Most fintech teams start with one question: should we use human translation or AI? That is the wrong place to start, because your content is not one thing. A money-transfer app has system messages that repeat thousands of times with a different amount in each. It has help articles that change every week. It has marketing pages that need to convince people. And it has terms and conditions that the legal team wrote word by word, on purpose.
If you treat all of that the same way, you pay twice. Use professional translators for everything, and you pay premium rates for messages nobody reads closely. Use machine translation for everything, and you pay later: in support tickets, lost customers and, for regulated content, legal risk.
A better question is this: for this piece of content, in this language, who checked the text before it went live?
Four service levels, each for a different job
Translation services come with a lot of acronyms: MT, MTPE, APE, LQA. Behind them are four service levels. None of them is better than the others in general. Each one fits a different kind of content, and most products use all four.
The real difference between them is who checks the translated text before it goes live. For regulated content, that is what should drive your choice.
AI Only
Machine translation that goes live without anyone reading it. Use it for content that only your colleagues will read: an internal knowledge base, an engineering wiki, internal process notes.
Do not use it for anything a customer sees. Errors, wrong terms and awkward phrasing all go live, and nobody notices.
What's inside this level
There are two versions. Raw AI is plain machine output. AI with automated post-editing adds a second AI pass that fixes grammar mistakes, mistranslations and missing words. It can make a large, low-risk body of text more readable, but the quality is uneven. In both versions, nobody has read the result. So neither belongs anywhere a customer or a regulator will look.
AI + Human
In the localization industry, this level is also called MTPE, short for machine translation post-editing. You will see that name in vendor proposals and quotes. It means the same thing.
AI translates first. Then a professional translator, a native speaker of the target language, reviews and corrects the text. This is the level most fintech products will use most. It works well for large amounts of content that keep changing and follow a pattern: the product interface, transaction emails, help articles, release notes. It usually costs 30–50% less than full human translation.
The quality is good: correct and natural, though not always elegant. That is exactly what you want for content that needs to be clear rather than persuasive.
What's inside this level
Which AI model. Models vary a lot by language pair and content type. A model that handles Spanish marketing text well may fail on Japanese legal text. A vendor who tests several models on your content before choosing one gets a better first draft. Less correction means lower cost.
How much correction. The translator can do a light pass (fix mistakes and terminology) or a full pass (also fix tone and flow). Choosing the depth by content type is one of the easiest ways to save money without moving content to a riskier level.
Publish first, correct later. For low-risk content, you can publish the AI translation right away and have a translator review it within two or three days. This works for help articles and release notes. Never do this with legal text, payment screens or major marketing content.
The standard. ISO 18587 is the international standard for post-editing of machine translation. Ask whether your vendor has it.
Human Only
Native-speaking translators translate from scratch and adapt the text for the local audience. This is what brand and sales content needs. Machines do not handle tone, humor or brand voice well. Translators do. Use this level for slogans, campaigns, your main landing pages and anything that carries your brand voice. High-volume marketing pages, such as app store descriptions or product feature pages, can often run on AI + Human with a full correction pass instead.
The rate is standard, but you get up to 70% off on text your translation memory has seen before. So the more you translate, the cheaper repeated text becomes.
What's inside this level
Some content needs more than translation. Jokes, wordplay, idioms and slogans often cannot be translated and have to be rewritten. Neither AI nor a literal human translation will save them. Plan for native-speaking copywriters on this part, and tell them how much freedom they have.
Human + LQA
Full human translation, followed by a quality check in the product itself. LQA stands for linguistic quality assurance: a native speaker looks at the translated text where it actually appears, in a test version of your app, not in a spreadsheet. This catches problems you cannot see on paper: text that gets cut off, layouts that break, a correct word that is wrong for that screen, a fraud warning where the word "not" was pushed out of view.
Use it for critical content: legal text, payment and checkout screens, onboarding. You pay for the translation plus an hourly rate for testing. Your team needs to provide a test build where testers can see the translations in place.
What sits on top of this level in fintech
For text with legal weight (disclosures, terms, pre-contract information), Human + LQA is the highest service level. But it is not the last step. Someone from legal or compliance in that market needs to sign off the translated text. That step is yours, not the vendor's. Make it a mandatory gate in your workflow for critical content, and keep a record. If a regulator asks how you made sure a disclosure was correct in Polish, "we had a very good translator" is not an answer.
The four questions that decide the level
Three of these apply to any product. The fourth is what makes fintech different.
1. What happens if it is wrong? A clumsy sentence in an internal document costs nothing. The same sentence on the signup screen costs you customers.
2. Is the text repetitive or creative? Machine translation is good at system messages that follow a pattern. It is bad at text meant to persuade, entertain or carry your brand's voice.
3. How much is there, and how often does it change? Ten thousand words once is one problem. Eight hundred words every Tuesday for two years is a different one. The second case is where AI + Human and a good translation memory pay off.
4. Does the text have legal weight? In most products, a bad translation is a user experience problem. In fintech, some text is the legal disclosure itself. It is what a regulator or a court reads to decide whether you informed the customer properly.
