The AI hype in networking is deafening. Every vendor claims their product uses "AI-powered" something. Most of it is marketing wrapped around basic pattern matching. But underneath the noise, there are tools that genuinely change how network engineers work. We have been using them in production for over a year. Here is what actually delivers results.
Phase 1: Code Completion (GitHub Copilot)
This is where most engineers start, and it is genuinely useful. GitHub Copilot sits inside your editor and suggests code as you type. For network automation specifically, it excels at a few things:
- **Boilerplate generation:** Writing Netmiko connection handlers, NAPALM getters, or Nornir inventory plugins involves a lot of repetitive code. Copilot autocompletes most of it after you write the first few lines.
- **Jinja2 templates:** If you start a Jinja2 template for a Cisco IOS config, Copilot often suggests the next several lines accurately because it has seen thousands of similar templates in its training data.
- **Test scaffolding:** Writing pytest fixtures for network automation tests is tedious. Copilot handles the structure so you can focus on the assertions.
**Limitations:** Copilot does not understand your specific network. It suggests generic code based on patterns. It will not know that your organization uses OSPF area 100 for underlay or that your naming convention is different from the Cisco default. You still need to review everything it generates.
Phase 2: Conversational AI (ChatGPT, Claude)
The next step up is using AI as a conversation partner for design decisions and troubleshooting. This is where things get interesting for network engineers who are not full-time developers.
Real example: you need to parse the output of "show bgp ipv4 unicast summary" across 200 routers and flag any neighbor that has been in Active state for more than 24 hours. You describe this in plain English, and Claude generates a complete Python script with TextFSM parsing, concurrent execution, and a summary report. The script works on the first try about 70% of the time, and needs minor corrections the other 30%.
This is not replacing the engineer. It is eliminating the hours spent reading library documentation and Stack Overflow answers. The engineer still needs to understand what the script should do, review the output, and validate it against real devices. But the time from "I need a script" to "I have a working script" drops from hours to minutes.
Phase 3: Agentic AI (Claude Code)
This is the jump that changed everything for us. Claude Code is not a chatbot — it is an agent that can read your codebase, run commands, edit files, and execute multi-step tasks autonomously. For network automation, this means:
**Building entire automation projects from a description.** You tell Claude Code "build a Flask API that accepts a webhook from PagerDuty, parses the alert, SSHs into the affected device, collects show commands, and posts the output back to the PagerDuty incident." It reads your existing code style, creates the project structure, writes the code, and tests it. You review and approve each step.
**Refactoring legacy scripts.** We had a collection of Expect scripts from 2018 that were still running in production. Claude Code read through all of them, understood the logic, and rewrote them as modern Python with Netmiko, proper error handling, logging, and unit tests. What would have taken an engineer two weeks took an afternoon of supervised AI work.
**MCP servers for live network interaction.** This is the cutting edge. MCP (Model Context Protocol) servers let Claude Code interact directly with network management systems. We built an MCP server for Cisco NSO that lets the AI query device configurations, check sync states, and even push config changes through NSO's transaction manager — all with human approval at every step.
What Does Not Work Yet
Honesty matters, so here is what AI still struggles with in networking:
- **Real-time troubleshooting under pressure.** When a network is down and you are on a bridge call, you do not have time to prompt an AI and review its suggestions. Muscle memory and deep protocol knowledge still win in break-fix scenarios.
- **Complex multi-vendor designs.** AI can help with individual device configs, but designing an end-to-end network that spans Cisco, Juniper, Arista, and Palo Alto with consistent policy enforcement still requires a human architect.
- **Anything involving proprietary or undocumented features.** AI models are trained on public data. If your vendor has a feature that is not well-documented online, the AI will hallucinate confidently wrong answers.
Where to Start
If you are a network engineer who has not touched AI tools yet, start with Phase 1. Get Copilot, use it for a month, and see how it changes your scripting workflow. Once you are comfortable, move to Phase 2 — use Claude or ChatGPT to help you design and build automation projects. When you are ready for the real power, move to Phase 3 with Claude Code and start building agentic workflows.
The progression is not optional. Engineers who skip AI tools entirely will find themselves competing against engineers who use them — and losing. Not because the AI is smarter, but because the AI-augmented engineer ships ten times faster.
Originally published at https://primeautomationsolutions.com
Top comments (0)