Three years into the LLM era, "AI in networking" has become a category badge rather than a description. Almost every vendor claims it. A useful exercise is to ignore the brochures and look at what AI is provably doing in production today, what is plausible in the next 12 months, and what is still wishful thinking.
What is provably working
- Anomaly detection on time-series telemetry — replacing static thresholds with model-derived "this is unusual for this device" signals.
- Multi-modal root-cause analysis — fusing telemetry, topology and history into a confidence-scored conclusion.
- Forecast-driven capacity planning — converting growth curves into procurement dates with usable accuracy.
- Predictive maintenance for optical and hardware components — SFP failure prediction, fan/PSU degradation.
- LLM-driven operational Q&A — "show me everything I know about ONT 4857" in seconds instead of minutes.
- Document understanding — RFP parsing, compliance matrix generation, automated knowledge-base build-out.
What is plausible inside 12 months
- Full closed-loop assurance for a defined service class — diagnose, fix, verify without human input in the loop.
- Natural-language configuration ("set up VLAN 500 with QoS for IPTV") with verification before push.
- Per-subscriber experience modelling — knowing which subscribers are about to churn by their network behaviour.
- Autonomous incident triage — every alert classified, scoped, and routed without a human first-pass.
What is still mostly hype
- Replacing the senior network engineer. Human judgment is still needed at the edges of policy.
- Designing the network. Topology design with novel constraints is mostly hand-built work.
- Choosing vendors. The decision is partly engineering, mostly procurement and politics.
- Generic "AI optimisation" of unspecified KPIs. If the vendor cannot tell you the metric, the metric is not real.
A useful test: who owns the loop
The crisp dividing line between real AI in networking and a chatbot is the loop. A chatbot answers a human's question. An AI in networking owns a loop — observation → reasoning → action → verification — and closes it without a human.
Vendor question that filters quickly
Ask: "Can your AI take a configuration action against a live device in production, and verify the action worked, without a human approval click?" If the answer is no — or "we can in roadmap" — you are looking at a chatbot, not an AI in networking.
Where AI fundamentally changes the operating model
The point of AI in networking is not to do the same work faster — it is to make work disappear. A network with closed-loop autonomy needs fewer engineers per million subscribers, but more architects per million subscribers. The job profile shifts from "operate the thing" to "design the policy that the AI operates the thing under."
“Automation makes the engineer cheaper per task. Autonomy makes most tasks not need an engineer at all.”
— NetXol design principle
Risk surfaces and how to manage them
- 1Confident wrongness — model is sure, wrong action taken. Mitigate with calibration tests, blast-radius limits, automatic rollback.
- 2Data poisoning — bad inputs in production drift the model. Mitigate with explicit data validation and red-team drills.
- 3Loss of human skill — engineers stop knowing the basics. Mitigate with deliberate runbook drills and education.
- 4Vendor lock-in via opaque models — your operational data ends up unique to the vendor. Mitigate by demanding open weights or open APIs to your own state.
