How GrowthLoop Wants to Make Marketing a Verifiable Science
Anthony Rotio on why experiments must become durable memory, why warehouse‑native CDPs beat stale copies, and the limits he won’t compromise to chase automation.
Topics
Anthony Rotio’s career travelogue reads like a primer on system thinking: from Harvard computer science and hands‑on robotics to steering high‑stakes marketing at AB InBev, and now co‑leading GrowthLoop, an agentic, warehouse‑native CDP built to turn marketing experiments into durable evidence. What unites those stops is a single conviction: complex systems improve only when they remember what worked, why it worked, and under which conditions it can be repeated.
GrowthLoop’s signature idea, the Causal Context Graph (CCG), is driven by the very same conviction. Rather than letting learnings die in slide decks, dashboards, or tribal memory, the CCG records four decision‑time truths: customer state, business and market context, the action taken, and the measured incremental outcome, so each experiment becomes reusable causal evidence. For Rotio, the payoff is profound: it creates a verifiable feedback loop that transforms marketing from pattern‑matching to evidence‑driven decisioning, and it supplies reward signals that can safely train AI agents to improve future choices.
In this conversation, Rotio, as Co-Founder & Co-CEO of GrowthLoop, walks through how the CCG works in practice, the thorny limits of automation he refuses to cross, and pragmatic migration strategies for CMOs stuck on legacy stacks.
His view is both hopeful and disciplined: AI can automate more of marketing, but only when decisions are rooted in causal evidence and human judgment remains the final check at the decision boundary.
In the Humans of Martech conversation, you introduced the “agent context graph” as a way to snapshot experiments and answer “what if?” before marketers hit send. What intellectual or practical gap were you trying to close with that model—and what are its limits?
The Causal Context Graph solves a common marketing problem that’s rarely recognised: learning from an experiment usually stops compounding once the campaign ends.
It sounds odd to position it that way, but it’s so often the case. A marketer runs a test, learns from the results, and then that logic is archived in a slide deck, a dashboard, or simply the shared knowledge of the marketing organisation. The learning freezes in time with that experiment, but the customer’s behaviour, the channel mix, and the competitive environment keep shifting.
At GrowthLoop, the Causal Context Graph records four pieces of info at the moment of a decision: the customer’s state, the relevant business and market context, the action taken, and the incremental outcome. Each experiment adds another piece of causal evidence to the enterprise’s data cloud. Over time, that creates a durable memory of what the company tried, for whom, under what conditions, and what actually changed.
Critically, the Causal Context Graph is built from causal evidence, not just observational patterns. Most marketing data tells you what customers did. An experiment asks a different question: what changed because we intervened, compared with what would have happened otherwise? The Causal Context Graph preserves that answer as reusable evidence about which customers respond to which treatments and under which conditions.
What excites me is the possibility of making marketing and growth a verifiable domain for AI. In coding, you can run tests. In math, you can check an answer. In science, you can test a prediction through an experiment.
Marketing needs that same feedback loop. By connecting interventions to measured incremental outcomes, the CCG provides the evidence from which we can derive reward signals for reinforcement learning. That creates a path to train AI to solve marketing and growth problems through repeated experimentation, using verified results to improve the next decision. The feedback is noisier and slower than a code test, so those rewards need to reflect uncertainty and the quality of the experiment.
Its limits are straightforward: a context graph is only as trustworthy as the underlying causal design and the data it has access to. Incomplete data, weak holdouts, or poorly chosen KPIs create bad evidence, and AI will amplify those weaknesses. Keeping the system close to governed enterprise data reduces stale copies and preserves control, but it does not rescue weak causal design.
GrowthLoop positions itself as an agentic, composable CDP and a Compound Marketing Engine that runs directly on a customer’s data cloud, rather than copying data into yet another silo. What problem in the CDP market were you most obsessed with fixing when you started building GrowthLoop?
When two of my fellow co-founders, David Joosten and Chris Sell, were working at Google, they felt the pain so many marketers feel today when launching campaigns: it took forever. It meant filing tickets with technical teams to build an audience and then waiting weeks to launch a campaign to that audience. Data was hard to access and fragmented across systems. We started GrowthLoop to give marketers direct, governed access to that data.
CDPs promised to solve this problem. But when we started GrowthLoop, every CDP on the market — even the ones marketers loved — copied customer data from the systems of record into a new proprietary store, which became the single source of truth, or “golden customer record.” But because it was a copy, that data was almost immediately stale and incomplete. And the CDP was just another system for the data team to manage.
