Next-best-action (NBA) is a decision policy for one customer at one moment. It chooses from the actions a business has made available, removes the ones that are not allowed or appropriate, and ranks what remains for the customer’s current context.
That is wider than choosing the next promotion. The answer might be an offer, a useful message, a service step, a delay, or no marketing contact. A churn score can inform the choice, but the score does not make the choice.
Next-best-action, in short. Start with a defined set of actions. Filter it with eligibility, suitability, context, and contact rules. Rank the remaining actions against a stated objective. Then compare the outcome with what happens when no action is taken. If a system only predicts who is likely to click or buy, it has a response model—not evidence that its chosen action changed retention.
What does next-best-action mean?
It means choosing what the business should do next for a particular customer, given the actions it can take and what it knows at that moment.
Pega’s description of next-best-action starts with a library of potential actions. It says business rules control which actions are eligible and relevant to the context, while the system selects among them. Its examples include an offer, a message, and support, and its categories extend beyond sales to service, retention, and nurture.
The useful unit is therefore not “Which campaign should this segment enter?” It is “Of the actions available now, which one should this customer receive?” A campaign can be one candidate in that decision. It is not the whole decision.
What counts as an action in retention marketing?
An action is something the business can actually deliver or deliberately withhold—a service step, a useful message, an incentive, or a decision to send nothing. For a retention team, the candidate set might include:
- explain a delayed order;
- show help for a product the customer already owns;
- send replenishment guidance;
- offer an incentive;
- route a service issue;
- wait and send nothing.
The list matters because a decision system cannot select an option that does not exist in its candidate set. If the library contains five discounts, the “best” result is still a discount. If it contains service and no-contact choices too, the decision can be about the relationship rather than only the sale.
Is next-best-action the same as predicting churn?
No. Predictive churn estimates a risk or probability. Next-best-action uses signals such as that risk to choose among possible responses.
A churn model might rank two customers in the same order regardless of whether the available response is a service call, an email, a discount, or no contact. The action policy has more work to do: decide which responses are allowed, what outcome matters, what each response costs, and which one should win now. The site’s explainer on predictive churn makes the same point from the model side: a score is not proof that acting on it retains anyone.
How does a next-best-action decision work?
A usable policy has five parts: candidate actions, filters, ranking, delivery context, and a comparison against no action.
1. Define the candidate actions and objective
Write down the choices the system is allowed to make and the outcome it should value. “Improve retention” is not yet a decision rule. The policy still needs to know whether it is choosing for a renewal, a repeat purchase, continued product use, margin after an incentive, or another stated outcome.
The objective also needs to carry the cost of the action. A response that would have happened without an incentive is not made more valuable by paying for it.
2. Remove actions that should not compete
Filtering comes before ranking. In Pega Academy’s implementation example, business rules and AI models sit in the same decision framework. Channel limits and constraints control exposure, while engagement policies determine whether an action is eligible, relevant at that moment, and suitable.
This prevents a high model score from overriding a hard rule. An action can be profitable and still be inappropriate. It can be relevant and still be suppressed because the customer has already received too many messages.
3. Rank what remains
After filtering, the policy compares the eligible actions. The implementation above calls this arbitration: it weighs response propensity, current context, action value, and business priorities.
Those inputs answer different questions. Propensity estimates the chance of a response. Context changes what is relevant now. Action value represents what the outcome is worth. Business levers, in Pega’s term, let the team apply priorities the model does not get to invent.
4. Make the decision in the current context
The same customer can need a different action when the context changes. A customer browsing a replenishable product presents one decision. The same customer asking about a failed delivery presents another. A useful policy is triggered with the current channel and situation, rather than replaying a segment label that was computed earlier.
5. Keep a no-action comparison
The final step is measurement. A team observes the outcome after the chosen action, but it cannot observe what that same customer would have done at the same time without it. That missing outcome is the counterfactual.
The practical answer is a control or holdout. BCG names randomised holdout groups or quasi-experimental designs as the training data an uplift model needs. Evaluation needs a comparison with customers who did not receive the action, using a randomised holdout or a quasi-experimental design. Without that comparison, the system can learn who responds while remaining unable to show that the action caused an incremental result.
Why is response propensity not enough?
Because a likely responder may act anyway.
BCG’s June 2026 explanation of next-best-action draws the line directly: propensity asks how likely a customer is to respond, while uplift asks whether the action changes behaviour compared with no action. A customer with high purchase intent can rank first in a response model even when a discount changes nothing except the margin.
