Playbook
The Next Best Action Playbook
Most of what B2C brands call personalization is a segment with a nicer name. The campaign still goes out on Tuesday, to everyone in the “lapsing high-value” group, with the same 20% off. The only thing that changed since 2015 is the size of the group.
A next best action program asks a different question. Instead of “which campaign goes to this segment,” it asks “which single action is right for this customer, right now, and would it change what they do?” That last clause is the one most programs skip, and it’s why they spend most of their budget on customers who would have bought, stayed or returned anyway.
This playbook is a ladder. Segments tell you who customers are. Propensity tells you what they’ll do. Uplift tells you whether your action changes it. Next best action picks the move, including the move of doing nothing. Four rungs, each with its own chapter, and each one you climb makes the budget work harder.
The chapters
- 1. Customer segmentation: from RFM to predictive segments.Demographic, behavioral and RFM segmentation, where each breaks, and what a predictive segment is.The four classic types of customer segmentationBehavioral segmentationRFM analysis, workedWhere RFM breaksPredictive segmentsFrom segments to decisions
- 2. Propensity modeling: who will act, and when.Framing the question, the data, the window, and turning scores into lists.Framing the question: who, doing what, by whenWhat the model learns fromThe three propensities most B2C brands runFrom score to list: capacity, deciles and calibrationPropensity is necessary and insufficientKeeping it accurateTwo propensity models with unusual deadlines
- 3. Next best action in practice: B2C examples by industry.Retail, subscription, gaming and banking decisions, step by step.Retail and ecommerce: the shopper who browsed twice and bought nothingSubscription and streaming: the subscriber whose hours collapsedGaming: the player stuck on level 34Banking and fintech: the paycheck that movedInsurance: seventy days before renewalWhat the five examples have in common
- 4. Uplift modeling: find the customers your offer actually moves.The four groups, how to build the model from a holdout, and the budget it saves.The four groupsWhy propensity targeting wastes budget: a worked renewalWhat an uplift model needs: a campaign with a holdoutHow uplift models are built, in plain wordsReading an uplift model: where to stop spendingUplift by industry: the actions it changes mostUplift and incrementality are the same idea at two scales
- 5. Measuring decisioning impact.Global holdouts, per-action lift, and tuning the arbitration.The global holdoutPer-action liftIncremental revenue per contact, and the contact budgetTuning the arbitrationThe dashboard that mattersA worked quarter
What is next best action?
Next best action is a decisioning approach that evaluates every possible action for an individual customer (an offer, a message, a service call, a product recommendation, or no contact at all) and selects the one most likely to improve that customer’s long-term value. It replaces segment-level campaigns with one decision per customer, made whenever the customer’s data changes.
The idea came out of banking and telecom, where the stakes per customer were high enough to justify real-time decisioning in the early 2010s. What’s new is that the same approach is now practical for an ecommerce brand with three million shoppers or a mobile game with ten million players, because the models that power it no longer need a data science team to build.
In practice, a next best action program has three parts: predictions about each customer (propensity to buy, to churn, to respond), an estimate of how much each action would change those outcomes (uplift), and an arbitration rule that picks between competing actions and respects limits like contact frequency and promo budget. The rest of this page takes those in order.
The decisioning ladder: segments, propensity, uplift, action
Segmentation is the first rung, and it’s where every brand already stands. Segments group customers by who they are and what they’ve done: new vs returning, high vs low spend, the classic RFM grid of recency, frequency and monetary value. They’re excellent for understanding a customer base and useless for deciding what to do with one customer tomorrow, because a segment is a description of the past. The chapter on customer segmentation, from RFM to predictive segments shows where RFM breaks and what replaces it.
Propensity comes next: a score for each customer that says how likely they are to do a specific thing inside a specific window. Buy from the fall collection in the next 30 days. Cancel in the next 60. Open a savings account this quarter. Propensity turns a segment of 400,000 into a ranked list where the top 20,000 matter most. The chapter on propensity modeling covers how the question gets framed and how the scores get used.
