From the Analyst Chair to the Inside: Nick Heudecker on GenAI, Data Infrastructure, and What the Market Gets Wrong.
Q1. You spent nearly eight years as a VP Analyst at Gartner covering data and analytics — advising hundreds of enterprises on their technology strategies from the outside.
You are now on the inside, in a hypergrowth company navigating one of the most turbulent and opportunity-rich moments in the data industry’s history. What has surprised you most about the difference between those two vantage points — and what did you think you understood about enterprise technology decisions from the analyst side that turned out to be significantly more complicated once you were operating inside a company trying to execute against them?
The biggest surprise is how much analyst work optimizes for being right in general, while operating inside a company optimizes for being right on time. As an analyst, I could tell a client “you should consolidate your observability tooling” and be correct on a five-year horizon even if it took them three years to act. Inside Cribl, being directionally right but early, or being right but slow, is functionally the same as being wrong, because a competitor or a budget cycle will close the window.
What I underestimated was the gravitational pull of sunk cost and internal politics on “rational” technology decisions. From the Gartner side, you see the RFP, the bake-off, the evaluation criteria and it looks like a logic problem. What you don’t see as clearly from outside is that the person championing a new tool inside the enterprise is spending political capital, not just budget. They’re betting their credibility on an outcome, and that changes the math in ways that have nothing to do with feature comparisons. While I understood this intellectually in my analyst days; I underestimated how dominant a factor it actually is until I was the one trying to get a customer’s champion to go to bat for us internally.
The other thing that’s more complicated in practice than in theory: as an analyst, “the market” is a set of vendors and use cases you can categorize cleanly. Inside a hypergrowth company, you’re not just executing against a market map, you’re trying to change the map itself in real time while customers, competitors, and your own product are all moving simultaneously. Advising on strategy from the outside assumes a reasonably stable target. Building the strategy from the inside means the target is often a function of decisions you yourself are making.
Q2. GenAI is simultaneously the most significant new workload hitting enterprise data infrastructure and the most significant disruptor of the business models of the companies that sell data infrastructure. From your current position at Cribl and your years watching this market, where do you see GenAI creating genuinely new and durable business value for enterprises — and where do you see it being adopted in ways that are adding cost, complexity, and risk without producing the outcomes organizations are promising their boards?
The durable value is showing up wherever GenAI is applied to a narrow, well-bounded problem with good data behind it. It falls flat when applied as a general-purpose replacement for judgment. Code generation and developer productivity is the clearest case. The task is well-specified, the feedback loop is fast, and the cost of a wrong answer is usually a failed test, not a business incident. Similarly, in my world of observability, security, and IT operations, using LLMs to triage alerts, summarize incidents, or draft the first pass of a runbook is durable value because you’re compressing a search-and-summarize task that a human would do anyway, and a human is still checking the output before anything consequential happens.
Where I see cost, complexity, and risk piling up without matching outcomes is in what I’d call ambient AI, or deploying GenAI broadly across a workflow because leadership wants an AI story for the board, rather than because there’s a specific, measurable job to be done. That shows up as a wave of pilots that never graduate to production, agent frameworks bolted onto processes that were never instrumented well enough to trust the output, and a real explosion in the volume and variety of machine-generated data that nobody budgeted to store, route, or secure.
Enterprises are discovering that every agent, every retrieval call, every model interaction generates telemetry and logs, and most data infrastructure was sized for a pre-agentic world. That’s a genuine architectural problem. The cost of ingesting and storing that volume of data in legacy observability and SIEM platforms will price out a lot of AI initiatives if enterprises don’t get smarter about what data actually needs to go where.
Q3. One of the defining characteristics of the current GenAI moment is that it is arriving everywhere simultaneously — in security, in observability, in analytics, in application development — while the underlying data pipelines that feed these systems are still largely the same fragmented, inconsistent, and governance-challenged infrastructure that existed before. From where you sit, what is the most consequential data infrastructure decision enterprises are getting wrong right now as they try to build AI-ready architectures — and what does getting it right actually require that most organizations are not yet doing?
