For decades, telecom operators built their networks around one assumption: the customer is a person. People watch videos, browse social media, make calls and then go offline. Traffic peaks in the evening, downloads outweigh uploads, and pricing, spectrum allocation and quality of service have all evolved around those predictable patterns.
The rise of AI agents is beginning to challenge that assumption.
Unlike human subscribers, AI agents do not stop after completing a task. They remain active, exchange information with enterprise systems, retrieve data from cloud services and increasingly communicate directly with other AI agents. As enterprises move from experimenting with a handful of AI assistants to deploying thousands or even millions of autonomous software workers, operators may find themselves serving a new type of network customer that behaves nothing like the people their infrastructure was originally designed for.
Why AI traffic runs uphill
The first sign of that shift appears in the direction data travels. For years, telecom networks have been engineered around downloads because that is how people use the internet. Streaming video, browsing websites and scrolling through social media all involve far more data flowing towards the user than away from them.
AI changes that balance.
Ericsson analysed traffic across 55 telecom operators worldwide during 2025 and found uploads were growing faster than downloads at 43 of them. At 17 operators, upload traffic was increasing more than one and a half times faster than download traffic.
The reason lies in how AI applications work. A traditional internet session mostly involves receiving information. An AI interaction starts by sending information. Every prompt, document, photograph or voice recording has to travel to the cloud before a response comes back. Ericsson estimates that a text-based AI conversation already operates at roughly a 29:71 split between upload and download traffic, a sharp departure from the traditional download heavy pattern. Voice assistants and autonomous AI agents move even closer to an equal split, while devices that continuously stream audio and video, such as AI glasses, increase uplink demand further.
Ericsson's modelling suggests the change could accelerate rapidly. Under its medium adoption scenario, uplink traffic in an urban area could triple between 2025 and 2031. Under a high adoption scenario, driven by widespread use of AI glasses alongside industrial devices such as smart cameras and autonomous vehicles, uplink traffic could increase fivefold over the same period.
The implication is that networks designed primarily to deliver entertainment are beginning to carry a growing volume of machine generated traffic moving in the opposite direction.
The shift is no longer theoretical. Recent benchmarking by Ookla found India's 5G performs strongly on download speeds but considerably less well on upload capacity and latency, the two characteristics becoming increasingly important for AI workloads. India's median upload speeds remain below the threshold Ookla considers suitable for more demanding AI applications, while latency is only just sufficient for today's text based AI experiences. More intensive workloads, such as augmented reality and real time multimodal AI, remain beyond the capabilities of virtually every network included in the study.
The challenge is partly architectural. Much of India's 5G network uses Time Division Duplex spectrum, where uploads and downloads share the same frequencies. Increasing capacity in one direction inevitably reduces capacity in the other, making the shift towards AI workloads more complicated than simply allocating additional bandwidth.
An agent does not behave like a person
The challenge extends beyond the volume of traffic. It is also about how AI agents consume network services.
Human users tend to generate predictable demand. They open an application, complete a task and move on. AI agents can repeatedly invoke the same network capability, combine multiple services in unexpected ways and interact continuously with enterprise applications and other AI systems. That makes traffic patterns significantly harder to anticipate.
Telecom executives are already preparing for that possibility.
At the GSMA's Open Gateway Summit earlier this year, executives from operators including Deutsche Telekom, Vodafone, Orange and Telefónica argued that autonomous AI agents will fundamentally change how network capabilities are consumed. Instead of occasional API requests from human developers, operators expect software agents to become continuous consumers of network services, creating new challenges around pricing, authentication, security and service guarantees.
The discussion also reflected a broader shift taking place across the industry. Networks are no longer being viewed simply as pipes that carry data. Increasingly, operators want them to expose programmable capabilities that software can discover and use automatically.
That has implications beyond traffic management.
