AI has moved fast from experiment to expectation. Enterprises use it to read and analyze data, automate everyday tasks, and support customers at scale. As adoption climbs, the pressure on IT and networking leaders isn't just to deploy AI. It's to make sure the infrastructure underneath it can actually carry the load.
When we surveyed IT and networking professionals, the most common enterprise use of AI was customer support, at 52.4%, followed closely by security threat detection and response, at 51.8%. Nearly half of enterprises, 48.7%, use it for network planning and capacity forecasting.
Here's the part that gets less attention. AI reads data, makes decisions, and executes workflows on its own, and that raises real questions: how people use it, how shared data is tracked, how the systems behave under pressure. Our research found that enterprises expect AI to push network operating costs up by around 12% by 2028, simply because AI is more demanding of the network and needs more capacity and performance to run well. Enterprises are already planning around this figure, and it isn't hypothetical.
Why AI Breaks the Old Network Playbook
For a decade, the job was straightforward: move to SD-WAN, optimize for cost, run it over broadband. The old approach worked because the traffic it carried, mostly SaaS and email, tolerated best-effort delivery.
AI traffic doesn't tolerate that. It's more bidirectional, more persistent, and far more sensitive to latency and packet loss. As inference moves closer to the edge, the network stops being simple transport and becomes part of the performance path itself. Best-effort, asymmetric broadband starts to fail the requirements AI places on it. So, the mission has to shift from cost optimization to performance engineering.
The shift is both technical and commercial. SD-WAN needs to identify AI traffic deterministically, prioritize it, and protect it end to end, over paths that meet a defined performance bar rather than whatever path happens to be cheapest. In practice, that means moving from cost-based routing to intent-based networking, aligned to AI workloads, with performance orchestrated across cloud, data center, and edge.
On the commercial side, SLA-backed access, symmetric bandwidth, predictable latency, and committed performance are coming back into focus. The shift pulls the underlay away from low-cost broadband and toward dedicated internet access and higher-quality connections, even on secondary links. Enterprises are increasingly buying a committed outcome, not a list of circuits, and AI is what makes that outcome measurable in real time rather than a promise on a slide.
We're not the only ones seeing this gap. A 2026 Cisco and Foundry survey of IT and networking leaders found AI-driven branch traffic up 34% year over year, but only 15% of organizations said their network was flexible enough to support AI at scale, and 73% expected to hit capacity limits within 24 months. The readiness problem shows up in one data point: Demand is already outrunning the infrastructure most enterprises run today.
What a Provider Should Actually Be Willing to Say
Plenty of providers will tell an enterprise that AI is going to transform their business, make them faster, leaner, more innovative. Some of that may even be true. But it isn't something a network provider controls, and it isn't something a network provider should be promising on someone else's behalf.
What a provider can control is narrower, and that's exactly why it matters more: uptime, latency, performance, security posture. Those are the commitments that should anchor every conversation with a client, because they're the ones to which a provider can be held accountable. If it's launched, say what it delivered. If it isn't, say when it will be and what it's expected to do. Vague commitments are worse than no commitment at all, because they erode trust the moment someone checks.
Accountability should sit in one place. Enterprises running across dozens of countries, often across dozens of carriers, clouds, and vendors, need a single relationship responsible for designing, delivering, securing, and operating the whole network, not a patchwork where no one owns the outcome. Independence from any one carrier or vendor matters here, too. It means building what's right for the client, not what happens to suit the provider's own footprint.
Security Is Part of That Responsibility, Not a Separate Line Item
The threat surface around AI is genuinely different from what came before it. Prompt injection, unsecured IoT devices, shadow AI: these are risks that basic firewalls were never built to catch, and most enterprises already know it. Our research found that data privacy and AI training data exposure concerns 47.1% of those surveyed, and 67% cite a lack of in-house expertise to manage AI in network operations as their single biggest concern.
