Sanctions screening: UN, OFAC, fuzzy matching and Urdu transliteration

Three ways to see it

1. 1

   Sanctions screening is the practice of checking every customer, every transaction counterparty, and every payment narration against published lists of sanctioned individuals, entities, and jurisdictions. For a Pakistani bank, the in-scope lists include the UN Security Council Consolidated List, the United States OFAC SDN List, the EU consolidated list, the UK HMT list, FBR's local proscribed-persons list, NACTA's Schedule IV list under the Anti-Terrorism Act, and the National Counter Terrorism Authority lists. SBP's AML/CFT Regulations require screening at four points: customer onboarding, customer due diligence refresh, every cross-border transaction, and every domestic transaction above defined thresholds. Sanctions failure is not a fine, it is an existential risk: a confirmed sanctions breach can trigger correspondent-bank de-risking and effectively cut a Pakistani bank off from US dollar clearing.

2. 2

   The technical heart of screening is fuzzy matching. Customer names rarely arrive in the exact spelling on the sanctions list. Common transformations include: missing or extra middle names, transliteration variants (Mohammed, Mohammad, Muhammad, Mohamed all map to one Arabic name), order swaps (last name first), abbreviated honorifics (Md, Mohd, M), spelling variants (Hussein, Hussain, Husain, Husayn), and pure typographical errors. A fuzzy matcher computes a similarity score between two strings using algorithms like Levenshtein edit distance, Jaro-Winkler similarity, or phonetic codes such as Soundex and Metaphone. The threshold is the score above which a match raises an alert. Lower threshold catches more sanctioned matches and produces more false positives; higher threshold misses sanctioned individuals.

3. 3

   Pakistani-name screening has its own engineering quirks that off-the-shelf western matchers handle poorly. Urdu names round-trip through Arabic, Persian, and English phonetic systems with multiple valid spellings. The single name Muhammad has at least 14 documented English transliterations. Tribe and clan names stack ambiguously: 'Khan' is a surname, an honorific, and sometimes a placeholder. The honest fix is a Pakistani-context match dictionary that explicitly maps known transliteration equivalents, plus a phonetic algorithm tuned for Urdu phonemes (the standard Soundex was designed for English and produces nonsense for Urdu sounds). Vendors like LexisNexis Bridger and Accuity have Pakistan packs. In-house teams typically build a curated mapping of the top 5,000 Urdu-origin names with all known English variants and merge it into the matcher.

The why-tree

Try this with Claude

AI-edge prompt: 'I am the AML head at a Pakistani bank with 3.2 million customers. Help me design a sanctions screening matcher tuned for Urdu and Punjabi name variants. List the components I need (transliteration dictionary, phonetic algorithm, fuzzy distance metric, threshold curve), recommend open-source libraries, and give me a worked example screening the name Muhammad Hussain Bhatti against a 12-name watchlist that includes one valid hit. Explain why your matcher catches the hit at the threshold I should choose.' Compare to your current vendor output.