18 November 2026
Why exact name matching fails for sanctions screening
Transliteration, word order, missing names and typos make exact matching miss sanctioned people. Here is what a modern name-matching approach looks like.
Imagine a sanctions list entry for Mohammed Abdullah Al Rashid. Over a week, your onboarding flow sees these customers:
- Muhamad al-Rasheed
- Al Rashid Mohammed
- Mohd Abdullah Rashid
- Mohammed Al Rashid, Dr.
To a person, these could all be the same individual. To an exact-match system, none of them are on the list.
Four reasons names do not match exactly
1. Transliteration. Names written originally in Arabic, Cyrillic, Devanagari or Chinese script have several legitimate Latin spellings. There is no single “correct” form, and lists themselves often record several.
2. Word order and structure. Surname-first and given-name-first conventions differ by country and by document. Patronymics, middle names and second surnames appear in some records and not in others.
3. Short forms and national variants. Alexander and Sasha, Mohammed and Mohd, Dmitry and its Greek form are the same name for screening purposes.
4. Noise. Typos, doubled letters, missing spaces, titles such as Dr or Sir, and punctuation all change the string without changing the person.
Why rules alone do not scale
The classic fix is to generate variations with rules: swap letters, reorder parts, drop the middle name. Each new rule multiplies the number of variations, and each variation is another lookup. Teams that go down this path see query volume grow exponentially, while still missing variations nobody wrote a rule for.
Why loose fuzzy matching is not the answer either
The opposite approach is to lower a string-similarity threshold until everything is caught. The result is an alert queue full of people who merely share a common first name, plus pairs like Maria and Marina, which differ by one letter but are different names. Analysts lose time, and customers wait.
A two-stage approach
Modern name screening separates recall from precision:
- Retrieve widely. Encode every name as a vector of character n-grams and use approximate nearest-neighbour search. Spelling variants and reordered names land close together, so candidates are found in milliseconds without generating variations.
- Verify precisely. Pass only the top candidates to a model that understands names. It accepts Alexander / Sasha and rejects Alexander / Alexey.
- Use the other signals. A matching date of birth, even a year or a range, raises confidence. A conflicting one lowers it.
- Explain the result. Every match should show which alias matched and how each stage scored it, so analysts can decide quickly and auditors can follow the reasoning.
This is the architecture behind AutoCompliance. The retrieval stage keeps recall high and is cheap to update when lists change. The verification stage can be improved independently, without re-indexing millions of names.
What to ask your screening vendor
- How do you handle transliteration and name order without generating variations?
- Can you show me why a specific match was returned?
- How do you treat single-word names and very common names?
- Which list version was used for a given result, and can you reproduce it later?