AI in Network Operations: Why Companies Are Speeding Up and Braking at the Same Time

From pilot project to business as usual

The figures are clearer than many would expect: more than three quarters of the companies surveyed already use AI-supported network solutions in production. A good one in four companies even describes its network operations as fully AI-integrated or autonomous. Only a barely measurable minority currently has no plans in this direction.

So the question is no longer whether AI is moving into network operations, but in what role, with which guardrails and with which partners. Those who draw the right conclusions gain a strategic advantage. Those who ignore them build in new dependencies.

Four contradictions you should know

The most interesting thing about the survey is not the approval ratings, but the tensions behind them. Four of them shape the current debate.

First: a larger attack surface and yet confidence. A clear majority expects AI to make networks more vulnerable, and around six in ten respondents believe this calls for a fundamentally different security concept. At the same time, about two thirds believe that defence will ultimately stay ahead in the race against AI-supported attacks. This cautious confidence is remarkable, and it relieves no one of the task of bringing architecture and governance up to date.

Second: distrust of AI decisions combined with a high willingness to delegate. Faulty automated security decisions are among the biggest concerns. Nevertheless, around seven in ten respondents would accept an AI system independently isolating substantial parts of the network when an attack is detected, provided a defined confidence threshold has been reached. This is not a contradiction born of carelessness, but a sign of how central clean threshold architectures and documented escalation paths have become.

Third: expertise in place and yet the biggest concern. The overwhelming majority credit themselves with the in-house know-how needed to run autonomous networks. At the same time, two thirds name exactly the opposite as their biggest risk: introducing AI without the ability to properly control it afterwards. A plausible explanation: companies trust themselves with day-to-day operations but fear the introduction phase, meaning architecture decisions, integration and the choice of the right partners.

Fourth: looking for partners, keeping control. External support is explicitly wanted, with model training and security monitoring at the top of the list of what companies would like to hand over. On one point, however, the willingness stops abruptly: the obligation to provide evidence and audits for AI security decisions is to remain in-house. This is perhaps the most important insight for any sourcing decision. Operations can be delegated, accountability cannot.

Where AI actually makes an impact today

Unsurprisingly, the most common area of use is customer service with intelligent assistance systems. Practically level with it, however, comes the detection and defence of security threats, followed directly by a field that is immediately relevant for network managers: capacity planning and forecasting.

Traditional capacity planning worked with historical utilisation values and patterns of experience. That is exactly where it fails today. Cloud usage, connected devices and AI applications themselves generate load profiles that defy simple extrapolation, resulting in bottlenecks, performance drops and, in extreme cases, outages. How quickly the network becomes a bottleneck usually only shows under load. AI-supported methods forecast demand much more precisely and shift network management from reaction to anticipation.

Other common fields of application:

• Optimisation of wireless networks and predictive maintenance
• Automated policy control in SD-WAN and SASE environments
• Management of satellite-based connections, with increasing adoption

Each of these applications shares one prerequisite: reliable, consistently measurable connectivity. A forecasting model working on patchy telemetry data from five countries and three carriers does not deliver reliable results. Intelligence in the overlay is only as good as the underlay it builds on, and in globally distributed environments that underlay is rarely available from a single source. How AI-enabled networks are built is therefore decided below the model level.

The real bottleneck is trust, not technology

Respondents name neither costs nor missing technology as the biggest brake on AI adoption, but security, data protection and data quality issues. In second place come technical integration problems, explicitly including legacy systems and vendor lock-in. Only then follow regulatory uncertainty and the shortage of skilled staff.

It is particularly revealing where the distrust actually takes hold: companies rate most critically those managed providers that pool data from several customers in shared models. Close behind are the security practices of the AI providers themselves. The practical consequence described by many respondents: adoption slows down, resources flow into due diligence and compliance, and some companies prefer to avoid external providers altogether. That is rarely the best answer economically, but an understandable one.

In the DACH region, this picture becomes even sharper. Data sovereignty, place of processing and the traceability of automated decisions are not secondary conditions here, but hard requirements, both from a regulatory perspective and in internal auditing.

Why the network gets more expensive at first

One finding is likely to spark debate in budget discussions: by 2028, respondents expect AI to make network operations not cheaper, but on average around twelve percent more expensive. This contradicts the original promise of efficiency, but on closer inspection it makes sense.

AI in the enterprise is network-hungry. It demands more capacity, more consistent performance and tighter cloud connections. On top of that come the operating costs of the AI models themselves: computing power, storage, data preparation, integration and qualified staff. The realistic calculation is therefore not that AI saves network costs, but rather: without intelligent control, the costs of poorly utilised bandwidth and oversized reserves would be even higher. In addition, a widely noted MIT study shows that the vast majority of generative AI pilot projects demonstrate no measurable financial effect, an argument for careful selection, not against AI.

If AI increases network costs in the short term, the purchasing of connectivity itself becomes the decisive balancing factor. This is where, in practice, the biggest untapped lever lies: price levels for business internet, DIA, Ethernet and site networking differ considerably between carriers, with identical technical performance.

Who decides and who is accountable

It is remarkable how high up AI decisions are now made. Most often, technical and security management are to decide jointly, followed in second place by executive management itself. Network architecture has thus definitively become a matter of corporate governance and not a specialist topic for the IT department. How strongly network sovereignty becomes a leadership issue shows in every contract negotiation.

A clear pattern is emerging for operational responsibility: a joint model of internal team and external partner is in the lead, while a third want to keep control exclusively in-house. When it comes to auditability, the picture tips even further inwards. Translated into contract requirements, this means: anyone buying AI-supported network operations needs enforceable rights to information, logging and evidence, not just a service level for availability.

When asked about their AI strategy in network operations, one model is clearly in the lead: a hybrid approach across several partners. Well behind follow hyperscaler platforms and specialised managed service providers. The logic behind this is understandable, as no single provider covers all requirements. The hybrid approach, however, comes at a well-known price: complexity. Several partners mean several control logics, several data sources, several sets of contracts and several escalation paths. Exactly the side effect that vendor sprawl also creates in traditional networks.

This conflict of objectives can only be resolved with a clear orchestration layer: a partner that consolidates the provider landscape without being a provider itself, and therefore has no vested interest in any particular technical solution. From the perspective of globally positioned companies with multiple sites, three models compete:

Predictive and simulation-based methods, meaning analytics and digital twins, are considered the most transformative technologies of the coming years, just ahead of autonomous, self-healing networks and generative AI for network design. When asked who will dominate the field by 2030, the answer is clear: hyperscalers are expected to lead, traditional telecommunications providers hardly at all. What remains open, however, is whether large platform operators can deliver the data sovereignty and local presence that many companies need for compliance reasons.