How To Use CDP Data For Next Best Action Models

Blog

11/16/26

How To Use CDP Data For Next Best Action Models

Use CDP data for next best action models by feeding the decision process unified customer history, deriving features such as recency and frequency, recording which actions were shown and what happened afterward, and applying eligibility, consent, suppression, and frequency rules before anything is delivered. Start with rules and propensity, then use control groups to determine whether actions actually change behavior.

A customer is eligible for four possible actions.

You could send a replenishment reminder.

You could offer a discount.

You could remind them that they are close to the next loyalty tier.

Or you could do nothing.

Which one should happen?

That is the problem a next best action system is designed to solve.

The mistake is starting with the algorithm.

Before a model can choose among actions, the enterprise needs reliable customer history, useful features, a defined library of candidate actions, current eligibility and consent, a record of what the customer has already received, and a way to determine what happened afterward.

A customer data platform can provide much of that foundation.

But the CDP is not automatically the entire next best action system.

What Is Next Best Action?

Next best action is a decisioning approach that chooses the most appropriate next step for one customer at a particular moment from several eligible actions, including the possibility of taking no action at all.

That action might be a message, an offer, a service intervention, a product recommendation, a sales task, or a decision not to contact the customer.

Next Best Action Is Broader Than Next Best Offer

Next best offer is a narrower problem.

It asks:

Which offer should this customer receive?

Next best action asks:

What should happen next?

An offer may be one candidate.

So might onboarding guidance, a service intervention, a renewal reminder, or no outreach.

This distinction matters because the economically best action is not always a promotion.

Next Best Action Is Different From Segmentation

Segmentation groups customers who meet similar criteria.

Everyone inside the group may receive the same treatment.

Next best action evaluates a customer and chooses among several possible treatments.

Segmentation can still be part of the system, especially in early versions, but it is not the same decision problem.

Next Best Action Does Not Automatically Require AI

A first version can use deterministic rules.

If the action library is small and the rules are understandable, a rule based system can be perfectly legitimate.

Machine learning becomes more useful as the number of customers, signals, actions, and possible interactions grows.

The objective is not to use the most sophisticated model available.

It is to improve the quality of the decision.

What A CDP Supplies To A Next Best Action System

The CDP provides the customer profile, history, identity, signals, consent, and often the features used by decisioning, but other layers usually define the actions, score them, arbitrate among them, deliver them, and measure what happened.

Some CDP products include more of these capabilities than others.

Enterprise teams should evaluate the specific platform rather than assume every CDP has the same boundaries.

The Unified Customer Profile

The CDP brings together the history required to understand the customer.

That can include:

  • Purchases
  • Digital behavior
  • Product usage
  • Messaging engagement
  • Service interactions
  • Lifecycle state
  • Consent and preferences

The decision system needs a coherent view of the customer before it can choose what should happen next.

Derived Features

Raw events usually need to become decision-ready features.

Examples include:

  • Days since last purchase
  • Purchases during the previous 90 days
  • Average order value
  • Change in product usage
  • Number of recent support interactions
  • Recent category interest
  • Time until renewal

These features may be calculated inside the CDP, warehouse, feature platform, or data science environment.

The important requirement is reproducibility.

If a model was trained on one definition of “recent engagement” but production calculates it differently, decision quality can deteriorate even when the model itself has not changed.

The Action Library

The CDP typically does not invent the universe of possible actions.

The business needs an action library.

Each action should define what the action is, who may receive it, what it costs, which channels can deliver it, and what rules restrict its use.

For example, a cross sell offer may be unavailable to customers who already own the product.

A retention discount may be inappropriate for a customer who would probably renew without an incentive.

A promotional message may need to pause when an unresolved service case is open.

Models And Scores

A model may estimate the probability of:

  • Purchase
  • Renewal
  • Churn
  • Expansion
  • Response to an offer

More advanced systems may estimate how much a specific action changes the outcome.

The CDP supplies the customer data underlying those calculations.

The model or decisioning layer produces the score.

Arbitration And Business Rules

Once several actions have scores, the system still needs to choose one.

