Operations | Monitoring | ITSM | DevOps | Cloud

AI observability closes the gap between “something’s wrong” and “here’s what to fix.” If you run AI in production, you might have felt the whiplash. Yesterday, your LLM answered in 300 milliseconds (ms). Today p99 crawls, costs spike, and nobody’s sure if the culprit is model behavior, data freshness, or GPUs stuck at the ceiling. Dashboards light up, but they don’t tell you which issue puts customers at risk. That’s the gap AI observability closes.

AI workloads break every assumption you have about infrastructure management. AI is everywhere. Machine learning-based tools are answering customer service questions, accelerating incident resolution, catching fraudulent transactions, spotting defects on production lines, and powering late-night searches that delve into the random topic that pops into your head right before bedtime. Behind every prediction, response, or generated sentence is massive computing power doing serious, continuous work.

Monitoring AI systems isn’t business as usual. Monitoring AI isn’t like monitoring traditional systems. You can’t just track uptime or response times and call it a day. AI models evolve, data shifts, and behavior drifts over time, which means your monitoring has to evolve, too. If you’re running AI workloads in production, you already know this. Your models might look healthy according to your infrastructure metrics, but they’re still making bad predictions.

The rise of autonomous AI attacks operating at machine speed demands that network security evolve beyond human capacity and manual processes. Kentik AI Advisor counters this threat by using AI for good, reasoning across full network context to proactively eliminate vulnerabilities and guide immediate, confident defense.

As AI agents move from copilots to autonomous systems, they’re generating and consuming data at unprecedented scale. The result is a new kind of infrastructure pressure — one that’s quietly reshaping how organizations think about data, cost, and control. Across IT, Security, and Observability, leaders are realizing a hard truth: too much data is too costly.

Non-deterministic UIs are the future and will replace your dashboards, but they’re not here yet. So until then, we’re stuck with conversational interfaces. In an effort to try and describe what I consider the future of UIs to look like, I wrote about how you (and I) have been designing dashboards wrong. The core insight was that we've been designing for static representations of data that sit on a TV in the office, when the actual use case is someone at a desk using them to debug an issue.

Cursor is fantastic at cranking out code changes. I recently used it to splice a brand-new downstream API call into one of our Go microservices, and the diff looked great. The unit tests finished before I lifted my coffee mug, yet I still had zero certainty the change would survive contact with real traffic. That gap is all about integration tests, so I paired Cursor with proxymock and the outerspace-go demo service to prove the behavior end to end.