How to pivot from AI predictions to true AI decision-making

Artificial intelligence is all around us, and the different types serve their own purposes for marketers. Generative AI, for example, is a reasonably accessible technology to test and use at most businesses to generate text and visuals. The same goes for predictive modeling, which can supply marketers with critical information, such as a customer’s propensity to churn. However, to reach the marketing zenith of one-to-one personalization, marketers must harness the power of another type of AI—AI decisioning. Supervised vs. reinforcement learning I often see marketers attempt to use predictive modeling to make decisions. That’s not a mistake, exactly, but it’s an older, clunkier way of doing business. Let’s unpack why this happens and the differences inherent in predictive models versus decisioning agents. Predictive models are built with a type of AI called “supervised learning.” These models are trained on “labeled” data—for example, a set of customers might have the labels “churned” and “did not churn.” The model then tries to predict the correct label to give new customers. Marketers use predictive models for tasks like churn prediction (How likely is this customer to churn in the next 90 days?) or category affinity (Which category is this customer most likely to purchase from next?). However, when marketers try to turn these predictions into decisions, they create a lot of unnecessary work and potential inaccuracies. This process is sometimes called “next best action” and is complicated, lengthy and complex to maintain. It requires marketers to divide people into segments, conduct manual A/B or multivariate tests to determine what’s best for each segment and then encode that information as rules. When you reduce people to segments this way, you end up with a group with similar churn scores but with different reasons for churning—maybe one moved to a different city, and another had a bad customer service experience. Yet, because everyone is in the same segment, they get marketed to in the same way even though they shouldn’t be. AI decisioning, on the other hand, is based on advances in reinforcement learning, a branch of AI that has been rapidly advancing in recent years. AI decisioning agents make decisions to achieve optimal results. These agents mimic humans’ trial-and-error learning process to achieve their goals, repeating the actions that work towards the goal and avoiding the ones that detract from it. AI decisioning is proliferating—and will likely be table stakes within the next five years—because it’s radically simpler. You don’t need segments, manual A/B tests or rules. You automate all of that work because instead of having models that make predictions, you have AI agents that make decisions. You tell the agent the goal you need to meet, the actions you’re allowed to take and then you let it loose and it starts learning. The AI agent gets more intelligent based on everything it knows about an individual customer and can make the best 1:1 choice for each customer to maximize the goal—not within months, but within days. As Veronica Moturi, senior VP of customer experience at Brinks Home, said, “By using OfferFit, we achieved a 200% improvement in the profit of our contract renewal offers in just weeks. If we were trying to get there through A/B testing, it would have probably taken us another 18 months.” And that, my friends, is how you achieve data-driven, personalized customer journeys—not predictions masquerading as decisions. How to get started with AI decisioning My first tip for getting started with AI decisioning is that an “out-of-the-box” solution likely won’t meet your needs, but building in-house is unlikely to succeed. First, no two businesses are the same; enterprises need a high level of customization regarding the KPI they want to maximize with AI decisioning. If you’re a restaurant, you want to personalize based on restaurant-specific information, such as what food people order. If you’re a bank, you want to personalize things based on financial information, such as where somebody spends money, which categories they spend money on or which banking services they use. Using an out-of-the-box product won’t work. Instead, you must customize the models and use cases to your goals and business operations. Second, because reinforcement learning is notorious within the machine learning research community for being difficult, it requires highly specialized expertise. For this reason, enterprises that work with OfferFit have a dedicated team, including a machine learning implementation engineer who’s an expert in configuring these products, an engagement manager and a customer success director, both during setup and throughout their entire use of our product. Once you have an eye toward the right solution and services to get you up and running with AI decisioning, I recommend you: Brands don’t need to stop relying on predictive models, which are great for things like churn prediction. But they do need to stop using those models for decisioning. AI decisioning can complement predictive modeling—and other types of AI—to form a robust toolkit marketers can leverage to learn all they can about their customers and take the best 1:1 action from there to drive results. About OfferFit OfferFit’s AI decisioning platform makes daily individual-level recommendations for message, product, incentive, message, channel, timing and frequency of communicating to maximize any success metric you choose. The AI learns from every customer interaction, and applies these insights to the next day’s recommendations.

How to pivot from AI predictions to true AI decision-making

Featured Stories