That decision may consider:

  • Eligibility
  • Expected value
  • Consent
  • Frequency limits
  • Customer fatigue
  • Channel availability
  • Action cost
  • Customer lifecycle state
  • Current service issues

The highest model score should not automatically win.

Delivery

The selected action eventually reaches another system.

That might be:

  • Email
  • Mobile
  • Website
  • CRM
  • Contact center
  • Sales workflow

The destination executes the action.

The Feedback Log

This is one of the most important and most easily omitted pieces.

The enterprise needs to record:

What action was shown, to whom, when, through which channel, and under which decision.

Without that record, the system cannot reliably learn which actions performed better.

The CDP Data A Next Best Action Model Needs

A useful NBA dataset combines customer history, current context, permission state, exposure history, and outcomes tied back to the same identity.

The last two categories are particularly important because they create the learning loop.

Purchase And Transaction History

Transaction data supports features such as:

  • Recency
  • Frequency
  • Monetary value
  • Product ownership
  • Category preference

A common failure occurs when important channels are missing.

If in-store purchases are not tied to the same customer identity as ecommerce purchases, the decision system may recommend something the customer already bought elsewhere.

Digital And Product Behavior

Recent product views, application sessions, feature usage, and similar behavioral signals help identify current intent or engagement.

The main limitation is identity.

Anonymous behavior provides little value for known-customer next best action unless it can eventually be connected reliably to the customer profile.

Messaging Engagement

Clicks, replies, unsubscribes, and other channel behavior can provide context about channel fit and fatigue.

Use some engagement metrics cautiously.

For example, email open data can be noisy because privacy features and automated image loading can affect whether an “open” represents real human engagement.

Service History

Open cases, recent complaints, and escalations can prevent the system from choosing the wrong action at the wrong time.

A customer dealing with an unresolved billing problem may not be a good candidate for an upsell message.

Service history therefore belongs inside the decision context, subject to the appropriate privacy and access restrictions.

Consent And Preferences

Consent and channel preferences should act as gates, not just model features.

The system should know whether the customer may receive the action before the action is selected or delivered.

Where permission can change quickly, the current state should be checked at the moment of action.

Action And Exposure History

The system needs to know what the customer has already been shown.

A useful exposure record includes:

  • Customer identifier
  • Action identifier
  • Decision time
  • Channel
  • Model or rule version
  • Reason code

This makes frequency control and causal analysis much easier.

Without exposure history, the model may repeatedly choose the same action without understanding that the customer already ignored it several times.

Outcome Events

The enterprise must also define what success means.

An outcome could be:

  • Purchase
  • Renewal
  • Expansion
  • Reply
  • Activation
  • Another business-specific event

The outcome needs to be tied back to the same customer identity and evaluated within a defined outcome window.

If the organization cannot connect the action to a later outcome, the learning loop is incomplete.

Build Next Best Action In Stages

Enterprises do not need to begin with agentic decisioning. A staged approach allows the data, measurement, and governance to mature before decision complexity increases.

Stable Kernel uses the following four-stage maturity ladder as a working model for planning. It is not an industry standard.

Stage 1: Rules

Rules use deterministic logic based on CDP segments, events, and eligibility.

For example:

If the customer is lapsed, has valid email consent, has no open service issue, and has not received a win back message recently, then allow the win back action.

Rules work well when the action library is small and the business needs a fast, interpretable starting point.

The limitation appears when rules multiply, overlap, and begin contradicting one another.

Stage 2: Propensity Plus Rules

A propensity model estimates how likely a customer is to produce an outcome.

The business rules still determine which actions are eligible.

The model helps rank opportunities.

This can improve prioritization when the organization has enough historical data and a clearly defined outcome.

But propensity creates an important problem.

The customer who is most likely to buy is not necessarily the customer whose behavior can be changed.

Stage 3: Uplift

Uplift modeling estimates the difference between what is likely to happen with the action and what is likely to happen without it.

This moves the question from:

Who is likely to buy?

to:

Whose probability of buying changes because we act?

That distinction becomes critical when the intervention has a cost.