We decided to bet on the data cloud (BigQuery, Snowflake, Databricks) because we knew the cloud would become the real system of record for customer data, and the right CDP wouldn’t compete with that. Instead, it would run natively on top of it. GrowthLoop lets marketers define audiences, orchestrate journeys, and activate data to the channels they already use while the data stays under the enterprise’s control.
That solved the access problem, but the CDP was never the destination. The larger opportunity is to close the learning loop: decide what to do, activate it, measure the incremental result, and feed that evidence into the next decision. The data cloud remains the foundation. The Causal Context Graph records what the enterprise learns, and AI decisioning uses that evidence in the next decision.
As GrowthLoop scales beyond core CDP into decisioning and agent orchestration, what will you not build—even if customers ask for it—because it would dilute the integrity of your causal AI thesis?
We will not build black-box autonomy or a system that optimises for activity instead of incremental outcomes. What might be impressive in a demo (an agent picks an audience, an offer, the send time, and reports results) could be disastrous in production.
Marketers shouldn’t have to hand over authority to an AI black box without any way to check the reasoning. The agentic system must show the evidence: what it expected, what it changed, and whether the action produced incremental value. Human control remains at the decision boundary.
You and your AI leadership team have argued that most marketing data today is correlational, not causal, which limits what AI can safely automate. In your writing on “agentic causal marketing,” you describe LLMs as “causal parrots”—excellent at dressing up correlations as causal stories without doing proper inference. How do you practically guard against that problem inside GrowthLoop’s own product roadmap?
We separate a hypothesis from a proven policy. An LLM can identify a pattern or propose a next action, but that does not make the recommendation causal. We use experiments, holdouts, and incrementality measurement to test whether the intervention changed the outcome. Until the evidence is strong enough, the recommendation stays a hypothesis for a marketer to approve and test rather than a rule the system scales automatically.
We’re also deliberate about keeping humans in the loop at the decision boundary, especially early in a new use case. After an action runs, the measured outcome goes back into the Causal Context Graph so the system can learn from what actually happened. We judge the agent by incremental business results, not by the volume of activity it produces.
“The self-deception among legacy providers is treating lock-in and switching cost as customer loyalty.”
You’ve been quite direct on LinkedIn about the difficulty of defending old CDP and marketing cloud logic in an era of open data clouds and AI decisioning. What do you think is the biggest piece of self‑deception happening inside legacy solution providers right now? Conversely, what is the biggest blind spot inside the “composable, warehouse‑native” CDP camp—including GrowthLoop—that you think will become obvious in hindsight?
The self-deception among legacy providers is treating lock-in and switching costs as customer loyalty. A lot of marketing clouds were built assuming that once a customer’s data lived inside their platform, they’d stay — not because the platform kept earning their loyalty, but because leaving was expensive and painful.
That assumption held for a long time. The data cloud changes that equation. Enterprises can keep control of their data and replace the application layer in stages, so incumbency has to be earned through product value rather than data captivity.
The blind spot in the composable camp, including at GrowthLoop, was assuming that open architecture alone would win. Open architecture solved the data-control problem, but marketers still want a complete workflow and often prefer fewer tools. CDPs and customer engagement platforms will keep converging.
As feature breadth becomes less distinctive, the durable advantage will come from the learning loop: the causal evidence that helps an enterprise improve each customer decision.
Many enterprises already have significant sunk costs in legacy marketing clouds and CDPs. What’s your most honest advice to a CMO who suspects they bought the wrong platform—but can’t rip it out tomorrow?
My honest advice is: don’t treat this as a single, all-or-nothing migration, because that’s usually how the wrong platform ends up staying in place for another three years. Identify the most expensive or restrictive point in the current workflow, then replace that layer with a use case whose value can be measured.
For many enterprises, the first target is the brittle process of copying data into a marketing cloud and keeping it in sync. That’s the part you can decouple first, without touching the execution tools your teams already know how to use.
More practically, start by making sure every new source of customer data lands in your data cloud first, not just in the martech vendor’s silo, so you’re not adding to the sunk cost while you plan your way out. Then run a warehouse-native layer alongside the existing stack for one high-value use case, such as lifecycle or loyalty. Set a holdout, measure the incremental result, and use that evidence — not a transformation roadmap — to guide the renewal decision.
You do not need to remove the entire stack to regain leverage. Start where the current architecture creates measurable cost or delay, prove a better path, and expand from evidence.
ALSO READ: Has RCS Solved Mobile Marketing’s Biggest Problem: Trust?
