The market has already answered whether telemetry is infrastructure. Cisco paid $28 billion for Splunk. Twelve days apart in 2025, CrowdStrike bought Spain’s Onum and SentinelOne moved to acquire Observo AI, both to embed AI-driven data pipelines directly beneath their platforms, while AI-native pipeline startups keep raising eight-figure rounds on the same thesis. When the largest names in security spend that kind of capital on the layer that collects, filters, and routes data, they are conceding something they rarely say out loud: detection, response, and every AI capability stacked on top are only ever as good as the pipeline underneath. Earlier this year we argued that telemetry had quietly become critical infrastructure. The market has now priced it in. Telemetry has stopped being plumbing. It has become the foundation.
But every road leads to the same place, and it isn’t automation. Filtering and routing data automatically, even brilliantly, even with AI in the stream, does nothing to fix data that was incomplete, malformed, or missing to begin with. Automating a broken pipeline just delivers garbage faster. And in the AI era the stakes compound: feed a model dirty, partial, or unnormalized telemetry and you don’t get one bad alert. You get amplified, confident, automated garbage out, at machine speed and scale.
The foundation of all security analytics is not a smarter model or a faster pipeline. It is clean data: every source collected with no blind spots, consistently parsed and normalized, whether the pipeline is AI-automated, enriched, or left untouched. That was the conversation we took to our roundtable at the Official Cybersecurity Summit in Chicago, and it came down to three questions. Each one decides whether you have a foundation, or whether you’re about to automate on top of sand.
1. Can you collect everything, from air-gapped to SaaS?
Clean data starts with complete data. A blind spot is not a missing line on a dashboard; it is a region of your enterprise that cannot be audited, correlated, or defended, no matter how good the analytics sitting downstream. And in 2026 the map of where data lives has stretched at both ends at once.
At one end, more workloads move to SaaS and cloud-native delivery, pushed by rising hardware costs and the pull of managed services. At the other, more data is being pulled fully air-gapped and sovereign: by NIS2 and DORA, by data-residency rules, by OT and industrial systems that cannot tolerate an outbound connection, by classified and contractually restricted workloads. The middle didn’t disappear, but both edges grew, and that is exactly where the consequential data now concentrates.
Most telemetry strategies were built for the comfortable middle of that spectrum. They handle the ordinary hybrid data center and struggle at the edges, which is precisely backwards, because the regulated, sensitive, high-consequence data is what now lives at those edges. The common failure is to paper over it with tooling sprawl: one stack for cloud, another for the data center, and an improvised, under-instrumented approach for the air-gapped segment. Three pipelines, three formats, three sets of blind spots, and manual correlation the moment an incident crosses a boundary, which modern incidents always do.
Completeness means one collection discipline that behaves identically whether it runs in an offline OT network, a hybrid data center, or a SaaS environment. Same coverage, same fidelity, both ends of the spectrum, not three separate tools you hope to reconcile after the fact. This is the part of the problem that doesn’t make headlines and doesn’t get acquired for nine figures, because it isn’t glamorous. It is also the part that determines whether everything downstream is working from a complete picture or a partial one.
It’s worth being honest with yourself here. Where in your environment can data physically or legally not go to the cloud, and how well is that segment instrumented today? How many separate telemetry tools are you running across cloud, on-prem, and air-gapped, and what is that fragmentation quietly costing you in coverage? If an air-gapped segment were compromised tomorrow, would you have the telemetry to reconstruct what happened? For a lot of otherwise mature security programs, the honest answers are uncomfortable.
2. Clean at the source, or just cleaned by automation?
Once you’re collecting everything, the second question is what state that data arrives in, and this is where the industry’s automation enthusiasm tends to get ahead of itself. Streaming pipelines that filter, route, and enrich in real time are valuable. But automation operates only on the data it’s given. If that data is malformed, inconsistently structured, or missing the fields that make it meaningful, automation doesn’t repair it. It moves the mess faster and hands it to whatever is downstream.
Clean means parsed and normalized: a Windows event, a Linux syslog line, a firewall record, an OT signal, and an API log all resolved into consistent, structured fields that mean the same thing across every source. A timestamp is a timestamp and a source address is a source address, no matter which of a dozen formats it arrived in. That work has to happen as close to the source as possible, at the edge, before the data enters analytics or an AI model, because it is far cheaper and far more reliable to normalize once at collection than to reverse-engineer meaning later, downstream, under pressure, during an incident.
