Call Transcript Multi‑Level Classification System (L1/L2 Categorisation using GenAI)
2025-08 – 2025-12
This project delivers an automated GenAI‑powered classification system that analyses customer–agent call summaries and assigns both **L1 (high‑level)** and **L2 (granular)** categories. Built as a production‑ready workflow on AWS, the solution enhances operational analytics, improves customer‑service insights, and supports downstream reporting teams with structured, high‑quality metadata extracted directly from unstructured call transcripts.
Scope
- Automatically classify call summaries into pre‑defined L1 and L2 categories used across Santander’s customer‑care operations.
- Replace manual tagging with a consistent, scalable, and auditable GenAI‑driven pipeline.
- Provide structured JSON outputs suitable for analytics, compliance, and quality‑assurance workflows.
- Deploy the full Proof of Concept (PoC) using AWS SageMaker, ensuring secure, enterprise‑grade experimentation and model hosting.
- Store classification results in Amazon DynamoDB for fast retrieval and integration with internal dashboards.
Technology Stack
- OpenAI GPT‑5 — core LLM used for semantic understanding and multi‑level category extraction.
- AWS SageMaker — model orchestration, experimentation, and deployment environment.
- Amazon DynamoDB — NoSQL storage for structured L1/L2 classification outputs.
- Python — prompt engineering, data processing, and pipeline logic.
- JSON — standardised output format for downstream systems.
Prompting Technique
The classification engine is powered by a chain‑of‑thought prompting strategy designed to guide the model through structured reasoning:
- The prompt begins by instructing the model to first summarise the call, ensuring it understands the context before classification.
- It then performs L1 reasoning, mapping the call to a broad category such as Banking Products, Savings Accounts, Loans, etc.
- Next, the model performs L2 reasoning, selecting the most granular sub‑category associated with the chosen L1.
- The prompt includes few‑shot examples demonstrating correct L1/L2 mappings, improving consistency and reducing ambiguity.
- The final output is formatted as a clean JSON object, containing:
call_summaryl1_categoryl2_categoryconfidence_notes(optional reasoning trace)
This structured prompting approach ensures the model produces reliable, explainable, and audit‑friendly classifications.
Outcome
The PoC successfully demonstrated that GenAI can classify call summaries with high accuracy, significantly reducing manual effort and improving consistency across customer‑care operations. The system provides a scalable foundation for future enhancements such as multi‑label tagging, sentiment analysis, and integration with real‑time call‑centre workflows.