What it does
Flagright automatically scans the free-text reference field of bank transactions (e.g., SWIFT MT103, ISO 20022) to extract person and company names. These extracted names are then passed to sanctions screening to check against watchlists.
How it works
An AI model reads the full reference string and identifies anything that looks like a real person or company name. It returns a list of extracted names — or an empty list if no names are found.
The model applies the following rules:
Extracted: Full person names (e.g., "John Doe") and company/entity names (e.g., "Globex GmbH") that appear in the reference text.
Ignored: Transaction codes, reference numbers, dates, invoice identifiers, standalone single letters, and other non-name content. If a reference only describes a payment type, order, or fee with no identifiable name, nothing is extracted.
No fabrication: The model only returns names that are explicitly present in the reference text. It will never guess or infer a name that isn't there.
Examples
Reference field | Extracted names |
| (none) |
| (none) |
| John Doe |
| Jane Smith, Globex GmbH |
Things to know
Extraction is based on the reference field only. It does not use other transaction fields such as sender or beneficiary name.
The model uses spacing and natural language patterns to distinguish names from codes and identifiers. References with unconventional formatting - such as names without spaces (e.g., "ALI" in "QUALIFY"), names with extra spacing (e.g., "J O H N D O E"), or names concatenated with codes - may not be reliably extracted.
Some legitimate names may be missed if they are ambiguous or embedded in unusual formatting. If you notice consistent gaps, contact your Flagright implementation manager.
Extracted names feed directly into your configured screening provider. A name appearing in the extraction output does not mean it is a sanctions match — it means it will be checked.