Sara Sebti, chief executive of digital identity company ShareID, argued at the same event that traditional authentication methods such as one time passwords are becoming less effective against AI powered fraud. Instead, she said operators would increasingly need to combine network signals, verified identities and biometric authentication into a layered trust model capable of serving both humans and autonomous software.
Lin Sun of Solo.io demonstrated another aspect of that future through the Model Context Protocol, or MCP. Rather than integrating separately with every service through conventional APIs, AI agents using MCP can discover available tools, invoke external systems and complete complex workflows through natural language interactions. In one demonstration, an agent configured infrastructure, updated network settings and generated software code through conversational prompts rather than manual API integration.
For telecom operators, the implication is straightforward. Network capabilities that cannot be discovered and invoked automatically by AI agents risk becoming invisible to a growing category of software users, regardless of how advanced the underlying network may be.
Ericsson reaches a similar conclusion in its white paper on AI agents. The company argues that operators will increasingly need to expose network capabilities, including location services and guaranteed quality of service, in forms that autonomous software can consume directly. That also requires new mechanisms for authenticating machine identities, discovering available services and governing communication between AI agents themselves.
Broadband feels this too
Much of today's discussion around AI connectivity focuses on 5G and, eventually, 6G. But the earliest impact may be felt on broadband networks, particularly in offices, factories, hospitals and data centres where enterprises are deploying AI applications at scale.
Unlike a consumer streaming films or browsing social media, an enterprise running AI assistants, software agents or industrial automation generates continuous two way traffic between cloud infrastructure, enterprise applications and connected devices. That places greater emphasis on upload capacity, low latency and predictable performance than conventional broadband networks were originally designed to deliver.
Fixed Wireless Access (FWA), where homes and businesses receive broadband over a 5G connection instead of a fibre cable, is expected to play a growing role in that transition.
Ericsson estimates global FWA connections will increase from 185 million in 2025 to 350 million by 2031, serving nearly 1.4 billion people. In India, Reliance Jio and Bharti Airtel together had already crossed nearly 17 million FWA connections by early 2026.
The same AI driven traffic patterns apply here.
A household using an AI assistant to analyse documents or generate videos uploads significantly more information than someone simply streaming entertainment. In enterprises, the effect is even more pronounced as AI agents continuously exchange information with cloud services, internal databases and other software systems throughout the day.
FWA already accounts for around 28 per cent of global mobile network traffic, a share Ericsson expects to rise to 36 per cent by 2031. As AI workloads become more common, broadband providers may face many of the same network challenges as mobile operators, particularly around uplink capacity and latency.
What operators can actually sell
For operators, the emergence of AI agents is not simply about carrying more traffic. It may also change what connectivity itself becomes.
For decades, telecom companies largely sold access to the internet through mobile and broadband subscriptions. AI applications increasingly require something more specific. They may need guaranteed latency for a factory robot, additional uplink capacity for a vision model, verified identity for financial transactions or temporary priority while completing a critical task.
Instead of paying for a fixed connectivity plan, future applications could request those capabilities automatically, consume them for a few minutes or hours and pay only for what they use.
That is the direction much of the telecom industry is moving towards.
Ericsson found that the number of commercial offerings based on 5G network slicing increased from 65 to 84 within six months. Rather than treating every user equally, network slicing allows operators to allocate dedicated network resources with predefined performance levels for specific applications or customers.
The company also found that businesses increasingly expect AI applications to rely on real time data. About 88 per cent of enterprises surveyed said real time connectivity would become important for their AI deployments. Yet only 18 per cent have implemented the mobile technologies needed to support those applications, while just 8 per cent have scaled AI across their organisation.
That gap presents an opportunity for operators to move beyond selling bandwidth towards selling assured network capabilities.
Whether that translates into meaningful new revenue streams remains to be seen. Many enterprises are still in the early stages of deploying AI agents, and standards around machine identities, network APIs and agent to agent communication continue to evolve.
What already appears clear, however, is that AI is beginning to change the nature of the network customer itself.
 "AI economy is creating a new kind of telecom customer. Are networks ready?")
