Next Best Action
“What product should this customer see next?”
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
By submitting, you accept the Terms and Privacy Policy.
Loved by data scientists, ML engineers & CXOs at
A real-world example
What product should this customer see next?
Banks send the same promotional offers to broad segments, resulting in 1-2% engagement rates on digital banners and email campaigns. A large retail bank running 40+ campaigns per quarter found that 78% of customers received irrelevant offers, leading to opt-out rates above 15%. The challenge is that the right action depends on a customer's full context: recent transactions, life stage, product gaps, service interactions, and real-time digital behavior. Rule-based decisioning engines cannot weigh thousands of signals across dozens of tables in real time.
Quick answer
The best next-best-action models for banking connect transaction patterns, life-event signals, product holdings, interaction history, and digital behavior into a relational graph. This lets the model match the right product to each customer's actual life context (new parent, pre-retirement, early career) rather than their demographic segment. Engagement rates jump from 1-2% to 4-6% because the offer is relevant to what the customer is actually experiencing.
Approaches compared
4 ways to solve this problem
1. Rule-based campaign logic
Assign offers based on segment rules: Premier customers get wealth reviews, Mass Market gets savings accounts, new customers get credit cards.
Best for
Simple to implement. No model needed. Campaign teams can execute immediately.
Watch out for
1-2% engagement rates. 78% of customers receive irrelevant offers. Opt-out rates climb above 15%, meaning you are actively driving customers away from your marketing channel.
2. Propensity model per product
Build a separate classification model for each product (mortgage propensity, credit card propensity, insurance propensity) and rank products by score.
Best for
Better than rules. Each product model can be tuned independently. Widely used at large banks.
Watch out for
Building and maintaining 15+ separate models is expensive. Models compete with each other (a customer might score high for both mortgage and insurance) with no coordination. Product-level models also miss cross-product timing: the right action depends on what just happened, not just what the customer looks like.
3. Collaborative filtering
Recommend actions based on what similar customers engaged with: 'customers like you accepted a wealth review.'
Best for
Good for digital channel recommendations with high-volume interaction data. Works without deep domain knowledge.
Watch out for
Ignores individual behavioral context. Two Mass Affluent customers in the same segment may need completely different actions if one just had a baby (insurance) and the other is planning retirement (wealth review).
4. KumoRFM (relational graph ML)
Connect customers to transactions, products, interactions, and life events. The GNN learns which action to recommend by reading each customer's full behavioral context, re-scoring daily as new data arrives.
Best for
Captures the full context: baby-related spend + income growth + no insurance = term life consultation. Prior offer responses + channel preferences = mobile app delivery. Lifts engagement 2-4x over segment-based campaigns.
Watch out for
Requires interaction-history data with outcome labels (accepted, clicked, ignored). Cold-start problem for brand-new customers with no behavioral data.
Key metric: NBA models powered by relational ML increase customer engagement rates from 1.5% to 6.2%, a 4x lift over segment-based campaign targeting.
Why relational data changes the answer
The 'next best action' depends on context that spans multiple tables. Transaction data reveals life events (baby spend, travel patterns, home improvement). Product holdings show gaps (no insurance, no investment account). Interaction history shows channel preferences and prior offer responses. Digital behavior shows real-time intent (rate comparison pages, investment calculators). No single table contains the full picture.
Relational models read all of these tables simultaneously and learn that Customer C-10042's baby-related transactions + income growth + protection product gap = 'term life insurance consultation via mobile app' with 72% acceptance probability. This is not a rule someone wrote. The model learned it from thousands of similar customers whose behavioral trajectories led to insurance adoption. The result: engagement rates jump from 1.5% to 6.2% because the offer matches the customer's actual life context.
Segment-based offers are like a waiter who recommends the same dish to every table. A good waiter reads the context: the couple looking at the wine list wants a bottle recommendation, the family with kids needs the children's menu, and the solo diner reading a book wants to be left alone. The relational model reads the customer's full context, not just their table number.
How KumoRFM solves this
Relational intelligence built for banking and financial data
Kumo builds a relational graph connecting each customer to their transactions, product holdings, service interactions, life events, and digital behavior. The model learns that Customer C-10042 just received a raise (higher direct deposits), started a family (baby-related spend), and has no life insurance. Instead of a generic credit-card upgrade offer, Kumo ranks 'term life insurance consultation' as the highest-propensity action. The NBA model re-scores daily as new transaction and event data flows in.
From data to predictions
See the full pipeline in action
Connect your tables, write a PQL query, and get predictions with built-in explainability — all in minutes, not months.