Research on causal contextual bandits for marketing describes the same failure from another direction. Optimising observed clicks or revenue can reinforce behaviour that would have happened without the campaign. The authors therefore formulate the problem around incremental outcomes and include a control state in which the customer is held out of marketing or receives a control treatment.
This is the minimum credibility test for an NBA claim: can the system distinguish “this customer responded” from “this action changed what the customer did”?
What does uplift modelling change?
Uplift changes the target from predicted outcome to predicted difference. Instead of asking which customer is likely to renew, it asks how much a particular action is expected to change renewal relative to no action.
That comparison makes “do nothing” part of the measurement design. It can also change which customer receives an incentive. Someone with high renewal intent but no incremental response may lose to someone whose renewal probability is lower but more changeable.
Uplift does not settle the full policy. The candidate actions, eligibility rules, cost, channel, and outcome window still have to be defined. It fixes one specific gap: mistaking predicted response for predicted impact.
When does a contextual bandit enter the picture?
When the policy has several actions and needs to learn which one works in which context while it is making decisions.
In BCG’s description, a contextual bandit balances exploitation—using the action currently expected to work best—with exploration: deliberately giving some decisions to less-certain actions so the policy can learn their value. That deliberate exploration is what separates it from a fixed propensity ranking.
Exploration is not permission to ignore constraints. Eligibility, suitability, contact pressure, and channel availability still define which actions may compete. The bandit learns within that allowed set.
What would this look like for one retention decision?
Suppose a customer has a high churn score, an unresolved delivery problem, and a recent promotion, and the candidate actions are a discount, a service update, a useful message, and no contact. A recent-promotion rule removes the discount. The open delivery problem makes the service update relevant now. The policy ranks the service update, the message, and no contact against its stated objective, with the churn score as one input, and delivers the winner in the channel available at that moment. A holdout keeps the comparison with no action.
The churn score matters, but it never becomes the instruction “send a coupon.” It is one input to a decision whose other parts are explicit and inspectable.
What should a retention team define before buying or building NBA?
Ask for the policy, not the label: the actions it can choose, the rules that remove actions before scoring, the objective with its outcome window and costs, whether it ranks by response or by incremental effect, where current context enters, how exploration is bounded, and the holdout that shows the no-action outcome.
Questions to ask
- What actions can the system choose, including service and no-contact outcomes?
- Which rules remove an action before scoring?
- What exactly is being optimised, over what outcome window, and after which costs?
- Is the ranking based on response propensity or incremental effect?
- Where does current context enter the decision?
- How is exploration bounded when the system is still learning?
- What control or holdout shows what would have happened without the action?
If a vendor or internal team cannot supply those answers, “next-best-action” may mean little more than ranking offers by likely response. That can still be useful, but it is a narrower system and should be judged as one.
Does next-best-action improve retention?
Not as a general result. The sources used here establish the structure of the decision and the measurement problem, not a retention lift from next-best-action.
One vendor documents how it filters and ranks actions. The independent analysis distinguishes response, uplift, and contextual learning. The research paper formalises the missing counterfactual and reports preliminary offline experiments on a retail-fashion dataset. None provides a universal retention lift, a universal reward window, or a reason to trust an NBA label without a holdout or comparable no-action comparison.
The defensible claim is narrower: next-best-action can organise a retention decision around an explicit action set, rules, context, objective, and learning method. Whether that policy retains more customers is an experiment the business still has to run.
Frequently asked questions
Is next-best-action always AI?
Not entirely. In the implementations described here, AI models contribute predictions or learning, but they are not the whole decision. A documented implementation combines business rules with AI models: rules decide which actions may compete, while models help rank the ones that remain. The NBA label does not tell you which model is used.
Can “do nothing” be the next best action?
A team can define it as a deliberate policy outcome. At minimum, no action must exist as the comparison used to measure incremental effect. Otherwise the system observes responses without establishing what the customer would have done untreated.
Is next-best-action just personalisation?
Personalisation changes an experience for a person. Next-best-action is the policy that chooses which action should be taken from a set of alternatives. A personalised message can be one candidate; it is not the decision system by itself.
What data does next-best-action need?
It needs the data required by the policy’s own parts: signals for the current context, fields used by eligibility and suitability rules, records of recent contact, the chosen action, and the outcome used for evaluation. The necessary fields therefore follow from the action set and objective, not from the NBA label.