AI for Optimizing Customer Lifetime Value

Summary

Most marketing analytics answers the wrong question.

That's the argument Tobi Konitzer, VP of AI at GrowthLoop, made in DataCamp's webinar on causal AI and customer lifetime value. Teams build models that predict who will churn, convert, or expand, then act on the prediction. Konitzer says that's backwards: the question that matters is what will change a customer's behavior, not who is likely to behave a certain way. A model that says viewers of a product page are worth more money describes a pattern already in the data. It says nothing about which action actually moves that number. That gap between correlation and causation, Konitzer argues, is where marketing strategies leak value they don't need to lose.

He traces the case from behavioral economics, bias and noise in human decisions, through the appeal of machine decisioning, auditable, repeatable, less biased, to a specific warning: AI agents left unsupervised reason over correlational data and make confident, plausible, wrong calls. His answer is architectural. Keep the actual decision, which treatment to send which customer, inside an explainable system built on reinforcement learning. Push agentic AI to the edges, where it can generate ideas and explain past decisions, but never make the causal call itself.

Key Takeaways

• Predictive models can identify who is likely to convert or churn, but they can't say which action caused that outcome, so building a strategy on the prediction alone risks scaling the wrong behavior.
• Causal inference has a hard limit: no business can observe both what happens with an intervention and what would have happened without it for the same customer, so causal claims need actual experiments, not just historical records.
• Konitzer points to two well-documented flaws in human decision-making, bias and noise, and argues neither can be fully corrected by auditing yourself, which is part of why he expects machine decisioning to keep replacing human judgment.
• Letting an AI agent act directly on a correlation, such as promoting a product because past high-value customers happened to browse it, risks repeating that mistake across an entire customer base.
• GrowthLoop's approach keeps treatment allocation inside an explainable reinforcement-learning system built on Bayesian bandits, and limits agentic AI to the edges of that system: generating candidate treatments, explaining past decisions, and reasoning over results.
• Lifetime value moves slowly and resists real-time experimentation, so Konitzer proposes tracking a faster-moving stand-in metric, closer to real-time purchase intent, built by combining sales data with event-stream data.
• Customer-facing teams often optimize for the next renewal cycle in ways that can work against a customer's value over time, a tension Konitzer admits has no clean fix.
• The practical starting point for any team is building a habit of running controlled experiments with real control and treatment groups, since an AI agent cannot infer cause and effect from historical data alone.

Deep Dives

Prediction Isn't the Same as Causation

Rhys Phillips opened the session by naming the problem directly: "Most marketing analytics answers the wrong question." Teams build models to predict who is likely to churn, convert, or expand, then act on that score. Konitzer's argument is that the score describes a pattern, not a cause, and treating the two as interchangeable is where marketing programs quietly lose money.

He illustrated the point with a statistic that pairs a country's McDonald's locations with its World Cup wins: France, with roughly 1,500 restaurants, beat Morocco, with 53. The pattern holds across tournaments, and no sophisticated audience mistakes it for cause and effect. But Konitzer pushed the example further: proving the relationship is purely correlational would require playing the same match twice, once with every McDonald's in France present and once with none, at the same time. That's impossible, and it's exactly the obstacle, known as the fundamental problem of causal inference, that shows up constantly in marketing data.

He applied the same logic to a familiar retail scenario: a brand notices that customers who frequently view a certain product tend to have higher lifetime value, so an agent promotes that product more aggressively to everyone. The view and the value may simply share a third cause, like income or intent, with no causal link between them at all. Conversion-rate targeting has the same flaw. A model that flags a 90 percent conversion probability doesn't tell a marketer whether that targeting changed the outcome. As Konitzer put it, the real question is "would those folks with high conversion rate have been more likely to convert with the targeting that we did based on that score?" Comparing targeted customers to a random sample answers a different, easier question, and it's the one most teams settle for.

The Case Against Human Judgment, and Its Limits in Machines

Before defending machine decisioning, Konitzer built the case against the human version. Drawing on behavioral economics, particularly Daniel Kahneman's work, he named two separate pathologies. Bias is the familiar one: documented patterns like racially skewed sentencing or disproportionate traffic stops. Noise is subtler and, in his view, more consequential: two interviewers evaluating identical candidates reach different verdicts depending on how their morning went. The outcome varies not because the evidence changed but because of what Konitzer called "an idiosyncratic pattern of divergence:" noise.

Machines don't inherit that particular flaw. "Machines are auditable," Konitzer said. "Whereas humans are not." A manager can't make an employee explain, days later, why one candidate got hired over an equally qualified one; the real reason is often inaccessible even to the person who made the call. A classifier, by contrast, can point to the exact inputs and weights behind a decision. Run it on the same data twice and it gives the same answer both times.