BCG makes the same distinction in its 2026 next best action research, arguing that propensity models answer predicted behavior while more mature systems need to understand changed behavior.

Stage 4: Adaptive Decisioning

More advanced systems may use contextual bandits or agent-based decisioning across a larger action library.

BCG describes a broader decision stack involving propensity and uplift models, contextual bandits, and foundation-model agents.

The attraction is adaptation.

The system can learn which actions perform well in different contexts rather than relying entirely on a fixed ranking model.

The governance requirement also becomes much stronger.

As decision autonomy increases, the organization needs tighter controls around action eligibility, measurement, monitoring, and explainability.

Why Propensity Alone Is Not Enough

Propensity predicts who is likely to act; uplift attempts to estimate whose behavior changes because of the action.

Consider a hypothetical customer who already has a very high likelihood of purchasing without intervention.

A propensity model may rank that customer first.

The system sends a discount.

The customer purchases.

The dashboard credits the offer.

But the customer may have purchased anyway.

The business paid for an outcome it was already likely to receive.

That is why next best action should eventually move beyond response prediction when interventions have meaningful costs.

A useful decision system needs to know not only whether the customer is likely to respond but whether the action changes the result enough to justify:

  • The discount
  • The sales effort
  • The communication
  • The service cost
  • The opportunity cost of not choosing another action

Control groups create the data required to start answering that question.

A Hypothetical Next Best Action Example

A practical NBA system first determines which actions are eligible, then chooses among them using customer context, business rules, and, where mature enough, predictive or causal models.

Consider a hypothetical retailer evaluating what to do with a customer who has not purchased recently.

Four actions are available.

Replenishment Reminder

This action becomes eligible if the customer previously purchased a product that commonly requires replenishment.

Before sending, the system checks whether the customer has already reordered and whether the customer has current permission for the channel.

If the customer already purchased the replacement in store yesterday, the reminder should disappear from the candidate set.

Win Back Offer

A win back offer becomes eligible when the customer exceeds the business's documented lapse threshold.

But eligibility alone does not make the offer best.

Before selection, the system can consider:

  • Customer margin
  • Recent offers
  • Frequency cap
  • Known offer sensitivity
  • Whether the relationship is actually lapsed

A customer likely to return without a discount may be better assigned another action.

Loyalty Points Reminder

A loyalty member close to the next tier may benefit from a reminder about the existing program rather than a new discount.

The system needs current points and tier information.

Stale loyalty data can create a poor experience by telling the customer they are “almost there” after they already crossed the threshold.

Do Nothing

Do nothing is a real action.

The customer may already have received several messages.

They may have a recent service complaint.

The model may find that none of the available treatments creates enough expected value.

A mature next best action system should be able to conclude that the best customer experience is no outreach at all.

Govern Next Best Action Before The Decision Is Delivered

Next best action should operate inside clear eligibility, consent, sensitivity, frequency, and explainability constraints rather than allowing the model to choose from every technically possible action.

Check Consent And Channel Permission

Permission should be current when the action is delivered.

A high model score does not override an opt-out.

Share Frequency Caps Across Programs

Next best action should not ignore messages sent by lifecycle journeys, campaigns, service programs, and other systems.

Frequency needs a customer-level view.

Suppress During Sensitive Service Moments

An open complaint or major unresolved case may justify suppressing promotional actions.

The service relationship can matter more than the marketing opportunity.

Control Sensitive Features

Sensitive information should not automatically become a model feature merely because the CDP contains it.

Features involving sensitive data or sensitive inferences should go through the appropriate privacy, legal, and governance review.

Keep A Reason Code

The system should record why an action was selected.

That does not require exposing every internal model calculation to a customer-facing employee.

It does require enough context for teams to investigate and explain decisions.

Test For Unfair Patterns

Compare action frequency and outcomes across relevant groups where legally and operationally appropriate.

Potentially unfair patterns should trigger review rather than being accepted because overall model performance appears strong.

Prove Next Best Action With Control Groups, Not Click Rates

NBA performance should be measured against what would have happened without the treatment, not only by counting clicks, conversions, or revenue after exposure.

