Report: Two-Thirds of Companies Running AI on Networks That Can't Handle It

Companies are spending millions on AI. Fewer are spending anything on the networks that AI actually runs on. That mismatch is now showing up as a measurable business problem, according to a Bloomberg study titled "The Future-Ready Enterprise," commissioned by Tata Communications.

The numbers are blunt. Three in four corporate leaders call AI a board-level priority, the study found. But nearly two-thirds of enterprises, 65%, are still running that AI on transitional or legacy network infrastructure never designed for it.

It's the gap between what executives say they want and what their systems can actually deliver.

Why Old Networks Can't Keep Up

Traditional business software could tolerate slow networks. A sluggish email server or a laggy CRM login, annoying but survivable, with 100 to 500 milliseconds of latency considered acceptable.

AI doesn't work that way. Mission-critical AI workloads now require latency under 10 milliseconds, according to Tata Communications.

"This isn't just an incremental improvement," said Kapil, Vice President of Global Network Services at Tata Communications. "It's a completely different performance paradigm that breaks traditional network design assumptions, where such extreme low latency was never a primary consideration."

Legacy networks were built static and rigid. They weren't designed to flex in real time. AI traffic, by contrast, is unpredictable and constant, driven by continuous inference, machines talking to other machines ("agent-to-agent communication" in industry jargon), and real-time data pipelines that never really shut off.

A Cisco study cited in the report found 80% of executives believe their company's competitive survival depends on agentic AI, meaning AI systems that act autonomously rather than just answering prompts. If that's true, and companies keep running those systems on networks built for a pre-AI world, the mismatch gets more expensive, not less, over time.

The Real Cost of a Slow Network

This stops being an IT problem and becomes a money problem. A fraud-detection model or a supply-chain optimization tool is only as good as the data it gets in real time. If network congestion delays that data, the model's output arrives too late to matter. The AI didn't fail. The pipe carrying its data did.

"Relying on a 'best-effort' network turns multi-million-dollar AI stack investments into a high-stakes gamble, where performance is left to chance," Kapil said.

Companies routinely test AI systems domestically, where performance looks fine, and then discover problems once data starts crossing borders or hitting international cloud platforms. The public internet, Kapil notes, is not built to function as a reliable global enterprise network. Performance that looks acceptable in one country can fall apart once the same workload has to route through multiple international carriers and cloud providers.

What This Means, and What It Doesn't

To be fair to skeptics of the "networks are the bottleneck" narrative: this study was commissioned by Tata Communications, a company that sells network infrastructure services. It has a direct commercial interest in convincing enterprises that their current network setup is inadequate and needs replacing. That doesn't make the underlying data wrong, but it does mean the framing deserves scrutiny rather than automatic acceptance.

Still, the core numbers, that AI adoption is outpacing infrastructure readiness at most large companies, track with a broader pattern seen across enterprise tech. Companies have a long history of buying software faster than they upgrade the hardware and networks underneath it. Cloud migration hit similar friction a decade ago. AI is just the latest, and most latency-sensitive, version of that problem.

The unresolved question is how much of this 65% figure reflects genuine technical incapacity versus companies simply not having gotten around to network upgrades yet. Tata's study doesn't break out how many of those enterprises have upgrade plans already budgeted, or on what timeline. Until companies start reporting actual AI deployment failures tied specifically to network latency, rather than general dissatisfaction with legacy systems, this remains a warning about risk rather than a documented wave of failures.

What is measurable: the latency requirement gap is real and specific. 10 milliseconds versus the old 100-to-500 millisecond tolerance is a difference of at least an order of magnitude. Any enterprise running latency-sensitive AI, like real-time fraud detection, on networks built for that older standard is operating with a documented technical mismatch, whether or not it has caused a visible failure yet.