The Network Cost of AI: From User Prompts to Agents

AI changes the unit of network demand.

The useful planning question is not “How much bandwidth does AI use?” It is “How much traffic does this workflow generate, how long does it remain active, and where does it cross the network?”

This brief is designed for enterprises with 600 or more workers. That does not mean every employee becomes an AI user on day one. It means the access layer has to absorb adoption as it spreads across teams, floors, buildings, and branches. At that scale, aggregate concurrency—not one person’s prompt—is what turns a small workflow into a material infrastructure decision.

AI does not dominate the enterprise link today. Cisco’s report says inference traffic remains negligible beside major categories such as video in the near term. The planning signal is the combination of rapid adoption and a different traffic shape: longer connections, more upstream traffic, and more overlapping model and tool calls.

Traditional web demand is often described as short, bursty, and mostly downstream. AI inference has a different signature. In Cisco’s 2026 AI Impact on Wide Area Networks report, measured AI inference flows lasted about twice as long as non-AI web flows, while the median regular web flow rate was ten times higher. AI was not always a bigger burst. It was a smoother connection occupying the network for longer.

Direction changes too. Nine percent of measured AI inference flows were upstream-heavy, compared with roughly 0.5% of ordinary web transactions. The median downstream-to-upstream ratio narrowed from 145.39:1 for non-LLM traffic to 21.11:1 for AI traffic. That matters at branches and campuses designed around download-heavy behavior.

Transport behavior adds another planning dimension. AI services in the study used both TCP and QUIC in a nearly even flow-count split, but QUIC carried 57% of AI data volume. That can change what monitoring and inspection tools can see. The takeaway is not that every AI application has the same profile. It is that a capacity model based only on yesterday’s short, downlink-heavy web session is incomplete.

Tokens price the model. Bytes load the network.

They often rise together in an agentic workflow, but they measure different things and should never be converted with one universal ratio.

Tokens are the units models process. Input tokens include the prompt and context sent to the model; output tokens are generated by the model. Providers may also distinguish cached input and reasoning tokens. API cost is commonly based on those token categories and the selected model’s rates.

Agentic AI can multiply token consumption in three ways. First, one task can require many model calls. Second, later calls may carry prior messages, tool results, source excerpts, or working state back into the context. Third, retries, verification passes, and parallel agents can repeat work that the user never sees. The useful unit is therefore tokens per completed task, not tokens in the final answer.

This is why a token forecast is not a bandwidth forecast. The same token count can travel with different serialization, encryption, streaming, caching, and tool data. Conversely, a large file transfer can add substantial traffic while contributing few model tokens if it is filtered or processed before inference. Measure both ledgers and correlate them by completed task.

Calculate the task before you size the link.

Start with a simple workload profile. Open the technical view when you need to control model calls, token categories, prices, traffic, duration, placement, and concurrency directly.

The workload card defines what happens inside one completed task. The adoption slider controls how often that task runs and how many workflows overlap: task frequency, peak-active share, and parallel workflows. Its five milestones are transparent planning positions—not an adoption forecast.

The default agentic profile is deliberately transparent: 600 enabled workers, two completed tasks per user per day, 12 model calls per task, 8,000 average input tokens and 1,500 output tokens per call, 25% cached input, and editable planning rates of $1 per million uncached input tokens, $0.25 per million cached input tokens, and $5 per million output tokens. Those are assumptions, not a claim about a particular model or deployment.

Switch to the Cisco research profile to load the measured 26.8 MB workflow-traffic value. Its model-call and token inputs remain illustrative because Cisco’s report publishes the network result—not a token ledger for that task. The calculator keeps that boundary visible instead of inventing a token count and labeling it measured.

Daily tokens, daily cost, and daily traffic describe cumulative consumption. Peak Mbps describes simultaneous pressure at one link. The same organization can have a large monthly model bill with little office-link impact if agent fan-out stays in a provider cloud—or modest token usage with a large local transfer if workflows move documents, images, or datasets across an enterprise boundary.

The quiet risk is concurrent state.

A lower-rate flow is not necessarily a lighter operational burden when it remains open longer and overlaps with hundreds of other tasks.

Longer-lived AI connections change more than utilization. They can increase the number of simultaneous sessions held by firewalls, proxies, network address translation, load balancers, and monitoring systems. Agentic workflows add fan-out: one user task can create multiple model and tool conversations before the answer returns.

That is why average daily volume and peak Mbps belong in the same model but answer different questions. Daily volume informs transit, egress, and cost. Peak Mbps informs immediate link pressure. Concurrent flows, duration, and transport mix inform the stateful infrastructure in between. None should be inferred from only one of the others.

Latency needs similar care. Cisco notes that today’s model inference time can dominate the network—for example, tens of milliseconds of network latency inside a response that takes seconds to generate. As inference hardware becomes faster, the network becomes a larger share of the experience. The practical response is not to blame or absolve the network in advance. It is to measure model time and path time together.

Build the baseline before adoption hides it.

The goal is not one universal “AI bandwidth” number. It is a repeatable measurement method for each important workflow and network boundary.

The result should be a living workload profile: completed tasks, model calls, input and output tokens, cache share, model cost, MB per task, active seconds, upload ratio, concurrent flows, transport mix, path placement, peak starts per minute, and user experience. Rerun it when the model, agent loop, tool set, data source, pricing, or hosting pattern changes.