Localisation bugs in a fintech app are not the same category of bug as a broken button. A misformatted IBAN field or a mistranslated consent string can mean a rejected transaction, a compliance gap, or a support queue full of confused users. This article from M21Global covers the business and linguistic side of that problem well: why regulatory terms like SCA, KYC, or AML need market-specific equivalents, and why error messages in a payments app deserve more scrutiny than marketing copy.
What it doesn't cover, and what most of us actually have to build, is the engineering side: how do you structure a codebase so these mistakes don't happen in the first place? This is the implementation layer underneath that decision.
Stop hardcoding number and date formats
The 1,234.56 vs 1.234,56 problem isn't a translation problem, it's an engineering problem. If your codebase ever concatenates strings to build a currency value, you already have a bug waiting to ship.
Use the Intl API that's been in every modern JS runtime for years instead of rolling your own formatter:
const amount = 1234.56;
new Intl.NumberFormat('de-DE', {
style: 'currency',
currency: 'EUR'
}).format(amount);
// "1.234,56 €"
new Intl.NumberFormat('pt-PT', {
style: 'currency',
currency: 'EUR'
}).format(amount);
// "1234,56 €"
Same currency, same value, different grouping and decimal separators, different symbol placement. This is exactly the class of bug the source article flags as a trust-breaker on a payment confirmation screen, and it's entirely preventable with a formatter instead of a template string.
Same logic applies to dates. new Date().toLocaleDateString('en-GB') and .toLocaleDateString('en-US') will give you 17/06/2025 and 6/17/2025 respectively from the same Date object. If a KYC form has a date-of-birth field and your validation regex assumes MM/DD/YYYY globally, you will silently reject valid German or Portuguese users.
IBAN and local ID formats need per-country validation, not one regex
The source article mentions IBAN formatting errors as a common failure point. Worth being specific about why: IBAN length and structure vary by country (DE: 22 characters, PT: 25, GB: 22 but different structure entirely), so a single regex pattern across all locales will either be too loose or reject valid input.
Don't write this by hand. Use a maintained library:
import { isValid, electronicFormat } from 'ibantools';
isValid(electronicFormat('DE89 3704 0044 0532 0130 00')); // true
isValid(electronicFormat('PT50 0002 0123 1234 5678 9015 4')); // true
ibantools handles per-country IBAN structure validation and MOD-97 checksum verification, which is the part most hand-rolled regexes skip entirely.
Same goes for national ID documents. The source article notes that "ID card" isn't a universal concept, a passport, a national ID (DNI, CNI, Personalausweis), and a residence permit are different documents with different field names. If your onboarding flow has a single idType enum with generic labels, that's a localisation debt item before it's a translation task.
ICU MessageFormat for pluralization and gendered strings
Simple key-value translation files ({ "balance": "Your balance is {amount}" }) break down fast once you hit pluralization rules, which differ wildly by language. English has two plural forms. Portuguese has two. Polish has four. Arabic has six.
ICU MessageFormat, supported by react-intl, i18next, and most serious i18n libraries, handles this correctly:
{
"transactionCount": "{count, plural, =0 {No transactions} one {# transaction} other {# transactions}}"
}
Writing your own plural logic per locale is a trap. Use a library that implements the Unicode CLDR plural rules instead of trying to special-case languages yourself.
Resource file hygiene: keep keys stable, never translate variables
The source article mentions preserving code variables and string placeholders across translation tiers. From the engineering side, this means your .json/.xliff/.strings files need discipline:
- Never interpolate raw strings into translation values that a translator could accidentally reorder (
"Hello {0}, you have {1} new messages"is fragile; named placeholders like"Hello {userName}, you have {messageCount} new messages"survive reordering during translation much better). - Keep keys semantic, not positional (
error.iban.invalid_format, noterror_12). - Never reuse a key for two different contexts even if the English string happens to match. "Cancel" the button and "Cancel" the subscription action may need different translations in other languages.
Pseudolocalization catches bugs before a translator ever sees the string
Before sending anything out for human review (which, per the source article, is where the real quality control budget should go for regulated content), run your UI through pseudolocalization. This expands every string by ~30-40% and wraps it in accented characters without actually translating it:
"Confirm payment" → "[Çönfîrm päymént ţéšţ]"
This surfaces two categories of bugs for free:
- Layout bugs: German and Finnish strings run 20-35% longer than English on average. If your confirm button truncates or overflows with a pseudolocalized string, it will truncate in production German too.
- Hardcoded string bugs: any text that doesn't get wrapped in brackets is a string that bypassed your i18n layer entirely, often a hardcoded error message buried in a catch block.
i18next has pseudolocalization plugins, and it's a cheap CI check to add before a release, not just a manual QA step.
Where this connects back to translation quality
None of the above replaces human linguistic review for consent text, regulatory disclosures, or KYC flows, that's a legal and compliance question, not an engineering one, and it's the part the source article covers well. But a lot of the "localisation bugs" that make it to production aren't translation failures at all. They're formatting, validation, and resource-management failures that happen before a translator ever opens the file.
If your pipeline passes clean, well-structured resource files with stable keys and correctly separated locale-sensitive logic, the human review step gets cheaper and more accurate. If it doesn't, no amount of linguist review fixes a hardcoded comma.
A practical starting checklist for a European fintech launch:
- Replace all manual number/date/currency formatting with
Intl - Validate IBAN and local ID formats per-country, using a library
- Migrate plain key-value translation strings to ICU MessageFormat
- Audit resource files for hardcoded strings outside the i18n layer
- Add pseudolocalization to CI before every release with new strings
Get these right and the translation review process becomes about terminology and tone, which is exactly where it should be spending its effort.
Top comments (0)