Treat different spellings as candidates to review, not automatic proof that two mentions refer to the same person. Build a local alias list, assign one placeholder to each confirmed identity and inspect the exported text in every language present.
Build an alias list before replacing
A bilingual document may use a local-script name, a Latin spelling, a shortened signature and a surname alone. List the forms actually present in the source, together with their context. Include email display names, headers and quoted correspondence.
Use a fictional example: Мария Волкова, Mariya Volkova and M. Volkova might denote one person in a file. Confirm that through the document rather than assuming it. Do not add real client identifiers to an online transliteration tool merely to generate more variants.
Search accents, punctuation and character variants
Search both complete forms and distinctive fragments. Hyphens, apostrophes, spaces and accent marks can change what an exact search finds. Copy a suspicious spelling from the source when retyping it produces no result.
Unicode describes normalization for equivalent character sequences; that is different from deciding that two personal names belong to the same person. Even normalised text may contain different transliterations, OCR mistakes or genuinely different names. Review the surrounding sentence.
Resolve initials and short forms in context
Initials are ambiguous. A. Silva could be Ana or André; a surname shared by two relatives is not a unique identity. Check the signature, role, addressee and chronology before assigning a placeholder. If the source cannot resolve the reference, preserve that uncertainty.
Company abbreviations need the same care. A trading name, registered name and abbreviation may refer to one organisation, while a similar abbreviation refers to another. Keep person and organisation placeholders distinct.
Preserve meaning across languages
Apply the same confirmed identity consistently across the case, while keeping different people separate. Check possessive forms, grammatical endings and titles around the replacement. A replacement can remove the right word but leave a phrase unreadable.
Read bilingual tables and translated quotations together. The local-language column may retain a full name after the English column is cleaned. Do not treat the English text as the master copy of what the AI will receive.
Review ClientSafe matches and the final copy
ClientSafe uses consistent placeholders within a case and lets you review detected data. Use that review to reject false positives and add missed forms. Do not assume automatic detection covers every writing system, transliteration or abbreviation.
Export a separate text copy and search it again using your alias list. Inspect signatures, headings and short isolated labels manually. Keep the alias list and replacement key out of the material sent to the AI service.
Check clues that survive the name
Removing every spelling can still leave a recognisable biography, rare role or precise address. Omit details that are not needed for the question, and reread the prompt along with the copy. The AI service and your organisation's sharing rules still need separate approval.
Before replacing more aggressively, check whether you are destroying a distinction the analysis needs. It is better to state that two references might denote the same participant than to silently combine them and obtain a confident but wrong summary.
Frequently asked questions
Should all similar spellings get one placeholder?
No. Similarity is a reason to investigate. Merge mentions only when the document supports that identity; keep ambiguous references explicit.
Does removing accents anonymize a name?
No. The name may remain recognisable. Replace identifying references consistently and check other details in the surrounding text.
Sources and further reading
Reviewed 3 November 2026