Every decade or so, a core assumption gets quietly wrong. Not obviously, not all at once, but wrong enough that the companies who catch it early build an advantage the others spend years trying to buy back. Cisco's new research on how AI changes network traffic is one of those moments. The numbers are striking. The implication underneath them is more important.

The study is based on real traffic data measured across live networks, plus hands-on testing of AI agents in a lab. Its central finding: when an AI agent handles a task instead of a human, the task generates 450% more network traffic. Without AI agents running in enterprise systems, corporate network traffic is expected to grow roughly 2.5 times over the next decade. With AI agents embedded in everyday workflows, that growth becomes approximately 9 times. Jeetu Patel, Cisco's President and Chief Product Officer, went further in a recent LinkedIn post: he believes even the 9x figure may prove wildly conservative, and that what the model projects for a decade could materialize in three years.

That is a precise claim from someone whose business depends on getting network planning right. It deserves scrutiny, not applause.

The Internet Was Built for Downloading, Not Thinking

For thirty years, the internet was designed around one basic pattern: a server somewhere holds content, and a person somewhere else requests it. A video streams down to your laptop. A file downloads to your phone. Almost all the traffic flows in one direction, from the source to the person.

AI agents break that pattern completely.

An AI agent doesn't just request information and wait. It thinks, searches, decides, acts, and then thinks again, sending large amounts of context back to the AI model with every step. Cisco's research finds that roughly 9% of AI-related network connections carry more data going up to the server than coming back down, compared to less than 1% for ordinary web traffic. The reason: every time an agent takes an action, it has to send everything it knows about the task back to the model so the model can decide what to do next. The network isn't just delivering answers. It's continuously ingesting the agent's running memory.

Most corporate networks were set up to handle heavy traffic coming in and light traffic going out. That was the right call when the heaviest users were humans downloading reports and watching dashboards. It is the wrong call when the heaviest users are software agents sending their working memory upstream thousands of times a day. The mismatch won't announce itself loudly. It will look like an AI performance problem, and the real cause will take time to find.

Most of That Traffic Is the AI Thinking, Not Moving Files

Of all the extra traffic that AI agents create, roughly 70% is the AI model doing its reasoning work, not files being transferred or videos being streamed. This distinction matters because reasoning traffic has different requirements than ordinary data.

When a human watches a video and a few frames drop, the picture gets blurry for a second. When an AI agent's connection to its model degrades, it produces a wrong answer that may already have been acted on before anyone notices. The connection between an agent and the model it runs on is, as Cisco puts it, the agent's "spinal cord." Interrupt it and the agent stops working correctly.

There is a second problem. Cisco finds that 57% of AI reasoning traffic uses a newer type of internet connection called QUIC, which is designed to be faster and more encrypted than older connection types. The side effect is that most enterprise security tools, which work by reading traffic as it passes through, cannot see inside QUIC connections easily. The traffic that most needs monitoring is increasingly the traffic that is hardest to monitor.

Networks Were Planned Around Human Work Schedules. Agents Don't Have One.

Corporate networks are typically sized based on when humans are most active: busy in the morning, quieter at lunch, tapering off in the evening. Those peaks and valleys give IT teams room to manage capacity.

AI agents have no schedule. They run at full speed around the clock. Cisco's research describes the difference as a shift from "bursty" traffic, which spikes and then drops off, to "sustained" traffic, which stays high continuously. A chatbot creates spikes when someone asks it a question. An AI agent embedded in a business process runs constantly, whether anyone is watching or not. A network sized for human peaks has no reserve capacity for a workload that never dips.

There is also a less obvious consequence for security systems. Most enterprise firewalls and security monitoring tools are designed to track individual connections and remember their state for a short time. AI reasoning connections last roughly twice as long as ordinary web connections and involve more back-and-forth. At scale, this means security systems will need to track far more active connections at once than they were built for.

Putting AI Closer to Users Is Not Yet About Speed

One common argument for deploying AI infrastructure closer to employees, in a regional office rather than a distant data center, is that it reduces the delay between asking a question and getting an answer. Cisco's report is direct on this point: that argument does not currently hold up.

Today, the delay in getting an AI response is driven almost entirely by how long the model takes to think, not by how long the data takes to travel. A response can take anywhere from half a second to several seconds depending on the model and the question. The network's contribution to that delay is typically 20 to 50 milliseconds, a small fraction of the total. Buying closer infrastructure to reduce that fraction is not the best use of the investment right now.

The honest case for placing AI infrastructure closer to users today is about cost, data control, and the ability to monitor what the AI is doing. The speed argument will become real as AI models get faster. When model processing time drops to milliseconds, the network delay becomes a meaningful part of the total experience. Cisco projects the period between 2029 and 2032 as when AI network traffic growth will be most aggressive, with a compound annual growth rate approaching 25%. Decisions made about network architecture in 2026 and 2027 are the ones that will either hold or fail in that window.

What Cisco Is Selling, and What the Data Still Shows

Cisco sells the networking equipment and software that enterprises would need to upgrade if this research is right. That commercial interest is worth naming. It does not make the underlying findings wrong, but it is context a CIO should hold when evaluating the urgency of any specific recommendation.

What makes this report more credible than typical vendor research is that it is based on actual traffic measurements from real networks, not just projections built from models. The work was done in collaboration with Opanga Networks, with lead researchers Johan Gustawsson, Javier Antich, and Waris Sagheer from Cisco. Cisco has also committed to repeating this study annually, which means the projections will be tested against real data over time.

The core argument is straightforward. The internet was built to move content from a few places to many people. AI agents reverse that flow. They are everywhere, running constantly, and sending large amounts of information back to centralized models to think. Buying more bandwidth helps with volume. It does nothing for the direction problem, the security visibility problem, the session tracking problem, or the scheduling assumptions baked into how networks were sized. Those are design problems, and they need design answers before 2029.