Konitzer was careful not to let that framing slide into blanket trust. Machines only avoid bias if the data feeding them represents an unbiased world, and the auditability advantage collapses once a model has billions of parameters that no person can trace by hand. His broader point: "I do not think that you can audit yourself to a world where you make better decisions. I don't think it's possible. And I think behavioral economists have proven that at nauseam." Machine decisioning still beats human decisioning on average, in his assessment, but only when someone keeps the specific failure modes of each in view rather than assuming the machine version is automatically safe.

An Onion, Not a Black Box: Where Agentic AI Should and Shouldn't Sit

Konitzer's model for a causal decisioning system looks like an onion. At the center sits the actual decision, whether to send Mary the ad, whether to email Bob today, built on reinforcement learning and Bayesian bandits: math that produces a number a person can check. Agentic AI never touches that center. It works the periphery instead: proposing candidate treatments, initializing the system with prior knowledge instead of random guesses, and explaining decisions after they're made.

That last function is what makes the architecture useful to a non-technical marketer. A reinforcement-learning system produces tables of probabilities and rewards that mean nothing to most people. But an agent can translate them on demand. Konitzer described a marketer asking why an email went to Bob, and the agent answering from the underlying data: "based on the probability distributions, Bob was two times as likely to have a good result, and his result, causally, was 1.5x compared to the standard." The explanation is generated by an agent, but it's grounded in a number the explainable core actually computed, not a guess dressed up as reasoning.

The same separation solves a cold-start problem. Reinforcement-learning systems typically explore at random when they have no history to learn from. Konitzer's platform instead seeds that exploration with results from real experiments: "we can use our prior causal knowledge that we painfully built through experimentation suites." In a live demo, he showed an outcomes-optimization flow where an agent proposed a set of treatments, a channel and content combination, based on that causal evidence, then handed the actual per-customer decision to the bandit system at the center. Agentic AI generated the options; it didn't choose among them.

Rethinking Lifetime Value as a Moving Target

When Phillips asked what data matters most for lifetime value models, Konitzer redirected the question toward a structural problem with LTV itself: "the problem with LTV in itself is that it's a very static, sticky thing. It's very hard to move this thing. It doesn't move much." Recency-frequency-monetary frameworks attempt to model it, but the underlying number changes too slowly to support the kind of real-time experimentation and decisioning Konitzer had spent the session describing.

His proposed fix is a faster-moving stand-in. He described it with a hypothetical: "if I had a magic wand and I could write a machine learning model that essentially shows my purchase intent or my dollar sign, proverbial dollar sign that I carry over my head," tracing a line that rises and falls by the second. That signal, not the static LTV score, is what he thinks decisioning systems should actually optimize against, because it moves fast enough to respond to a treatment and be measured.

Getting there means combining data sources that usually sit apart. As Konitzer put it, "combining sales data with event stream data maybe into a machine learning model" produces what his team calls real-time LTV: a stand-in outcome close enough to true lifetime value to anchor decisioning, but responsive enough to actually experiment against. The distinction matters beyond marketing. Any team trying to optimize a slow, sticky outcome, patient health, loan default, customer churn, faces the same problem: the metric that matters most is often the metric that's hardest to move fast enough to learn from.

Where to Start, and the Guardrail That Matters Most

Asked where a marketing team should begin, Konitzer didn't point to a tool or a model. He pointed to a habit: "I wanna build autonomous machines that make better decisions. The first step is to build that causal collateral and build some muscle of experimentation data where I have a control group, I have a treatment group." Without real experiments generating real causal evidence, an agent has nothing to reason over except the correlational data already sitting in a warehouse, and that data can't answer the questions decisioning requires.

He was equally direct about where good intentions go wrong in practice. Customer-facing teams face constant pressure around renewal cycles, and that pressure produces short-term fixes with long-term costs: "I've seen customer facing teams essentially do everything possible for the next renewal cycle. But in some ways, this may be counterproductive for the long term, customer lifetime value." He didn't offer a clean resolution, calling it a genuine cross-pressure rather than a solved problem.

The session closed on the same warning that opened it. Machine decisioning is coming regardless, in Konitzer's view, because humans can't fully audit or fix their own decision-making flaws. But he was blunt about the condition attached to that shift: "any future of agentic decisioning without guardrails will lead to disaster." In marketing, he noted, the stakes of getting a decision wrong are comparatively low, a customer sees the wrong product recommendation. In fields like medicine or criminal sentencing, they aren't, and the same architecture, causal evidence at the center, agents at the edges, matters more the higher those stakes climb.