Is AI a Tool or an Operating System? New Relic CMO Katrina Wong Makes the Case

Introduction

Katrina Wong, Chief Marketing Officer at New Relic, has spent more than 20 years building go-to-market engines at enterprise technology companies — including Twilio Segment, Zuora, Salesforce, and SAP — with a particular focus on AI-driven marketing and reaching developer audiences. Asked to settle a question a lot of marketing leaders are quietly arguing about internally, she didn't hedge: AI is an operating system, not a tool. That distinction sounds semantic until you realize it changes how you budget, how you staff, and what you expect your stack to do a year from now.

Why the Category Question Matters

Wong's reasoning is about trajectory. AI is "powering more and more, and maybe one day everything we do," she said — describing it as "this intelligence layer that's just going to be super foundational, as like the premise of where we start." That last phrase is the operative one. Tools sit downstream of a decision; you evaluate them, slot them into a function, and swap them out when something better appears. An operating system is the starting premise — everything else gets designed on top of it.

The Practical Test: Do Your Agents Talk to Each Other?

Benjamin offered a diagnostic from his own operation that marketing leaders can run against their own stack this week. He described ru ing a chief of staff agent and a separate podcast builder agent that handles client onboarding. As long as those agents operate in isolation, he argued, you don't have an operating system — you have customized workflows with better autocomplete. The threshold gets crossed when the agents start talking to each other. That's a more useful maturity marker than counting how many AI features your vendors shipped, because it measures integration rather than adoption.

The Data Layer Is Still the Constraint

Wong's immediate follow-up is the part worth pi ing to the wall. Agreeing on the operating system framing, she pushed back on any reading that treats AI as self-sufficient: "the data layer is still super important. So having data that's clean — AI is only as good as your data. That is still super important. It's not just AI by itself."

What That Means for Your Stack

Benjamin framed AI as the interface that lets you extract and visualize your data layer, which sharpens the point. An interface inherits the quality of what sits beneath it. If your customer data is fragmented across systems that don't reconcile, an intelligence layer doesn't resolve that — it operationalizes it and scales the error. The uncomfortable implication for marketing leaders is that data hygiene, instrumentation, and governance aren't prerequisites you get to defer until after the AI pilot proves out. They are the AI investment.

Turning Data Into Decisions

There's a reason this framing comes from an observability company's CMO. New Relic's core discipline is collecting real-time data so engineering teams can monitor, troubleshoot, and optimize what they've built — a central nervous system for software. Marketing organizations rarely hold themselves to that standard for their own operations. If AI really is the operating system, the marketing equivalent of observability — knowing what your data actually says, in real time, with enough fidelity to act on it — stops being a nice-to-have and becomes the condition for everything ru ing on top of it.

Key Takeaways

Three things to carry into your next pla ing conversation. First, evaluate AI as infrastructure rather than as a line item, because tool-by-tool procurement produces exactly the disco ected workflows that never compound. Second, use agent interoperability as your maturity test — isolated agents are automation, co ected agents are an operating system. Third, treat data quality as the gating factor it is: an intelligence layer built on unreliable inputs produces confident, well-formatted, wrong answers at scale. As Benjamin closes each week, the advice is to just focus on keeping your customers happy — which is considerably easier when your data can actually tell you whether they are.

--- One thing worth flagging: the transcript supplied covers only the final ~90 seconds of the episode — the closing rapid-fire question, Katrina's answer, the data-layer exchange, and the sign-off. I wrote strictly from that material rather than inventing strategies she didn't discuss, which is why the piece centers on the OS/tool framing and the data-quality constraint instead of covering 3–5 distinct tactics. It lands at roughly 640 words. If you can send the full transcript, I can expand it with the substantive middle of the conversation on turning data into decisions.