The most consequential mistake is treating the data pipeline as plumbing rather than as the control point it actually is. Enterprises are racing to plug GenAI and agentic systems into their existing data estate as if the pipeline’s only job is to move bytes from A to B. But when you connect a model to your data, you’re no longer just moving data, you’re deciding what that model gets to see, in what shape, with what context, and with what guardrails. Most organizations are making that decision by default rather than by design, because their pipelines were built for a world of dashboards and static reports, not for feeding autonomous or semi-autonomous systems that will act on what they’re given.
The practical consequence is that a lot of “AI-ready architecture” is really just old, fragmented pipelines with a vector database and an LLM stapled onto the end. The governance problems that existed before, like inconsistent schemas, duplicated and stale data, PII scattered across systems nobody fully inventoried, don’t go away when GenAI enters the picture. They get amplified, because now those inconsistencies aren’t just producing a wrong number on a report, they’re feeding decisions an agent might act on autonomously, at machine speed, without a human in the loop to notice the number looks off.
What getting it right actually requires is treating data-in-motion as a governance layer, not just a transport layer. That means having the ability to inspect, transform, mask, route, and enrich data before it lands anywhere, so that sensitive data is redacted at the source rather than after it’s already been indexed into a dozen downstream systems; so that only the right data reaches the right model or agent for the right purpose; and so that there’s an audit trail of what data went where and why. Most enterprises are not doing this today because it requires investing in the unglamorous middle of the stack rather than the flashier bits, like the data sources and the AI application. Everyone wants to talk about their model, their agent, or whatever. Almost nobody wants to talk about the plumbing. But the plumbing is where trust, cost control, and compliance actually get decided, and it’s the layer most organizations are still under-investing in relative to how much weight they’re now asking it to bear.
Another aspect most enterprises are overlooking is pipeline ownership. Getting value from your data means sending it to multiple downstream destinations. If your pipeline is locked into a given vendor, your options suddenly narrow and you may not be able to capitalize on a best of breed approach that’s essential as technology continues to evolve quickly.
Q4. You have a rare combination of experiences — deep analytical knowledge of how enterprise technology markets form and evolve from your Gartner years, and now direct operational responsibility for category creation and go-to-market strategy at a hypergrowth company. Category creation is one of the hardest things in enterprise technology — most attempts fail or collapse back into existing categories. What does GenAI do to the dynamics of category creation in enterprise data right now — does it make it easier because everything is in motion, or harder because every vendor is claiming an AI story and the signal-to-noise ratio has collapsed?
It sounds like a contradiction, but it’s really both. GenAI makes category creation easier to attempt but harder to win. It’s easier to attempt because customer budgets and attention are genuinely unlocked right now. Buyers who would have taken eighteen months to evaluate a new category are willing to have the conversation today, because nobody wants to be the CIO or CISO who missed the shift. That’s an opportunity for anyone with a legitimately new architecture to propose. Cribl’s own experience creating the observability pipeline category is a great example. Category creation works when you’re naming something the market already feels but doesn’t yet have language for. Right now, GenAI has put a lot of enterprises in exactly that position with their data infrastructure. They feel the strain but they just don’t have a name for the gap yet.
It’s harder to win because, as you say, the signal-to-noise ratio has collapsed. Every vendor slide now has an AI-shaped logo on it, and buyers have gotten fast at discounting “AI-powered” claims that aren’t backed by an actual architectural difference. You can’t create a category on vocabulary alone anymore. You have to create it on a demonstrable, defensible difference in outcome. A few years ago, being early and loud with an AI narrative was itself a differentiator. Today it’s table stakes, and the differentiation has to come from something a customer can verify.
The other dynamic I’d point to is that GenAI is collapsing the shelf life of a category faster than before. In the old model, you might get several years of relatively stable category definition before the market matured and consolidated. Now, the underlying technology is shifting quarter to quarter, so the category you define at launch may need re-definition twelve months in because the ground is moving underneath you. That means category creation today is less a single campaign and more a continuous act of re-explaining yourself as the technology and the buyer’s understanding of it both evolve. That’s a much harder discipline to execute than the classic “define it once, defend it forever” playbook analysts used to describe.