Uplift is the third rung, and the one that separates programs that make money from programs that make noise. Uplift estimates how much your action would change a customer’s behavior. A customer with 90% propensity to renew doesn’t need a discount; they’re renewing. A customer at 40% who jumps to 65% when offered a pause option is where the money is. The chapter on uplift modeling explains how to find those customers, and the ones your offer would push away.
At the top of the ladder sits the action itself: an arbitration layer that looks at every eligible action for a customer, weighs the predicted value of each (propensity times uplift times margin, roughly), applies the constraints (one message a week, this promo budget, that compliance rule), and picks one. Or picks none. Our chapter on next best action in practice across B2C industries walks through what the decision looks like in retail, subscription, gaming and banking, and the measuring decisioning impact chapter covers how to prove the ladder paid off.
Propensity models: scoring who is likely to act
A propensity model answers one question with a deadline: how likely is this customer to do X by date Y? The deadline is what makes it useful. “Likely to buy again” is a personality trait. “Likely to buy again in the next 30 days” is a campaign.
The model learns from your historical data: every customer who did X in a past window, every customer who didn’t, and everything you knew about them at the start of that window. Order history, browse behavior, session frequency, support contacts, payment events, past campaign responses. It finds the patterns that preceded the behavior and applies them to every customer today. If the mechanics are unfamiliar, our explainer on predictive modeling covers how these models get trained and validated.
Three propensities cover most of B2C: propensity to buy (overall, or per category or product), propensity to churn, and propensity to respond to a given kind of contact. A mature program runs all three on every customer and refreshes them daily. Most teams start with one, usually the one attached to the biggest line in the budget.
The output is a score between 0 and 1 for every customer, plus the drivers behind it. Rank by score, draw a line where your team’s capacity runs out, and you have a target list. That’s already a large improvement over a segment. It’s also where most programs stop, and stopping here is the expensive mistake.
Uplift modeling: finding the customers your action actually moves
Take 100,000 subscribers with high propensity to renew. Send all of them a 20% off renewal offer. Most renew, the campaign reports a 91% renewal rate, everyone celebrates. Then someone asks what the renewal rate was for the subscribers who didn’t get the offer, and it was 88%.
The offer moved three points. It paid 20% to all 100,000. That’s the propensity trap: targeting by likelihood to act means you spend the most on the customers who needed you the least.
Uplift modeling fixes it by predicting the difference an action makes to each customer’s behavior. Every customer lands in one of four groups. Sure things act whether or not you contact them. Lost causes don’t act either way. Sleeping dogs act less if you contact them (the subscriber who forgot they were paying until your renewal email reminded them). Persuadables act only because you reached out. Only the last group is worth a dollar of budget.
Building an uplift model takes one thing most teams already have and don’t use: a past campaign with a holdout, so the model can learn which customers responded differently with and without the action. If you’ve never run one, the principle is the same one behind incrementality measurement, applied at the individual level. Start a holdout on your next campaign and you’ll have training data in a cycle.
The payoff is a different-shaped budget. Instead of 100,000 offers that move three points, you send 25,000 offers to persuadables that move twelve, skip the sure things entirely, and leave the sleeping dogs asleep. Same retention. A quarter of the offers, and about a sixth of the discount spend.
Next best offer vs next best action
The terms get swapped, and the swap hides a real difference in what the program is allowed to decide.
| Next best offer | Next best action | |
|---|---|---|
| Chooses among | Products, promotions, discounts | Any action: an offer, a message, a service call, a nudge, a wait, or nothing |
| Optimizes for | Conversion or revenue from this contact | The customer’s long-term value |
| “Do nothing” | Rarely an option | Always on the menu |
| Typical owner | Merchandising, promotions, recommendations | CRM, lifecycle, decisioning |
| Origin | Retail recommendation engines | Banking and telecom decisioning |
| Example | Which coupon to show at checkout | Whether to show a coupon at all, resolve the failed payment first, or wait a week |
Strictly speaking, next best offer is a subset of next best action. It’s a good place to start if your biggest decision is what to recommend. It becomes a problem when “best offer” is the only tool, because then every decision ends in a promotion, and the sleeping dogs get woken on schedule.
