Next Best Action Recommendation System

Overview

The Next Best Action (NBA) system is an AI-driven decisioning engine that recommends the most effective action a commercial user should take for a given account or healthcare provider (HCP). Rather than focusing on isolated message recommendations, the system prioritizes what to do next—such as outreach, channel selection, and follow-up strategy.

The system was deployed globally and used across multiple regions, supporting sales and marketing teams with consistent, data-driven guidance. It enabled organizations to move from manual, intuition-driven workflows to scalable, prioritized engagement strategies grounded in behavioral data.

Problem

• Sales and marketing decisions were manual, inconsistent, and dependent on individual experience
• No unified prioritization existed across accounts, channels, or actions
• Large volumes of interaction data were underutilized due to noise and fragmentation
• Feedback signals were sparse and indirect, making optimization difficult
• Naive recommendation approaches led to repetitive actions and reduced effectiveness over time

Constraints

• Large-scale datasets with high-cardinality entities (accounts, users, actions)
• High-volume interaction logs requiring distributed batch processing
• Noisy behavioral data requiring filtering, aggregation, and feature extraction
• Sparse feedback signals where engagement is only a proxy for success
• Need to balance relevance with diversity to avoid over-targeting and fatigue
• Computational constraints when generating recommendations at scale

Approach

Framed the problem as a ranking task: for each account, generate and rank a set of candidate actions based on expected impact.

Built a batch-driven recommendation pipeline consisting of:

• Distributed data processing using Dask on Kubeflow to scale pandas-based transformations
• Interaction preprocessing to filter low-signal events, aggregate behavior, and extract features such as recency, frequency, and engagement strength
• A Neural Factorization Machine (NFM) model to capture higher-order feature interactions across users, accounts, and actions
• Post-processing steps to ensure diversity and avoid repetitive recommendations

In addition, introduced configurable business-rule controls (“knobs”) that allowed sales and marketing users to tune recommendation outputs based on strategy, campaign goals, or regional needs.

Prioritized:

• Data quality and feature engineering
• Scalable batch processing over real-time systems
• Flexibility to adapt recommendations to business constraints

System Design

The system is a distributed batch pipeline built on Kubeflow and Dask.

Data Processing Layer

• Interaction data is processed using Dask clusters deployed via Kubernetes
• Jobs are orchestrated through Kubeflow Pipelines
• Data is partitioned by region and segment to support global deployments
• Preprocessing includes filtering irrelevant interactions, aggregating events, and constructing behavioral features

Modeling Layer

• Neural Factorization Machine (NFM) is trained offline on aggregated interaction data
• Feature space includes behavioral signals, contextual attributes, and temporal patterns
• Model captures non-linear interactions between users, accounts, and actions

Recommendation Layer

• Candidate actions are generated per account in batch
• Actions are scored and ranked using the trained model
• Business-rule tuning and post-processing adjust outputs based on strategic priorities and diversity considerations

Recommendations are materialized and consumed downstream in sales and marketing workflows.

Key Decisions

Frame the problem as action-level ranking

Shifting from message-level recommendations to action-level decisioning aligned the system with real-world workflows and enabled better prioritization across channels and strategies.

Use Neural Factorization Machines for modeling

NFM allowed the system to capture higher-order feature interactions while remaining practical for large-scale batch training and inference.

Introduce business-controlled tuning mechanisms

Exposing configurable knobs enabled:

• Alignment with evolving business strategies
• Flexibility across regions and user roles
• Faster iteration without retraining models

However, this also introduced variability depending on how effectively users tuned the system.

Build a batch-first distributed system

A batch architecture provided:

• Scalable processing of large interaction datasets
• Operational simplicity and reliability
• Flexibility to adapt pipelines across regions

Results & Impact

• Adoption: Recommendation utilization exceeded 90% in several deployments, indicating strong user trust and alignment with workflows
• Field Productivity: 93% increase in top-tier accounts achieving call plan targets
• Engagement:
  • 4x increase in rep-triggered email send rates
  • 14x increase in marketing consent opt-ins
• Scale: Deployed across global teams, supporting thousands of users and adapting to regional engagement patterns

The system enabled a shift from manual prioritization to consistent, data-driven decision-making across commercial operations.

Tradeoffs

• Batch processing limited responsiveness to recent behavioral changes
• Neural models increased dependency on feature quality and training data consistency
• Business-rule tuning improved flexibility but introduced user-dependent performance variability
• Lack of a closed feedback loop meant the system did not automatically learn from user adjustments or outcomes
• Compute cost increased with large-scale batch scoring

Learnings

• Combining ML ranking with user-controlled business logic is often necessary in enterprise systems
• Model sophistication alone is insufficient—how outputs are controlled and consumed is equally important
• User-level tuning improves adoption but creates uneven outcomes across users
• Absence of a feedback loop limits long-term system optimization
• Neural models are effective but remain heavily dependent on upstream data quality

Future Work

• Introduce a closed feedback loop to learn from user actions and outcomes
• Replace manual tuning with adaptive, data-driven optimization
• Implement contextual bandits to balance exploration and exploitation
• Improve consistency across users through guided tuning strategies
• Incorporate sequence-aware models for multi-step action planning