Q5. Looking at the enterprise data and observability landscape over the next three to five years as GenAI matures and agentic AI systems begin generating their own data, logs, and telemetry at machine scale — what do you think is the most underappreciated infrastructure challenge that enterprises are not yet planning for, and what decisions should data and security leaders be making today that most are not yet making?
The underappreciated challenge is that agentic AI is about to make machine-generated data a much bigger and much stranger category than anything security and observability teams have dealt with before. We’ve spent two decades building infrastructure around data generated by humans clicking, applications logging, and infrastructure emitting metrics, and all of it was relatively predictable in volume and shape. Agents don’t behave that way. An agent can spin up sub-agents, retry failed reasoning steps, call the same tool a dozen times debugging itself, and generate logs, traces, and intermediate outputs that have no human-scale analog. Volume isn’t the only issue. The shape and semantics of the data are new too, and most SIEM and observability platforms, and the budgets behind them, were sized for the old shape. Enterprises are planning for AI adoption; very few are planning for the fact that every agent they deploy is also a new, prolific data source that needs to be ingested, retained, secured, and made sense of.
The second piece, closely related, is that agentic systems create a new and much harder security surface: not just “is our data protected,” but “can we reconstruct exactly what an autonomous system decided and why, after the fact.” When an agent takes an action, the audit trail needs to capture not just the action but the reasoning chain and the data the agent had access to at that moment. Most security teams are instrumented to answer “who did what,” a question built for human actors with stable identities. They are not yet instrumented to answer “which agent, running which prompt, with which retrieved context, took this action.” That gap is going to matter enormously the first time an agent does something wrong at scale and a regulator or a board asks for the reconstruction. And if that weren’t bad enough, all of these challenges compound when talking about multi-agent systems expected to share context, goals, and processes.
What data and security leaders should be doing today, and mostly aren’t: treating agent telemetry as a first-class data category now, before volume forces the issue, by building the routing and filtering layer that decides what agent-generated data is retained in full, what’s sampled, and what’s summarized before it hits expensive storage — because the naive approach of “log everything” will not survive contact with agent-scale volume economically. They should also be establishing an identity and provenance standard for non-human actors now, while the number of agents in production is still small enough to retrofit, rather than waiting until there are thousands of agents with inconsistent or nonexistent identity trails. The organizations that will handle this well are the ones treating agentic AI’s data exhaust as a known, sizable infrastructure requirement to architect for today — not as a problem to solve once it’s already caused an incident.
Qx. Anything else you wish to add?
Just one thing I’d add. For all the disruption GenAI is causing, the fundamentals I learned covering this market as an analyst haven’t changed. They’ve just gotten more expensive to ignore. Data quality, governance, and knowing what you actually have and where it lives were always the unglamorous prerequisites for good decision-making. The only difference now is that the systems making decisions on top of that data can act autonomously and at machine speed, so the cost of skipping those fundamentals shows up faster and bigger than it used to. My advice to enterprises hasn’t really changed direction from my Gartner days; it’s just gone from “you should really get to this” to “you no longer have the luxury of not getting to this.” I’d rather see the industry spend the next few years getting genuinely excited about that unglamorous work than chasing the next model headline.
This part has been one of the most interesting aspects of working with IT and security teams instead of traditional data analytics folks. IT and security teams don’t think of themselves as data managers, so the larger conversation around data strategy, governance, and quality are generally new to them, but they also see the value and are eager to learn.
…………………………………………………………..
Nick Heudecker leads market strategy and competitive intelligence at Cribl. Prior to joining Cribl, he spent over seven years as a VP industry analyst at Gartner, covering the data and analytics market. With over twenty years of experience, he has led engineering and product teams across multiple successful startups in the media and advertising industries.and advertising industries.