How a next best action model is built: propensity, uplift meta-learners and a ranking function that scores expected incremental margin per action.
Frequently asked
Related resources
Solutions
See it in the product
The same loops this note describes run 24/7 against your customer base. Watch the workspace decide, experiment and execute 1:1.
Vocabulary
Definitions an answer engine can quote, each one a page of its own.
Compare
Neutral reads on the categories buyers evaluate against a decision layer.
All comparisonsA next best action model is the scoring function that decides, for one customer at one moment, which eligible action carries the highest expected incremental value. It is not a single algorithm. It is a stack: a propensity layer that estimates what is likely, an uplift layer that estimates what an action changes, and a ranking layer that converts both into money and picks a winner.
Most teams build the first layer, skip the second, and let the third become a spreadsheet of business rules. That is why so many programs report impressive AUC and flat revenue. This guide covers the model families that work in production, the objective function that should sit on top of them, and how to know the model is right rather than merely accurate.
The output is not a score per customer. It is a matrix: for every customer, one expected value per eligible action, including the null action of doing nothing. The decision is the argmax of that row, subject to constraints.
Writing the objective this way forces three things into the open that a propensity model hides: the candidate set has to be enumerated, every action needs a margin, and doing nothing has to be allowed to win. In a mature program the null action wins a large share of decisions, and that share is a feature, not a defect.
On tabular customer data, gradient-boosted trees (XGBoost, LightGBM, CatBoost) remain the production default. They handle mixed types and missingness, train in minutes, and reach most of their quality ceiling with little tuning. Sequence models earn their cost only where event ordering carries more signal than event counts, typically long behavioural traces in streaming and gaming.
Propensity answers who is likely to convert, churn or upgrade. It does not answer whether the action changes anything. That is a causal question, and the practical defaults are meta-learners: T-learner (separate models on treated and control), X-learner (more stable under treatment imbalance) and R-learner (Neyman-orthogonal, more robust to nuisance error). Each wraps the same boosted-tree workhorse in a causal frame, so the engineering cost of adding uplift on top of an existing propensity pipeline is modest.
The failure mode to watch for is propensity relabelled as uplift. If the top of the ranked list is dominated by customers who were going to convert anyway, the program is paying for behaviour it already had.
The ranking layer applies eligibility, consent, frequency caps, quiet hours, budget and cannibalisation rules, then takes the argmax. Constraints belong here rather than inside the model: they change weekly, they are auditable, and they should never be laundered into learned weights.
Contextual bandits allocate traffic across actions while learning, which is useful when the candidate set changes constantly and rewards arrive quickly. They are weaker when the reward horizon is long, as it is with subscription margin, and when you need a clean causal readout for finance. A common production compromise: bandits for creative and offer variants inside an action, uplift models for which action to choose, and a preserved holdout on top of both.
Weeks 1 to 4: enumerate the candidate action set and attach a margin to each, then ship a propensity baseline with a holdout. Weeks 5 to 10: add a T-learner or X-learner on the two or three highest-volume actions, and switch ranking from propensity to expected incremental margin. Weeks 11 onward: extend the candidate set, add constraints, and let the null action compete. Most of the lift arrives in the second phase, when the objective changes rather than when the algorithm does.
For how the models sit inside the wider operating loop, see the guide to next best action, and for the retention-specific version of the same stack, see customer churn prediction.
Markin builds and retrains the propensity, uplift and ranking layers against your base, and reads every decision against a preserved holdout. Explore Next best action to see it live.
