Structuring unstructured address data for ISO 20022 compliance: Design, assurance and governance of an open-source AI utility for cross border payments

Abstract

The migration to ISO 20022 requires financial institutions to move from fully unstructured postal address fields towards structured or hybrid address formats, including explicit town and country elements. This creates a practical challenge for banks and corporates that hold large volumes of legacy free-text address data across jurisdictions, languages and local postal conventions. This paper examines that challenge through a case study of a narrowly scoped artificial intelligence (AI) utility designed to infer town and country from unstructured payment address data. The approach combines a compact machine learning model with reference-data matching, postal-code logic, synthetic training data and confidence-based outputs, enabling local deployment without reliance on external large language models or remote inference services. Validation across internal, external and independently reviewed datasets showed strong performance on representative payment-style data, while also identifying predictable failure modes in ambiguous, incomplete, non-standard and transliterated addresses. The case demonstrates that targeted AI can support ISO 20022 data remediation and cross-border payment efficiency where its operating boundary is explicit, its limitations are documented, and its outputs are subject to appropriate governance, validation and human review. This article is also included in the Business & Management Collection which can be accessed at http://hstalks.com/business.