Selected writing

Governed Agentic Automation for Chargebacks: A Prompt-First Architecture for Policy-Driven Enterprise Workflows

Authors: Nataraj Agaram Sundar; Tejas Morabia.

Generative AI adoption is outpacing reliable productionization, especially in enterprise workflows that combine structured inputs, policy constraints, and operational risk. This technical report describes a prompt-first architecture for chargeback automation that replaces brittle, retrieval-heavy integrations with a governed orchestration layer for prompt management, policy enforcement, and observability. Using chargeback dispute handling as a case study, the report examines the transition from a Retrieval-Augmented Generation (RAG) design to modular prompt chaining and agentic execution over structured inputs. In internal deployment measurements, the prompt-first approach reduced generation latency by 35%, improved internal quality scores by 20%, and enabled near-immediate policy updates without re-indexing. A centralized guardrail framework auto-resolved 85% of policy issues before human review, while inter-rater agreement between reviewers improved from Cohen's kappa 0.65 to 0.82 across three optimization cycles. The results suggest that for decision-grade automation over structured enterprise data, governed prompt orchestration can outperform more complex retrieval-heavy designs in speed, controllability, and operational fitness. Some system names and implementation details are generalized.

Abstract page · PDF