Propensity models and RFM segments answer 'is this customer likely to respond' or 'which bucket does this customer fall into.' Next-best-action (NBA) systems go a step further: for a given customer, at a given moment, they decide which specific action — among a defined set of candidate actions (an offer, a piece of content, a channel, a message, or sometimes 'no action') — is predicted to produce the best outcome right now. This reframes marketing decisioning from segment-level campaign targeting into per-individual, context-aware decisioning, evaluated continuously rather than at a fixed campaign-planning cadence.

The architecture: candidate generation, scoring, and arbitration

A working NBA system has three stages. Candidate generation narrows the full universe of possible actions down to those actually eligible for this customer right now (business rules: don't offer a discount to someone who already redeemed one this month, don't recommend an out-of-stock product). Scoring predicts the expected value of each remaining candidate for this specific customer — often using uplift modeling (predicted incremental effect of the action, not just predicted baseline propensity) rather than a plain propensity score, since the goal is the *marginal* benefit of taking the action, not just whether the customer is generally likely to convert regardless. Arbitration picks the winning action, applying business constraints (frequency capping across channels, budget limits, legal/compliance rules) on top of the raw score.

next-best-action-arbitration.ts

typescript

interface CandidateAction {
  actionId: string;
  type: "offer" | "content" | "channel_message";
  predictedUpliftScore: number; // expected incremental value, not raw propensity
  costToExecute: number;
}

interface ArbitrationConstraints {
  maxActionsPerWeek: number;
  actionsAlreadyTakenThisWeek: number;
  excludedActionIds: Set<string>; // e.g. compliance or recent-redemption exclusions
}

function selectNextBestAction(
  candidates: CandidateAction[],
  constraints: ArbitrationConstraints
): CandidateAction | null {
  if (constraints.actionsAlreadyTakenThisWeek >= constraints.maxActionsPerWeek) {
    return null; // frequency cap reached - no action this cycle
  }

  const eligible = candidates.filter(
    (c) => !constraints.excludedActionIds.has(c.actionId)
  );

  if (eligible.length === 0) return null;

  // Rank by uplift net of cost, not raw predicted propensity.
  return eligible.reduce((best, current) =>
    current.predictedUpliftScore - current.costToExecute >
    best.predictedUpliftScore - best.costToExecute
      ? current
      : best
  );
}

Uplift modeling, not propensity, is the correct scoring foundation

Scoring candidates by raw propensity-to-convert overweights customers who were going to convert anyway regardless of the action — sending a discount to someone who'd have purchased at full price is pure margin loss, not incremental value created. Uplift modeling explicitly estimates the *difference* in outcome probability between taking the action and not taking it (typically via a randomized treatment/control holdout used as training data), which is the metric that actually matters for an NBA system deciding whether an action is worth taking at all.

Propensity scoring vs. uplift modeling for action selection

Dimension	Propensity scoring	Uplift modeling
What it estimates	P(convert), regardless of whether action was taken	P(convert \| action) - P(convert \| no action)
Risk if misused for targeting	Targets people who'd convert anyway, wasting incentive spend	Correctly targets people the action actually moves
Training data requirement	Historical conversion outcomes	Requires a randomized treatment/control group to isolate causal effect
Typical use	Lead prioritization, general ranking	Deciding whether an incentive/action is worth the cost for this individual

NBA systems need hard governance rails, not just good models

A purely model-driven NBA system with no business rule layer will happily recommend sending a discount offer to a customer three times in one week because each individual scoring pass looks optimal in isolation — frequency capping, channel-fatigue rules, and compliance exclusions have to be enforced as hard constraints in arbitration, not left to the model to learn implicitly, because the model has no inherent concept of cumulative annoyance or regulatory limits.

What's Next?

A single next-best-action decision is useful in isolation, but the real operational leverage comes from wiring many such decisions together into a persistent, continuously running decisioning engine that governs an entire customer relationship over time.

I build these systems professionally.

Whether it's a RAG pipeline, analytics migration, or AI workflow — let's talk.

Work With Me →

Need production AI or MarTech expertise? I build custom LLM workflows, server-side GTM tagging, and analytics infrastructures professionally.

Need custom AI or MarTech setup? Let's build together.

Let's talk →