Next Best Action explained: what it is, how it works, how it differs from segmentation, and what a modern NBA stack looks like in 2026.
Next Best Action (NBA) is a decisioning approach that picks, for each customer at each moment, the single action with the highest expected incremental value. It replaces segment-and-blast campaigns with per-customer, per-moment routing. In 2026, most large B2C growth programs either run some version of NBA or are actively migrating toward it.
The term has been overloaded by vendors — Salesforce Einstein, Pega Customer Decision Hub and half a dozen CDPs all claim it — so it helps to strip it back to the underlying idea and rebuild from first principles.
Three things are load-bearing in that definition. Per customer means the unit of decision is one person, not a segment. Per moment means the decision is re-evaluated when state changes, not on a weekly cadence. Incremental value means the score reflects what the customer does because of the action, not what they would have done anyway.
Segmentation groups customers into buckets and applies a shared treatment to the bucket. NBA scores every eligible action for every individual customer and picks the winner per customer per moment. The distinction matters because segments systematically undershoot per-customer uplift: two customers in the same segment can have opposite treatment effects, and a segment-level treatment averages them into a mediocre outcome for both.
Segments can still exist as guardrails or eligibility filters (region, plan tier, consent state) but they should not decide what to send. That is the job of the per-customer, per-action score.
A modern NBA stack ingests three families of signals in near-real time: behavioral (product events, session recency, engagement decay), transactional (purchases, billing events, plan changes), and contextual (device, location, seasonality, upstream marketing exposure). These land in a feature store that resolves the state of every customer at any moment.
Two model families do the real work. Propensity modelsestimate the probability a customer takes an action; useful for eligibility and prioritization. Uplift models — meta-learners such as T-learner, X-learner and R-learner — estimate the causal effect of a specific intervention on a specific customer. Uplift is the model family NBA actually needs, because the whole point is to rank actions by their incremental effect, not their expected level.
The decisioning engine takes the scored actions and picks a winner per customer, subject to constraints: channel capacity, frequency caps, consent, product eligibility, holdout assignment. It emits one decision per customer per moment with a full provenance trail — the signal, the segment, the model version and the prior experiment it descends from.
The chosen action is routed to the appropriate channel (email, push, in-product, call center, paid) and the outcome is measured against a preserved holdout of customers who received no action for the same window. The read is treated-versus-holdout, per action, per cohort. Anything less than that is not causal.
Every category leader has an NBA product: Salesforce Einstein Next Best Action, Pega Customer Decision Hub, Adobe Journey Optimizer, and now most modern CDPs. The stacks differ in detail but converge on the same conceptual shape. What distinguishes them in practice is (a) whether the models are genuinely uplift-based or wrap propensity under a rebrand, (b) whether a preserved holdout is a first-class object or a bolted-on afterthought, and (c) how much of the loop is actually agentic — meaning proposing candidate actions, not just picking among a fixed set the marketer wrote in advance.
A future post in this series will do the head-to-head on the big vendors; for now, the buyer's rule of thumb is: if the product cannot show you the incremental margin per decision against a control, it is not doing what NBA is supposed to do.
If you're evaluating NBA as an operating model rather than a feature to buy, the two pieces that pair with this one are continuous decisioning (why the calendar has to go) and how to increase ARPU (which specific decisions actually move the number).
Markin runs Next Best Action as a native operating model: per customer, per moment, ranked by incremental margin, measured against a holdout. Explore Growth optimization to see it live.
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.