The cost of getting this wrong is climbing fast, not staying flat. IBM's 2026 Cost of a Data Breach Report found shadow AI incidents more than doubled year over year, from 20% to 43% of breached organizations, averaging $5.39 million per breach. The risk isn't a slow build that enterprises can plan around at leisure; the numbers got materially worse in a single year.
A responsible provider treats AI security as inseparable from network security. In practice, security work includes runtime protection and data loss prevention for AI workloads, plus gateway controls that manage how many tokens a system can exchange, a cost control as much as a security one. An AI system with no guardrails, no human on the decisions that matter, and no visibility into what it's doing is the risk. An observable, guard-railed, cost-capped system is the fix.
Responsible security also means being precise about what autonomy actually is, because the word gets used loosely. Autonomous is not the same as fully automated. The credible near-term system investigates, diagnoses, and recommends on its own, and acts within guardrails, while a human signs off on anything consequential. The limitation doesn't need explaining away, and the position isn't just ours. Gartner's own adoption forecasts for automated network operations still track well below full autonomy over the next several years, which lines up with what we're seeing directly with clients. Across a multi-vendor, multi-carrier, multi-jurisdiction network, one wrong autonomous action can carry an enormous blast radius, and that's exactly why the human stays in the loop. Accountability is what the client is actually paying for.
The Underlay Squeeze Enterprises Are About to Feel
Enterprises are investing heavily in AI while also funding the infrastructure, cloud capacity, security, and expertise it takes to run it. Increased AI investment doesn't automatically mean spending more on the network. It means spending more deliberately.
The risk is that enterprises cut the underlay to free up budget for AI, which is self-defeating. AI doesn't perform well on a network that wasn't built to carry it, no matter how good the model sitting on top of it is. Cutting the underlay, the physical infrastructure that actually transports data and supports latency and capacity, undermines reliability, user experience, cost efficiency, and the network's ability to adapt as needs change. Don't cut the network to fund the AI. The AI just underperforms.
This is where a managed, accountable partner earns its place: helping enterprises balance underlay and AI spend against what they're actually using, not what the budget assumed going in.
What Enterprises Should Expect Going Forward
Over half of enterprises in our research, 50.4%, believe predictive and simulation technologies will transform networking over the next three to five years. Autonomous networking follows at 49.6%, generative AI at 48.7%.
The technology will change how networks are operated. What changes more is what enterprises expect from the organizations responsible for running them. Most aren't looking to handle this alone: 41.6% want shared management of AI decisions with an external partner, and 39% are already running a hybrid model in network operations.
We act as that single point of contact when a network or security issue crosses vendors, carriers, or borders. We use AI to help networks run more efficiently, and we help enterprises secure AI workloads, strengthen security tooling, and work out what needs to move where, while guarding against prompt injection and shadow AI. We're asset-light; there's no fiber build, no GPU farm to justify, so the capital goes toward what actually improves outcomes for the client.
As AI usage grows, the network carrying it becomes more important, not less. The providers responsible for it need to be precise about what they can control and honest about what they can't. Holding to that standard is harder than making a bigger promise, but it's the one that holds up.
Author
Gina Nomellini - Chief Operating Officer
Gina brings extensive communications and technology leadership experience across startups and large global organizations to her role as Chief Operating Officer at Globalgig.
Before Globalgig, Gina served as CMO at GTT Communications, leading brand development, marketing communications, and product functions, and as CMO at One Source Networks prior to its acquisition by GTT. Before those roles she served as Vice President of Global Product and Marketing at Telstra, with earlier positions at Level 3 Communications and Broadwing. She holds a Bachelor of Arts from the School of Foreign Service at Georgetown University, an MBA from the University of Texas at Austin, and serves as a Global Ambassador for Télécoms Sans Frontières, the world’s first emergency technology NGO.
At Globalgig, Gina leads operations, engineering, service delivery, product, marketing, and vendor management. The breadth of that remit is intentional: these are not separate disciplines at Globalgig. They are the same conversation, led by someone who has run each of them at scale.