Predictive marketing: how AI anticipates your customer’s next move

Most brands I see are still running their CRM on numbers pulled out of thin air. A win-back flow that fires at 30 days of inactivity. A replenishment email that lands 60 days after the first order. When I ask where the number comes from, the answer is always the same. An internal average, a benchmark someone read somewhere, or something a guy said once in a meeting.

These static rules treat everyone the same and never adjust on their own. Predictive marketing fixes that, and almost nobody in eCommerce is actually using it.

What predictive marketing is, without the bullshit

Predictive marketing, in plain terms, is AI that predicts what a customer is going to do next, then triggers messages off that. When they'll reorder, how much they'll spend across their entire lifetime, and the odds they're about to walk.

Classic segmentation groups people on what they've already done: bought recently, opened an email, spent over $100. It only looks backward. Predictive estimates what each customer is going to do, based on their own patterns.

What people forget is that you already have this data. Every purchase, every browsing session, every click in an email is a signal. It's sitting in your CRM, you're just not using it.

Predictive marketing vs classic marketing, what actually changes

I'm starting here because it's what I get asked every single time.

Classic marketing reacts to the past. It looks at what happened and applies a rule. Predictive anticipates. It looks at what the customer is about to do, and that's exactly what shifts your CRM timing.

Here's an example. Classic marketing makes you say "this customer spent over $200, so I'll drop them into my VIP segment." Predictive makes you say "this customer is going to spend $800 over the next 12 months, so I treat them as a VIP now, before they've proven anything." Same data to start with, but you steer differently.

Why static rules don't hold up

Take two customers on a skincare site. A pro makeup artist who rebuys her products every week. And a regular customer who restocks twice a year. If you send them both the same reminder at 60 days, it's mistimed for both. Too late for one, way too early for the other.

Static rules lump everyone together, and since they don't correct themselves, the tuning falls back on you. You have to look back every month, test, adjust the window, start over. A predictive model does this continuously, recalibrating on each customer's behaviour without anyone touching the flow.

With a static number, you say, I haven't seen this person in 30 days, I need to re-engage them. Based on what? A market average. With predictive, you re-engage at the exact moment that specific customer is about to rebuy (or leave). Your timing finally matches real behaviour.

What data you need to make it work

This is what stops most brands before they even start. Good news, you don't need to build out a data science team or plug in a data lake.

The data that feeds the models, you already have it. Each customer's order history, purchase frequency, amount spent, recency of their last order, and email engagement. That's exactly what your CRM collects continuously.

The problem isn't raw quantity. What you really need is volume and a clean base. The model needs enough customers and enough history to produce a reliable prediction instead of falling back on an average. And it needs a clean base, proper segmentation and data that isn't full of duplicates or dead profiles. I'll come back to the exact thresholds right below.

The predictive metrics worth knowing

Klaviyo builds machine learning models directly into the platform. Here are the ones I use most with our clients.

Predicted LTV combines historical LTV with estimated future spend. You spot your best customers earlier, before they've even proven their value. Expected next order date is a per-customer forecast based on their own cadence, not a store average, and you can even customise the windows to match your purchase cycle. Churn risk gives you the probability a customer is slipping through your fingers, modeled on their own frequency. Next best product and cross-sell date tell you what they'll buy next and when they'll be receptive, based on real purchase sequences across your catalogue. And RFM sorts your customers (Loyal, At Risk, Needs Attention) by purchase frequency, recency and amount spent, and more importantly it tracks how they move from one category to another over time.

One important point, next best product and RFM require Marketing Analytics or the Advanced Data Platform. And when your account doesn't have enough data, some predictions fall back on averages. For it to run well, you need at least 500 customers who have ordered, 180 days of order history with orders in the last 30 days, and a handful of customers with 3 or more orders. Below that, the model doesn't have enough to work with to be reliable.

Klaviyo's edge here is that these predictions behave like any other profile property. You use them directly in your segments and flows. I've worked on other platforms before, and that's not the norm at all. Elsewhere, acting on predictive data usually means extra tools and a tighter set of options.

The tactics that actually pay off

Here's what I set up, roughly in the order I prioritise it.

Replenishment flows tuned to each customer

First thing I build as soon as a brand sells consumables. Instead of five flows with five different timelines for five products, you build a single flow triggered by the expected next order date. Klaviyo already knows which product to rebuy and when.

The logic is simple. If a customer runs out of your product and you're not in their inbox at that moment, someone else is. A well-timed replenishment email keeps you top of mind right when they're ready to check out again.

Oral care brand Smile Brilliant uses predicted next order date for its replenishment flows. Combined with other personalisation efforts, that produced a 39x ROI on the platform in one year.

Cross-sells triggered by predicted receptivity

My favourite combination, next best product plus cross-sell date. Rather than blasting a generic promo to your whole list, you send each customer the product they're most likely to buy next, at the moment they'll be most receptive to it.

The result, conversion rate goes up, because you're hitting the right offer at the right time.

RFM segmentation to stop treating everyone the same

RFM is a priority for most of our clients. You segment your base by value and purchase frequency, and you adapt your message to each group. A loyal customer doesn't need the same thing as an at-risk one.

Win-backs triggered by RFM movement

The approach I prefer: trigger the win-back when a customer changes RFM category, instead of a plain day counter. When a customer shifts from "Loyal" to "At Risk," you catch them at that exact moment.

So you re-engage on a real behavioural signal, not a day counter that doesn't care what the customer is actually doing.

Prioritising acquisition budgets with predicted LTV

Predicted LTV isn't just for CRM. You can push your best potential customers up to Meta and Google to feed your lookalike audiences with the right input. You stop optimising for conversion volume and start targeting customers who'll actually be worth something in 12 months.

How to know if your model is actually working

A detail people often forget, a predictive model can drift. Buying behaviour shifts, your base evolves, and a prediction that was accurate six months ago can end up way off.

So you don't set it on autopilot and forget it. You check whether the predictions hold up against reality. Do the customers the model flags "at risk" actually leave? Do the next order dates land at the right time? If the gap between the prediction and what actually happens keeps widening, your model is short on fresh data or your base has gotten dirty. You go back to data quality before touching the flows.

Where to actually start

If I had to give you the running order, first you check you have the data volume for the models to be reliable (the 500 customers and 180 days of history). Then you build your replenishment flow on next order date, that's the most obvious quick win if you sell consumables. Then you set up your RFM segmentation and tie it to real actions.

Because that's where most brands stall. They look at their cohorts, they go "huh, interesting," and that's it. Predictive is worthless if you don't turn the insight into concrete CRM actions. A model that tells you a customer is going to churn, if you don't trigger anything off it, is just a nice dashboard.

You already have all this data in your CRM. The only thing that matters is wiring actions onto it. Otherwise it just sits there asleep in your CRM.