BCG's 2026 measurement guidance similarly emphasizes global and local control groups for separating incremental contribution from activity that would have happened anyway.

Create A Persistent Control

Withhold NBA treatment from a randomized share of eligible customers where appropriate.

The control can receive:

  • No intervention
  • The existing business-as-usual treatment

depending on the decision being tested.

Measure A Business Outcome

Use the outcome that the system is designed to change.

Examples include:

  • Incremental purchase
  • Renewal
  • Retained margin
  • Expansion
  • Activation

Opens and clicks can explain behavior.

They should not be treated as proof of incremental economic value.

Protect The Control Across Channels

If the customer is held out from an NBA email but receives the equivalent offer through another channel, the experiment is contaminated.

Control assignment should therefore be visible across the activation architecture.

Do Not Assume Universal Lift

Next best action performance varies by business, action library, customer population, baseline treatment, channel, and decision quality.

Stable Kernel does not use a universal lift assumption for NBA.

The organization's own experiment should decide whether the system creates sufficient incremental value.

Where Next Best Action Programs Go Wrong

NBA programs fail when architecture, modeling, operations, measurement, and customer data mature at different speeds.

BCG organizes many NBA challenges into four structural gaps: architecture, science, operating model, and measurement.

The customer-data layer adds several related failure modes.

The Architecture Becomes A Web Of Overlapping Rules

Every team creates its own campaigns, journeys, and exceptions.

Eventually, several actions compete for the same customer and nobody owns the arbitration.

The fix is a governed action library and one decision point.

The Model Predicts Likelihood Instead Of Impact

A propensity model finds customers likely to purchase.

The organization then gives incentives to customers who may have purchased anyway.

The fix is experiment data and, when maturity allows, uplift or other causal approaches.

The Organization Is Still Structured Around Campaigns

Teams own their individual programs but nobody owns the customer's combined decision environment.

The result is local optimization and global inconsistency.

The action library needs shared ownership and conflict rules.

Measurement Tracks Activity

The program counts impressions, opens, clicks, or attributed revenue.

None of those measures establishes incrementality.

The fix is a control architecture designed before launch.

The Action Log Does Not Exist

This is one of the most damaging data-side gaps.

If the enterprise cannot reconstruct which action a customer received, it cannot train or validate future decisioning properly.

Outcomes Are Not Connected To Identity

An order occurs, but the organization cannot reliably connect it with the customer who received the action.

The learning loop breaks.

Features Cannot Be Reproduced At Decision Time

A data science team builds a model using historical features that production cannot calculate consistently.

The model looked good in development but operates on different information in production.

The feature pipeline needs point-in-time reproducibility.

A Seven-Step Method For Building Next Best Action With CDP Data

Step 1: Pick One Decision

Start with a narrow question.

For example:

What should we do next for a customer approaching lapse?

Avoid trying to solve all personalization decisions at once.

Step 2: Define The Candidate Actions

List the possible actions, including do nothing.

For each one, document:

  • Eligibility
  • Cost
  • Channel
  • Suppressions
  • Expected outcome

This creates the action library.

Step 3: Validate The CDP Data

Confirm that the required transaction, behavior, service, consent, and identity data actually exists.

Check match rates and important gaps.

Do not build the decision model on a customer history you have not validated.

Step 4: Build The Action And Outcome Log

Do this before launch.

Record every selected action and the later outcome tied to the same identity.

This data becomes the foundation for better future decisioning.

Step 5: Start With Rules And Add Propensity When Useful

Rules can validate the operating model before machine learning adds complexity.

Add propensity scoring when history, volume, and outcome definitions support it.

Step 6: Launch With A Control Group

Choose one business outcome and preserve a comparison group.

Separate technical delivery metrics from incremental business impact.

Step 7: Review Before Increasing Sophistication

Review the program after enough outcome data matures.

Example review points such as 30 and 90 days can be useful planning placeholders, but the correct period depends on the customer decision cycle.

Move toward uplift or adaptive decisioning only when the experiment history, action volume, governance, and measurement justify it.

How Stable Kernel Approaches Next Best Action

Stable Kernel treats next best action as a customer-data architecture and decisioning problem rather than a model implementation in isolation.

