AI Adoption in Networking

I recently joined Greg Bryan to talk about the real state of AIOps in network management. What's working, what's hype, and where engineers should actually focus their energy.

The short version? Despite all the noise, adoption is still low. Surveys suggest only about 15% of organizations have deployed AIOps tools, and for good reason. Engineers don't trust output they can't verify. LLMs are inherently indeterminate and they can hallucinate confidently, sending you chasing ghosts through your logs. Layer in years of unformatted legacy data that most organizations are sitting on and you've got a real problem before you even get started.

That said, the tools are getting genuinely useful in specific areas. Root cause analysis is the sweet spot right now. Instead of spending an hour combing through logs, you can ask "why can't Sally connect to the Wi-Fi?" and get an immediate, actionable answer about password failures or signal strength. Morning summaries that flag overnight circuit flaps and anomalies are another practical win. The kind of thing that makes your first cup of coffee more productive.

For teams looking to get started, my advice is simple: use what you already have. Vendors like Juniper Mist and HPE Aruba Central have been baking AI capabilities into their platforms for years. If you want to go further and integrate your own data with LLMs, look into the Model Context Protocol (MCP). It lets LLMs query your databases via API calls without permanently ingesting your data.

But here's the non-negotiable: start with a zero-trust, read-only mindset. AI agents should be able to look but not touch. No write access, no configuration changes, no deletions. Not yet.

The bigger picture here is that the Tier 1 support noise and rote analysis can be automated so we can focus on the complex stuff that actually requires judgment. An AI can flag a maxed-out circuit, but it doesn't know the office is closed for a holiday. That contextual, intent-driven decision-making is still ours. The engineers who learn to wield these tools effectively are the ones who'll thrive.

A closer look at MCP

MCP is Anthropic's protocol for letting an LLM reach data outside itself. Think of an MCP server as a translator: on one side it speaks the model's language, on the other it makes REST calls or SQL queries against your systems, and it tells the model what it is capable of so the model can ask for what it needs. There are hundreds if not thousands of them now, many written with AI itself; I have built a couple. Network to Code's Nautobot has integrations, and Itential announced Flow AI at Autocon, a marketplace of agents for their automation platform. RAG and RAFT are heavier lifts. Try MCP first.

Selling it internally, and not breaking anything

Ask Scott Robohn's question first: what problem are you trying to solve? Then prefer buy over build where you can, so there is someone to call when it misbehaves. On security, MCP's early versions were weak and the ecosystem is now catching up with single sign-on, proper identity, permissions and accounting. My advice for right now is blunt: read-only access, one hundred percent of the time, and only to the subset of data the task needs. There are enough stories of an LLM with file system access wiping a directory. Think about that at database scale. And test the load, because a chatty model can quietly denial-of-service your own API endpoint in the middle of the day.

Check out the full conversation below.