Two examples of text with legal weight
Under the EU's PRIIPs rules, the Key Information Document for an investment product must be written in an official language of the country where it is sold, or in a language accepted by that country's regulator, or translated into one of them. The regulation states that the translation must faithfully and accurately reflect the content of the original.1
Under MiCA, the EU crypto regulation, a crypto-asset white paper must be drawn up in an official language of the home country and of any country where the asset is offered, or in a language customary in international finance, which today means English. The people behind the offer are liable to holders for information in the white paper that is not complete, fair or clear, or that is misleading. The regulation names translations directly: a summary, "including any translation thereof", can trigger that liability if it is misleading, inaccurate or inconsistent with the rest of the document.2
For content like this, a named person has to be accountable for the translation.
A content map for a fintech product
This is how most fintech content inventories look once you sort them by risk instead of by team. It builds on the general map we use for software products and adds the content types that fintech has. Treat it as a starting point. Your own risk assessment and your legal team's view come first.
| Content type | Risk | Recommended service |
|---|---|---|
| Terms, privacy policy, regulatory disclosures, key information documents, white papers | Critical | Human + LQA + legal sign-off in the target language |
| Onboarding, identity checks (KYC) | Critical | Human + LQA |
| Payment, transfer, top-up and checkout screens | Critical | Human + LQA |
| Fee schedules and rate disclosures | Critical | Human + LQA + legal sign-off in the target language |
| Product interface (main screens) | High | AI + Human or Human + LQA for your biggest markets |
| Security and fraud alerts | Critical | Human + LQA |
| Brand and campaign copy: slogans, taglines, main landing pages | High | Human Only |
| High-volume marketing: app store descriptions, product and feature pages, ad variants | High | AI + Human, full correction pass, or Human Only |
| Error and decline messages | Medium–high | AI + Human, full correction pass |
| Transaction emails and push notifications | Medium | AI + Human |
| Help center, FAQs, support templates | Medium | AI + Human |
| Release notes and changelogs | Low–med | AI + Human, publish first and correct later is fine |
| Blog and educational content | Low–med | AI Only or AI + Human |
| Internal knowledge base | Low | AI Only |
Two things to keep in mind. First, the same content type can sit at different levels in different languages. Your two biggest markets may justify Human + LQA for the interface, while smaller markets run on AI + Human. Second, "transaction emails: AI + Human" assumes the email is a standard one. If the email announces a fee change, a new interest rate or a complaint deadline, it is a regulatory notice and belongs in the critical row, whatever system sends it. The level depends on the content, whatever channel delivers it. The same logic splits marketing in two: text that has to persuade or carry your voice goes to translators, while pages that mostly inform, such as app store descriptions, can start with AI as long as a translator does a full pass afterwards.
How to decide without guessing
Test on your own content. Take a real sample, not a marketing paragraph you happen to like, and have it translated at two levels. Comparing AI + Human with Human Only on your actual text settles arguments that otherwise go on for months. Send the glossary, the style guide and reference material along with the test. A vendor who follows your instructions and asks questions is showing you how they will work later.
Ask which AI models the vendor tested for your languages. "Which models did you try on our content, and how did they do?" tells you whether the vendor tunes their setup for each client or uses one engine for everyone. You can check some of this yourself. Evaluation tools based on large language models score grammar, vocabulary and style, and explain the errors they find. That gives a team without native speakers a first opinion.
Change levels deliberately. Start cautious with anything customers see. Measure. Then relax the level where the data supports it. Moving content down a level should be a recorded decision with a named owner.
What a mixed setup looks like in practice
Most teams end up using several levels at once. Here is an example from our own work. It comes from a different industry, but the decision is the same one a fintech team would face. Our client Vizor had all eleven of its languages translated by native-speaking translators and wanted to cut costs by 30% without losing quality. It kept full human translation for German and Japanese, its most valuable markets, and moved the other nine languages to AI + Human. Six months later, monthly localization costs were down 35–50%, and the apps kept top ratings for localization quality on the App Store and Google Play.
The logic transfers even where the exact split does not: protect the markets and content where quality brings in money, and automate the rest.
Where AI should not go, even when the budget is tight
- Anything with legal weight. Disclosures, terms, pre-contract information, regulated marketing claims. The translated text is the one that binds you.
- The screens where people sign up and pay. Registration, identity checks, adding money, checkout. Mistakes here cost more per word than anywhere else in the product.
- Brand and campaign copy. Jokes, wordplay, idioms and slogans need a native-speaking copywriter to adapt or rewrite them.
- Any language nobody on your side can read. If no one in the process reads the language and no testing is planned, AI Only becomes a risk you cannot measure.
Regulatory references
The two regulatory points in this article refer to the following texts. They are summarised for orientation and are not legal advice.
- Regulation (EU) No 1286/2014 (PRIIPs), Article 7(1). Full text on EUR-Lex.
- Regulation (EU) 2023/1114 (MiCA), Recital 25 and Article 6 (language of the white paper), Article 15(1) and 15(5) (liability for white paper information, including translated summaries). Full text on EUR-Lex.
AI Video Studio