From November 2026, the Swift network will implement a significant change by removing unstructured postal addresses from payment messages. Participants and customers using the ISO 20022 HVPS+ guidelines must transition from free-text address fields, especially in debtor/creditor and agent sections of MT 103 and pacs.008 messages, to the structured ISO 20022 CBPR+ format, which mandates the inclusion of Town and Country fields. 

The transition is part of the larger Standard Release 2026 and follows the ISO 20022 CBPR+ guidelines. 

Currently, many banks and businesses use free-text or semi-structured address fields. This makes it hard to comply with regulations and to work together in a global payment system that increasingly depends on digital standards. The removal of unstructured addresses will enhance data quality, ensure regulatory compliance, and improve operational efficiency by standardising address formats.

Swift’s AI-based address structuring model

To support this transition, Swift has built an AI-driven address structuring model using Natural Language Processing (NLP). The model extracts structured data, like Town and Country, from unstructured legacy addresses in corporate databases or payment files. 

It also offers confidence scores and resolution options for automated processing and expert review. It ensures transparency and auditability.

The model is trained using data from over 200 countries and regions, including sources like GeoNames and OpenStreetMap. Organisations can adjust the model with their proprietary data or regional registries to meet specific jurisdictional requirements.

The model focuses only on extracting the Town and Country from debtor and creditor fields in MT 103 and pacs.008 messages. It does not extract full postal addresses or manage certain agent fields, which are typically identified by BIC codes. 

The model performs optimally with English or transliterated inputs and may produce unreliable results with non-Latin scripts such as Arabic, Chinese, Cyrillic, or Japanese.

It is intended for financial institutions, service bureaus, Swift partners, and corporate ERP systems that need to clean and structure address data to meet the ISO 20022 migration deadline.

How will the implementation work? 

Swift will offer a free, open-source AI address structuring model in November 2025, free of charge. The model can be integrated into internal systems or used for batch-processing legacy payment messages.

It uses a Conditional Random Field (CRF) approach to effectively tokenise address strings and accurately identify location entities using linguistic and contextual features. Training is based on diverse synthetic datasets to ensure global applicability.

No personal data was used in training the model, which operates offline within user systems.

Users need to adhere to data protection regulations.

While Swift offers documentation and installation guidance, ongoing support and maintenance will not be provided.