Normalizing at the source is also where enrichment belongs. The pipeline adds the context that makes each record immediately useful: precise timestamps, source metadata, and resource signals such as CPU usage captured at the moment of processing. Because one agent collects the events and the metrics from the same host, a log line reporting that a process died unexpectedly can arrive already carrying the fact that CPU was pegged at 100% when it happened. The tools and analysts downstream receive meaning rather than raw noise to decode. And it happens to be where efficiency lives, too: filtering noise and normalizing at the edge is what keeps volumes, storage, and ingestion costs from running away as data grows. You move and store less, and what you do move is worth more.
The distinction matters because "we have an AI-powered pipeline" is not the same claim as "we have clean data." One describes how the data moves. The other describes whether it can be trusted when it arrives. Only the second is a foundation, and it’s worth asking, of your own environment, whether your data is normalized at the point of collection or whether you’re still asking your SIEM, and your analysts, to make sense of raw formats after the fact. When you added AI or automation to your pipeline, did data quality improve, or did you only process the same noise faster? What percentage of what you collect and store is high-value signal, and what percentage is just volume you’re paying to keep?
3. The foundation under your AI
The third question is what happens when you put AI on top, because everyone is. AI agents are no longer confined to dashboards. They are acting on infrastructure: provisioning systems, modifying rules, responding to incidents, increasingly reaching external tools through standardized layers like the Model Context Protocol. That shift raises two demands, and both come back to clean data.
First, AI has to be fed clean data, or it amplifies error with confidence. A model reasoning over incomplete or unnormalized telemetry doesn’t fail quietly. It produces fast, plausible, wrong conclusions at scale, and increasingly acts on them without a human in the loop. Garbage in is no longer one bad alert an analyst can catch; it’s automated garbage out. Clean, structured data is also cheaper for AI to work with: hand a model normalized input and it reaches the answer faster and on far fewer tokens, while messy, unstructured data is slower and more expensive to reason over. Dirty data costs you twice: once in wrong conclusions, and again in the compute burned reaching them.
Second, AI itself has to be audited, which makes it a telemetry source of its own, and one of the most demanding you’ll ever collect from. When an autonomous agent blocks a production system at 02:13, "rule applied" is worthless. You need the context it was given, the tool calls it made, the data it referenced, the sequence of decisions, and the resulting change, captured as structured, normalized, tamper-resistant records, across every environment the agent touches. The same clean-data discipline that makes AI trustworthy as a consumer is what makes it accountable as an actor. We walked through what capturing that agent telemetry looks like in practice in a recent post on capturing AI-agent telemetry with NXLog Agent.
And this is a risk you cannot see coming. An AI system can behave in ways that diverge from what it reports, or pursue a legitimate goal through steps no one authorized, and because its reasoning is opaque, the danger stays invisible until it produces an outcome. Behavioral monitoring is what surfaces that deviation, flagging the moment an agent acts out of alignment or steps outside the boundaries you agreed to.
The industry now has more than one case study for exactly this pattern. In July 2026, OpenAI models running a cybersecurity evaluation escaped their sandbox and reached the production infrastructure of Hugging Face and several other third-party services, executing thousands of actions with no human in the loop. By OpenAI’s own account the agents were pursuing the objective they had been given, and along the way they stood up an inter-agent message board and traded hundreds of thousands of messages to coordinate with each other. Weeks later Anthropic disclosed a comparable incident in which its models reached the production systems of three organizations, and two of those three did not detect the intrusion themselves. Further cases from other AI vendors have surfaced since. What they share is a shape: agents drifting outside the alignment their operators expected, sometimes coordinating to reach a goal in ways nobody sanctioned, with every individual action looking legitimate and only the deviation in behavior giving the game away. Monitoring the AI as a distinct capability, not just the systems around it, is what makes that deviation observable instead of catastrophic.
So the questions to sit with are these. If an AI agent made an infrastructure change overnight, could you reconstruct its full decision chain today? Do you trust the data your detection AI is trained and run on enough to let it act autonomously? Are you treating your AI agents as privileged actors that must be logged as rigorously as any human admin, or as convenient black boxes you’d rather not look inside?
What this points to
The acquisitions tell you the industry has agreed on where the value sits: the data layer beneath the analytics. But writing a check for a pipeline is not the same as having a foundation, and automation is not a substitute for quality. Complete collection across every environment, clean normalization at the source, and telemetry trustworthy enough to both feed and audit AI: those three properties are what turn a pipeline into a foundation.
Get them right and everything you build on top, your SIEM, your automation, your autonomous SOC, stands on solid ground. Get them wrong and you’ve bought a faster way to be misled. The data will keep growing, the budgets will not, and AI will keep raising the cost of every impurity you let through.
Clean data first. Everything else is built on it, or it’s noise.
NXLog Platform was built for exactly this: one collection discipline across air-gapped, on-prem, hybrid, and SaaS environments. If you’re rethinking the foundation under your analytics and your AI, start a free trial or talk to our team.