Next best action examples in B2C: retail, subscription, gaming, banking
In retail and ecommerce, the decision is usually what to send a shopper who browsed but didn’t buy. A shopper with high propensity to buy from that category in the next two weeks and low uplift gets content: the restock alert, the styling guide, nothing with a discount attached. A shopper with middling propensity and high uplift gets the coupon. A shopper predicted to return half of what they order gets nothing promotional at all, because the margin on that order was never there.
In subscription, the decision triggers when a subscriber’s churn propensity jumps. The menu is a pause offer, a plan downgrade, a content recommendation that matches what they used to watch or cook, a call about the failed payment, or silence. Uplift sorts the menu: price-driven risk responds to the pause, engagement-driven risk responds to the recommendation, and the subscriber who forgot they’re paying responds to a reminder by cancelling, so they get nothing.
Gaming compresses the same decision into days. A player predicted to lapse before day seven gets one of: a bonus level, a social invite from a friend who’s still playing, an easier next stage, or a small amount of in-game currency. Uplift matters doubly here, because handing currency to a player who was about to buy it is the fastest way to lose a payer.
Banking and fintech, where next best action grew up, own the classic decision: what to do when a customer’s direct deposit is predicted to move elsewhere. A rate bump on savings, a fee waiver, a call from a banker, a credit line review, or nothing. The same arbitration handles the happier decisions too: which customers get the card upgrade offer, and which ones would have applied anyway.
Insurance runs the same decision on an annual clock. Sixty days before renewal, a policyholder predicted to shop around gets one of: a proactive rate review, a bundling offer, a call from an agent, or nothing, because for a large share of policyholders a renewal reminder is the thing that prompts the shopping. The uplift estimate decides, and the global holdout keeps score.
What every one of these has in common is the shape of the decision. One customer, a short menu of actions, a prediction of what they’ll do, an estimate of what each action would change, a few constraints, and one choice. Industry changes the menu. It never changes the shape.
How to launch a next best action program without a data science team
Start with one decision, and make it the one with the most money attached. For most B2C brands that’s “which customers get a retention offer,” because it’s a decision you’re already making badly, with a segment and a blanket discount.
Build the two predictions that decision needs: churn propensity for everyone, and uplift from your last retention campaign (or from the one you’ll run next, with a holdout). Write the arbitration rule in plain language: “offer the pause to persuadables with price-driven risk, offer the recommendation to persuadables with engagement-driven risk, contact nobody else.” Keep a global holdout of 10% who get no action at all. Run it for one cycle and compare.
The models are the part that used to require a data science team. That’s where we come in, in one sentence: you bring your historical customer data and the decision you want to make, the Predictive AI Agent reads that data as it is, trains and validates the propensity and uplift models, and we own the outcome, with scores and drivers live in the CRM, ESP or decisioning tool your team already uses. When the first decision is paying for itself, the second one (usually who gets the cross-sell, or who gets the win-back offer) runs on the same foundation.
Predict the right move for every customer:
Book a demo
FAQ
What is next best action in marketing?
Next best action in marketing is deciding, for each individual customer, which single action (an offer, a message, a service touch, or no contact) is most likely to improve their long-term value, based on predictions of what they’ll do and how much the action would change it. It replaces segment-level campaigns with one decision per customer.
What is the difference between next best offer and next best action?
Next best offer chooses among products and promotions to maximize the value of one contact. Next best action chooses among all possible actions, including service touches and doing nothing, to maximize the customer’s long-term value. Next best offer is a subset of next best action.
What is a propensity model?
A propensity model predicts how likely an individual customer is to take a specific action within a specific time window, such as buying in the next 30 days or cancelling in the next 60. It’s trained on your historical customer data and outputs a score for every customer, plus the drivers behind it.
How is uplift modeling different from propensity modeling?
Propensity modeling predicts whether a customer will act. Uplift modeling predicts whether your action changes what they do. A customer can have high propensity and zero uplift (they’d act anyway) or moderate propensity and high uplift (your offer tips them). Uplift needs a past campaign with a holdout to learn from; propensity doesn’t.
