AI agents. This helps to identify breaches more quickly and limits the amount of damage if an attack is successful.
One of the great challenges in network security is a lack of visibility – as the saying goes, you can’t secure what you can’t see. This is often caused by systems and tools that don’t work well together, leaving network and security teams switching between different interfaces and struggling to get a comprehensive overview. This is another area where agentic AI offers a potential solution, by tying together disparate network and security tools, components, and systems into a single operational platform. This approach provides comprehensive visibility and unified orchestration, analytics and reporting, faster identification and resolution of security issues, and more stringent application of data security rules and policies.
Enterprises Are Confident About Governance, but Less Sure About AI Making Decisions
Many early adopters of agentic AI in cyber security have focused on using agents to help human teams triage and investigate risks and suggest appropriate responses.
This cautious approach is helping lay the foundations for future agents with greater autonomy, that can make and implement security decisions without human intervention.
More than two-thirds (68.8%) of organizations said they would trust AI to automatically quarantine a major part of their network without human review if it detected a threat, so long as a confidence threshold is met. This figure is surprisingly high given that 39.4% say their biggest security concern with AI-driven networks is AI making incorrect security decisions.
Perhaps this contradiction is explained by the confidence enterprises feel about their ability to put in place the policies, protocols, processes, safeguards, and tools that make up effective AI auditing and governance. Our survey showed an average score of 8.05 on a scale of 1-10 when it comes to organizations’ confidence in their ability to govern AI.
Several of the companies we spoke to said that their confidence had been bolstered by working with trusted partners. This is supported by the fact that 41.6% of organizations favor a hybrid approach to AI-decision-making, with an internal team and an external partner taking joint responsibility.
Organizations Need AI Partners, but They’re Not Convinced They Can Be Trusted
Our survey surfaced an interesting contradiction: 85% of respondents said they believe they have the in-house expertise to manage autonomous AI networks, but 67% admitted that their biggest worry is implementing AI in netops without the necessary expertise.
This could reflect a confidence gap in the reality of rolling out AI in network operations, or it could show that organizations are concerned about the implementation phase but comfortable about the prospect of routine management.
While enterprises are looking to third parties for help navigating AI deployment, they have significant misgivings about vendor practices. Nearly two out of five (39.4%) want to outsource AI model training and optimization to a trusted partner, but the biggest trust gaps in AI security are managed AI providers which aggregate multi-client data (33.8%) and AI vendors and their security practices (31.4%).
This issue is likely to be compounded by the fact that 39.0% of enterprises are relying on multiple partners to help them realize their AI-driven network strategies, since a hybrid model risks introducing operational complexity and could hinder regulatory compliance.
Enterprises are understandably worried about losing control of their data to providers without clear visibility of how it’s used, stored, and secured. They’re looking for partners that can demonstrate the very highest levels of data security and compliance, or that don’t collect sensitive customer data at all – but in the meantime, some respondents in our survey said the vendor trust gap was holding back their AI adoption ambitions.
AI-Driven Networking Will Become More Proactive, but Full Autonomy Is Some Way Off
What do enterprises think the future holds for AI-driven networks? Our survey showed that more than half (50.4%) of organizations believe predictive and simulation technologies will transform networking in the next 3-5 years, closely followed by autonomous networking (49.6%).
Both these technologies have the potential to help organizations shift away from what’s often their daily reality: a reactive, fire-fighting model where problems are only tackled once they’ve begun to have an operational effect.
AI-powered analytics and simulation technologies like digital twins allow organizations to spot likely network problems before they affect users, and to explore ways to improve network performance without having to test changes on a live network.
Autonomous networking technologies relieve human staff of the more mundane aspects of network management, and open up the possibility of self-healing networks that can continually optimize themselves for best performance.
These are heady prospects, but if our survey respondents are right, we won’t see full and widespread autonomy in the near future. By 2028, organizations believe that only 29.4% of their network operations will be AI-autonomous, and the rest will either be AI-assisted (38.4%) or managed by people in the traditional way (32.2%).
Standing Still Comes at a Cost - But So Does AI Deployment
Conversations about AI revolve heavily around efficiency and productivity. But our survey showed that organizations don’t expect AI in network operations to save them money – in fact, they expect it to increase costs by 12% by 2028.
The likely explanation for this is that implementing AI across the business puts more strain on the network, so organizations need to plan capacity needs more accurately and for their networks to perform better.
Already, nearly half of organizations (48.7%) are using AI to improve network planning and capacity forecasting. This improves network cost-effectiveness – but designing, implementing, and managing these models also incurs expenses in terms of compute power and skilled professionals. Without AI-in-networking, though, it’s likely that the cost to the business of poor network performance or unused capacity would be even higher.
Providers Need To Step Up To Give Enterprises the Support They Need
Despite the hurdles, the reality is that AI adoption is a business imperative for most organizations.
There are a lot of moving parts here: unavoidable security concerns; the race to keep up with skills and approaches in a field that’s moving incredibly fast; and the fact that costs are rising so quickly that Goldman Sachs has reportedly said many large organizations are overrunning their AI inference budgets “by orders of magnitude”.
If providers can overcome the trust gap, enterprises will almost certainly be searching for their support to navigate this complexity – but this raises the question of who’s best suited to provide that expertise.
According to our survey, 44.9% of respondents expect cloud hyperscalers to lead the AI-in-telecoms race by 2030. Whichever direction this technology takes next, the providers that enterprises choose to partner with must be able to offer the data security, governance, sovereignty, price, and local presence that are all needed to make AI-in-networking a demonstrable success.