Your data
The relational tables Kumo learns from
CUSTOMERS
| customer_id | age | life_stage | products_held | segment |
|---|---|---|---|---|
| C-10042 | 34 | New Parent | 3 | Mass Affluent |
| C-10078 | 52 | Pre-Retirement | 6 | Premier |
| C-10115 | 28 | Early Career | 1 | Mass Market |
TRANSACTIONS
| customer_id | merchant_category | amount | frequency_30d | timestamp |
|---|---|---|---|---|
| C-10042 | Baby & Kids | $420 | 8 | 2025-09-10 |
| C-10042 | Daycare | $2,100 | 1 | 2025-09-01 |
| C-10078 | Travel | $3,800 | 3 | 2025-09-05 |
PRODUCT_CATALOG
| product_id | product_name | category | eligibility_segment |
|---|---|---|---|
| P-101 | Term Life Insurance | Protection | Mass Affluent+ |
| P-102 | 529 College Savings | Investing | All |
| P-103 | Travel Rewards Card | Cards | Premier |
INTERACTION_HISTORY
| customer_id | channel | offer_shown | outcome | timestamp |
|---|---|---|---|---|
| C-10042 | Credit Card Upgrade | Ignored | 2025-08-15 | |
| C-10042 | App | Savings Rate Boost | Clicked | 2025-08-20 |
| C-10078 | Branch | Wealth Review | Accepted | 2025-09-02 |
Write your PQL query
Describe what to predict in 2–3 lines — Kumo handles the rest
PREDICT ARGMAX(INTERACTION_HISTORY.OUTCOME = 'Accepted') FOR EACH CUSTOMERS.CUSTOMER_ID WHERE CUSTOMERS.SEGMENT IN ('Mass Affluent', 'Premier')
Prediction output
Every entity gets a score, updated continuously
| CUSTOMER_ID | RECOMMENDED_ACTION | PROPENSITY | CHANNEL | PRIORITY |
|---|---|---|---|---|
| C-10042 | Term Life Insurance Consult | 0.72 | Mobile App | 1 |
| C-10042 | 529 College Savings | 0.58 | 2 | |
| C-10078 | Travel Rewards Card Upgrade | 0.81 | Branch | 1 |
Understand why
Every prediction includes feature attributions — no black boxes
Customer C-10042 (James Whitfield)
Predicted: Term Life Insurance Consultation (72% propensity)
Top contributing features
Life-stage signal (baby-related spend)
8 txns/30d
30% attribution
Income growth (direct deposit increase)
+18%
22% attribution
Protection product gap
No insurance
20% attribution
Prior offer response pattern
Savings clicked
16% attribution
Peer cohort adoption rate
34% of similar
12% attribution
Feature attributions are computed automatically for every prediction. No separate tooling required. Learn more about Kumo explainability
PQL Documentation
Learn the Predictive Query Language — SQL-like syntax for defining any prediction task in 2–3 lines.
Python SDK
Integrate Kumo predictions into your pipelines. Train, evaluate, and deploy models programmatically.
Explainability Docs
Understand feature attributions, model evaluation metrics, and how to build trust with stakeholders.
Frequently asked questions
Common questions about next best action
What is a next best action model in banking?
An NBA model predicts the single most effective product, offer, or interaction for each customer at any given moment. Unlike broad campaign targeting, NBA considers each customer's full context: recent transactions, life stage, product gaps, channel preferences, and prior offer responses. The goal is to present the right offer through the right channel at the right time.
How do you improve banking offer engagement rates?
Move from segment-based campaigns (1-2% engagement) to individual-level NBA scoring (4-6% engagement). The key is connecting transaction data (life-event signals), product holdings (gap analysis), and interaction history (channel and response preferences) into a model that scores each customer-action pair. Relevance drives engagement.
What data do you need for a next best action model?
Customer profiles, product holdings, transaction data (for life-event detection), interaction history with outcome labels (accepted, clicked, ignored), and a product catalog with eligibility rules. Digital engagement data (app usage, page views) adds real-time intent signals. The more behavioral context, the more relevant the recommendations.
How often should NBA scores be refreshed?
Daily at minimum. Life events (new baby, raise, home purchase) show up in transaction data within days, and the optimal action changes accordingly. Banks running weekly or monthly re-scores miss time-sensitive windows. Relational models that ingest streaming transaction data can re-score overnight with updated behavioral context.
What is the ROI of a next best action model?
A large retail bank running 40+ campaigns per quarter can expect $25-50M in incremental annual revenue from NBA-driven engagement. The math: 4x engagement lift on 10M customer touchpoints, with each converted interaction worth $50-200 in incremental product revenue. The cost of irrelevant offers (opt-outs, brand erosion) also drops significantly.
Bottom line: Increase customer engagement rates from 1.5% to 6.2% by matching the right product to each customer's life context, generating $25-50M in incremental annual revenue.
Related use cases
Explore more financial services use cases
Topics covered
One Platform. One Model. Infinite Predictions.
KumoRFM
Relational Foundation Model
Turn structured relational data into predictions in seconds. KumoRFM delivers zero-shot predictions that rival months of traditional data science. No training, feature engineering, or infrastructure required. Just connect your data and start predicting.
For critical use cases, fine-tune KumoRFM on your data using the Kumo platform and Research Agent for 30%+ higher accuracy than traditional models.
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
