AI and Machine Learning in Networks
The newest exam topic in the CCNA v1.1 blueprint. After a network is instrumented and automated, the next problem is scale: a modern campus produces millions of telemetry points an hour and no human can read them all. This lesson explains what machine learning actually is, how predictive AI finds problems before users notice, where generative AI helps, and why the two are not the same thing.
Why networks need AI/ML
Automation solves the problem of doing the same task many times. It does not solve the problem of knowing what is wrong. A large network streams continuous telemetry — interface counters, CPU and memory, wireless signal strength, application latency, flow records — and the volume grows far beyond what an operations team can watch. The result is alert fatigue: so many alarms that real failures hide among the noise.
Three pressures push operations toward AI and machine learning:
- Volume. Millions of telemetry samples per hour cannot be read by people.
- Complexity. Cause and effect are indirect; a slow application may trace back to a single congested uplink or a flapping wireless channel.
- Speed. The business expects problems found and fixed before users report them.
Volume
Millions of telemetry samples per hour cannot be read by people, so a real failure is easy to lose in the noise.
Complexity
Cause and effect are indirect: a slow application may trace back to one congested uplink or one flapping wireless channel.
Speed
The business expects problems found and fixed before users report them, not after the ticket arrives.
What machine learning is
Artificial intelligence (AI) is the broad goal of machines performing tasks that normally need human intelligence — recognising a fault, understanding a question, choosing an action. Machine learning (ML) is one way to get there: instead of programming an explicit rule for every situation, you feed a model examples and let it learn the pattern.
Three ideas make ML work:
- Training data. Historical records — interface counters, past incidents, labelled faults — that contain the pattern to be learned.
- The model. A mathematical function, adjusted during training, that maps inputs (features) to an output (a label, a number or a category).
- Inference. Running live data through the trained model to get a prediction. Training happens offline; inference happens in real time.
Two common training styles appear in networking:
| Style | How it learns | Network example |
|---|---|---|
| Supervised | From labelled examples (input → known answer) | Classify a flow as benign or malicious using past labelled traffic. |
| Unsupervised | From unlabelled data; finds structure on its own | Group “normal” behaviour into a baseline so anomalies stand out. |
Predictive AI in network operations
Predictive AI analyses the past and present to forecast what is about to happen. It is the workhorse of network assurance. Typical uses:
- Anomaly detection. Flag behaviour that no longer matches the learned baseline — a sudden rise in CRC errors, a wireless channel getting noisier at a certain hour.
- Root-cause analysis. Correlate many weak signals to point at the one device or link that explains a cluster of symptoms.
- Forecasting and capacity planning. Project growth in bandwidth or client count so a link is upgraded before it saturates.
- Predictive maintenance. Estimate when a component is likely to fail from trends in temperature, errors or power draw.
- Dynamic baselining. Learn that “normal” differs at 9 a.m. on a weekday versus 3 a.m. on a Sunday, so thresholds adapt instead of producing false alarms.
Generative AI in network operations
Generative AI produces new content — text, summaries, code or configurations — usually from a large language model (LLM) trained on huge amounts of text and code. In network operations it acts as an assistant rather than a detector:
- Natural-language interaction. An engineer asks “which access switches in Building B have uplinks over 80% utilisation?” and the assistant turns that into the right API query.
- Summarisation. Turn a wall of syslog events or a long incident timeline into a short, readable explanation.
- Configuration assistance. Draft a command sequence from an intent such as “put ports 1–12 in VLAN 20 and enable port security”.
- Documentation and translation. Explain an unfamiliar CLI output, or convert a CLI task into an API call.
Generative output must always be verified. A language model predicts plausible text; it can be confidently wrong (a hallucination) or propose a command that is syntactically valid but operationally unsafe. In production the draft is reviewed, tested in a lab or on a non-production device, and applied through controlled automation — never pasted blindly into a live device.
Where AI/ML fits the automation stack
AI/ML does not replace the controller and the APIs from earlier in this week — it sits on top of them. Data flows up, decisions flow down:
Telemetry flows up to a data store, the ML model turns it into a prediction, and the controller converts the decision into policy applied through APIs. The outcome is measured and fed back — the loop that keeps a baseline current. This is why AI/ML arrives after the topics earlier in the week: you need telemetry, a controller and APIs before a model has anything to learn from or any safe way to act.
AI vs ML, generative vs predictive
| Term | What it is | Think of it as |
|---|---|---|
| Artificial intelligence | The broad goal of machine intelligence | The umbrella term |
| Machine learning | A subset of AI: learn patterns from data | Training a model |
| Predictive AI | Forecasts, classifies or detects from data | “Something is about to go wrong” |
| Generative AI | Creates new text, code or configuration | “Here is a draft and an explanation” |
| Baseline | A learned picture of normal behaviour | What the model compares against |
| Hallucination | Confident but wrong generative output | Why drafts need review |
Review
Check what you have learned with these review questions.
| Key term | Meaning |
|---|---|
| Artificial intelligence (AI) | The broad goal of machines performing tasks that usually need human intelligence. |
| Machine learning (ML) | A subset of AI in which a model learns patterns from data instead of explicit rules. |
| Training data | Historical examples used to adjust a model. |
| Model | The trained function that maps inputs (features) to an output. |
| Inference | Running live data through a trained model to get a prediction. |
| Baseline | A learned picture of normal behaviour used as the reference for anomalies. |
| Anomaly detection | Flagging behaviour that deviates from the baseline. |
| Predictive AI | AI that forecasts, classifies or detects from data. |
| Generative AI | AI that creates new content such as text, code or configuration. |
| Large language model (LLM) | A generative model trained on large amounts of text and code. |
| Hallucination | Confidently stated but incorrect generative output. |
| Alert fatigue | The condition where so many alarms fire that real failures are lost in the noise. |
- Cisco. CCNA Exam v1.1 (200-301) Exam Topics, Objective 6.4 — AI and machine learning in network operations (2024). learningnetwork.cisco.com.
- Cisco. What Is Artificial Intelligence (AI) in Networking? cisco.com.
- Cisco. AI Network Analytics & Cisco AIOps (intent-based networking). cisco.com.
- ISO/IEC. ISO/IEC 22989:2022 Information technology — Artificial intelligence concepts and terminology. iso.org.
- Russell, S. & Norvig, P. Artificial Intelligence: A Modern Approach, 4th ed. — AI vs machine learning.
Sources & references
- Cisco. CCNA Exam v1.1 (200-301) Exam Topics, Objective 6.4 — AI and machine learning in network operations (2024). learningnetwork.cisco.com.
- Cisco. What Is Artificial Intelligence (AI) in Networking? cisco.com.
- Cisco. AI Network Analytics & Cisco AIOps (intent-based networking). cisco.com.
- ISO/IEC. ISO/IEC 22989:2022 Information technology — Artificial intelligence concepts and terminology. iso.org.
- Russell, S. & Norvig, P. Artificial Intelligence: A Modern Approach, 4th ed. — AI vs machine learning.
Target exams: CCNA 200-301 v1.1 · CyberOps Associate 200-201. Last content verification: 2026-10-11.