Build The Data And Feature Foundation First

Stable Kernel can help enterprise teams connect identity, transactions, product behavior, service data, consent, and outcome events into the profile and feature pipelines required by next best action systems.

That includes designing the exposure and outcome logging needed for continuous learning.

Connect CDP Data To Governed Decisioning

Stable Kernel can also help teams integrate customer data with decisioning tools, define action eligibility and suppression requirements, establish real-time or batch delivery patterns where appropriate, and design the measurement needed to determine whether selected actions create incremental value.

If your organization wants next best action but cannot answer which actions are available, what each action costs, what the customer has already seen, why one action was chosen, what happened afterward, and what equivalent customers did without the treatment, adding a more sophisticated model will not solve the underlying problem.

Request a next best action data readiness review from Stable Kernel to evaluate the customer profile, feature pipelines, action and outcome logging, governance rules, and measurement architecture required before scaling decisioning complexity.

Three Questions Before Building A Next Best Action Model

1. What Decision Are We Trying To Improve?

Define one customer decision rather than beginning with a general goal such as “personalization.”

A narrow decision creates a clearer action library and outcome.

2. Do We Know What Actions Customers Already Received?

If exposure history is missing, add it before trying to build a learning system.

A model cannot reliably learn action effectiveness from an action it cannot observe.

3. How Will The Control Group Tell Us It Worked?

If the program measures only customers who receive the selected action, it cannot separate changed behavior from behavior that would have happened anyway.

FAQ

What Is Next Best Action?

Next best action is an approach for choosing the most relevant next step for one customer at a particular moment. The action may be a message, offer, service step, product experience, sales task, or no contact. It can use simple business rules, predictive models, or more advanced decisioning.

How Does A CDP Support Next Best Action?

A CDP provides the unified customer profile, behavioral and transaction history, identity, features, consent state, and other context required by next best action models and rules. The final action may be selected by a separate decisioning layer and delivered through email, mobile, web, CRM, or another channel.

What Data Do You Need For A Next Best Action Model?

Typical inputs include transaction history, digital or product behavior, service history, messaging behavior, customer identity, consent and preferences, and derived features such as recency and frequency. The system also needs an exposure log showing which action was selected and outcome events showing what happened afterward.

Do You Need Machine Learning For Next Best Action?

No. Rules built from CDP customer data can be an effective first version when the action library is small and the business logic is understandable. Machine learning becomes more useful when there is enough historical data, a clear outcome, and many possible customers or actions to rank.

What Is The Difference Between Next Best Action And Next Best Offer?

Next best offer chooses among possible offers. Next best action chooses among a broader set of possibilities, including product guidance, service actions, messages, offers, sales tasks, and doing nothing.

What Is The Difference Between Propensity And Uplift Models?

A propensity model predicts how likely a customer is to produce an outcome. An uplift model estimates how much a specific intervention changes that likelihood compared with no intervention. Uplift is particularly useful when an action has a cost and the business wants to avoid rewarding customers who would have acted anyway.

How Do You Measure Whether Next Best Action Works?

Maintain a randomized or otherwise credible control group that does not receive the NBA treatment and compare the primary business outcome. Measures such as opens, clicks, and attributed revenue can provide context, but they do not by themselves demonstrate incremental value.

Why Do Next Best Action Programs Fail?

Common causes include overlapping rules, weak customer identity, missing action and outcome logs, models that predict likelihood rather than treatment impact, teams organized around separate campaigns, and measurement that tracks activity without a control group. Decisioning complexity often grows faster than governance and measurement.

Does Next Best Action Need Real-Time Data?

Not always. Many next best action decisions can operate successfully on daily or scheduled customer data. Faster decisioning matters when the opportunity changes quickly, such as current-session web or application experiences, immediate service context, or other moments where delay makes the action stale.

How Is Next Best Action Governed?

Governance should check consent and channel permission, shared frequency limits, service suppressions, sensitive-data restrictions, eligibility, and other applicable policies before the action is delivered. The system should also retain a reason code and enough decision history to explain and audit why an action was selected.