Telecom Next Best Action Reference Architecture

Establish a telecom architecture that enables real-time next best action at scale

Data and platform flows:

Next Best Action (NBA), Next Best Experience (NBX)

1. Data Ingestion from Telco Operational Systems
   Customer behavior, billing, CDRs (call/data/SMS records), CRM, recharge, and fintech transaction data are collected from hybrid/on-prem systems via BSS/OSS integration interfaces or third-party data connectors like Informatica and LakeFlow. These data streams are sent in near real-time or batch to the Streaming Ingestion Layer (e.g., Kafka, Azure Event Hub), where lightweight preprocessing occurs (e.g., PII filtering, protocol conversion).
2. Unified Lakehouse Storage and Raw Data Staging
   Ingested data lands in Delta Lake Bronze tables, partitioned by attributes like OpCos, site ID or channel. At this stage, operation like schema enforcement and PII masking may occur. From there data progresses into Silver tables, where it's deduplicated, joined, and enriched to build Customer 360 profiles, usage summaries, and churn indicators—creating a harmonized, analytics-ready foundation.
3. Feature Engineering and Model Training
   Cleaned Silver data feeds into advanced transformation pipelines using Lakehouse Pipelines. This stage includes:
   • Deriving intent features and registering them in a centralized Feature Store decoupled from channel- or region/country-specific logic.
   • Enabling point-in-time enrichment and audience identity stitching.
   • Training NBA models using Mosaic AI, including use cases like churn prediction, offer acceptance scoring, and Reinforcement Learning for dynamic offer optimization.
   • Models are trained and managed within the unified platform with lineage, observability, and governance through Unity Catalog.
   • A critical feedback loop in this architecture is A/B testing:
     1. Models are continuously assessed via batch lift testing embedded into CICD process measuring business impact (e.g., uplift, retention, conversion).
     2. These insights inform whether the current models are outperforming baselines to enable objective decisions on rollout or retraining.
4. Business Insights and CX Activation
   Scored outputs from ML pipelines are stored in Gold tables, ready for BI and real-time CX applications:
   • Databricks SQL enables real-time KPIs, subscriber behavior analysis, and ROI measurement of NBA campaigns.
   • Executive and marketing teams access this data via familiar tools like Power BI, Tableau, and Looker.
   • Lakehouse apps enable audience segmentation, campaign simulation, and cross-sell/up-sell targeting for CX teams.
5. Model Deployment and Real-Time Inference
   Trained NBA models are exposed for real-time activation via:
   • Model Serving APIs hosted on Mosaic AI or Databricks Model Serving, supporting hybrid deployments.
   • APIs are deployed or proxied into cloud-based and/or on-prem operational environments for direct integration with SMS systems, call centers, mobile apps, or IVRs.
   • NBA outcomes are continuously monitored with dashboards, A/B test results, and model performance tracking to ensure relevance and minimize drift.