Custom AI translation software development: from e-commerce localisation to legal and medical translation

Anyone scaling into multiple EU markets quickly has thousands of product titles, legal clauses, package leaflets or support conversations to translate, at a quality that respects brand consistency, terminology and compliance. Generic translation buttons won't cut it. We build custom AI translation software that combines NMT engines, LLMs, terminology databases and human post-editing into a pipeline that fits your product catalogue, legal domain or medical record flow.

E-commerce localisation Legal translation Medical translation Live chat translation Document translation Meeting transcription
Discuss your translation case View applications
EN: Pre-order now NL: Nu vooraf bestellen DE: Jetzt vorbestellen FR: Precommandez ES: Reserva ahora QE 92 TBX OK Glossary COMET QE

Why your own AI translation pipeline delivers more than a button in your CMS

Generic translation plugins work for individual sentences, but they fall short as soon as terminology has to be consistent across thousands of SKUs, as soon as a legal contract can't afford to lose the difference between "agreement" and "contract", or as soon as a package leaflet has to match the registered product information word for word.

EU companies exporting to six or more markets face a continuous stream: new product pages on the webshop, contract amendments, policy documents, knowledge base articles, support tickets, customer conversations, annual reports. Manual translation doesn't scale, and pure machine output without review damages the brand. The gain lies in a hybrid pipeline in which neural machine translation (NMT) does the first 80%, a terminology database ensures product names and legal terms don't slip through, a quality estimation model flags which segments a human needs to see, and a translator with domain expertise reviews only those segments.

We build that pipeline around your existing systems: Shopify, Magento, Salesforce Commerce, Akeneo PIM, Contentful, a proprietary DMS or a legal contract management platform. The translation engine isn't the product; the integration with your content flow is the product. We design the architecture so that Dutch content flows to EN, DE, FR, ES, IT and beyond without editors or lawyers having to copy and paste manually with every release.

Application areas where AI translation software really makes a difference

From product content to legal clauses and medical records: every domain has its own quality criteria, its own terminology and its own risk profile. We tailor the pipeline architecture accordingly.

🛒

E-commerce product content i18n

Product titles, descriptions, variants, attribute values, SEO meta data and category copy in five to ten languages. Pipeline: PIM export to XLIFF, NMT translation with a brand glossary, automatic SEO keyword mapping per market, and publishing back to Akeneo, Magento or Shopify. Includes delta detection, so only changed segments are retranslated.

⚖️

Legal translation with terminology control

Contracts, NDAs, general terms and conditions, privacy statements and compliance documents. A TBX terminology database defines how legal concepts ("termination", "default", "force majeure", "penalty clause") are translated consistently across all languages. DNT tags protect party names, amounts and clause numbers from being altered by the model.

🏥

Medical translation and compliance

Package leaflets, IFUs, clinical protocols, patient forms and medical records. MDR- and EMA-compliant terminology via MedDRA and your own TBX. Mandatory review by a medically trained linguist for each segment, with an audit trail. Back-translation as an extra QA step where regulations require it.

💬

Live-chat translation in customer support

A Dutch support agent chatting in real time with a German, French or Spanish customer. Bidirectional translation with sub-second latency, glossary injection for product names and a correction loop when the QE score drops below a threshold. Integrates with Zendesk, Intercom, Salesforce Service Cloud or your own helpdesk.

🎙️

Meeting transcription and translation

Whisper for speech-to-text, followed by DeepL or Claude to translate the transcript into your participants' languages. Speaker labelling, automatic summaries per participant language and export to Notion, SharePoint or your own knowledge base. Useful for international teams, customer calls and compliance meetings.

📄

Document translation with layout preservation

PDF, Word and InDesign files that keep their formatting: columns, tables, footnotes, headers and footers, and images stay exactly where they are. Pipeline: extraction to XLIFF with segment IDs, translation, re-injection via a document engine and an automatic QA check for text overflow. No re-typesetting and no shifted layouts.

How Appfront builds an AI translation pipeline

Our opening question is never "which model should we use?" but "where does your content live today, where does it need to go and who reviews what?". The technical choice (NMT engine, LLM, post-editing tool) follows from the content flow and the quality requirements, not the other way round. A furniture brand's product catalogue can tolerate more machine output than the patient information leaflet for an infusion solution, so the pipeline choices differ fundamentally.

In practice, that means we first map your source and target systems, identify content types by quality class, and choose the right mix for each class of NMT (DeepL Pro, Google Translate Advanced, Microsoft Translator), LLM translation (OpenAI GPT-4o, Claude for stylistic translation), terminology enforcement (TBX import, Phrase TMS, MemoQ) and human post-editing (MTPE). Then we build the orchestration: webhooks from your CMS or PIM, queueing, retry logic, audit logging and dashboards where you can track turnaround times, machine cost and post-editor effort.

The end result is a system that works invisibly: an editor publishes Dutch content, a few minutes later the other language versions are ready for review, a post-editor handles the flagged segments and the site or document goes live. No more manual export-import cycles, no copy-pasting into spreadsheets, no scattered emails to a translation agency.

Architecture: NMT, LLMs and the orchestration in between

An AI translation pipeline is not a single model call. Quality comes from the combination of translation engine, terminology layer, quality estimation and review routing. These are the building blocks we use as standard and tailor to your situation.

NMT vs LLM-based translation

Neural machine translation (DeepL, Google, Microsoft) is fast, cheap and strong at sentence level, which makes it ideal for product content and support cases. LLMs such as GPT-4o and Claude excel at stylistic translation, marketing tone, idiomatic expressions and domain context. We combine both: NMT as the baseline, and an LLM for segments where tone of voice or cultural adaptation matters most.

Glossary, TM and TBX pipelines

A Translation Memory (TMX) reuses previously approved translations, so fuzzy matches save costs and ensure consistency. A TBX terminology database enforces that "warranty" always becomes "garantie", never "borg". We feed both into NMT engines via glossary APIs, or into LLM prompts via dynamic retrieval, so agreed terminology doesn't slip through the cracks.

Quality estimation with COMET

COMET (and in older setups BLEU/chrF) calculates a quality score per segment, without needing a reference translation. Segments above a threshold are published immediately; segments below it are routed to a post-editor. The result: less review time, predictable turnaround and measurable quality per content type and language.

DNT tags, back-translation and QA

Do-Not-Translate markers protect brand names, codes, legal article numbers and units of measurement. Back-translation (target back to source) is an extra QA step for medical and legal content, where differing interpretations are risky. Automated QA checks (length overflow, missing placeholders, number mismatches) run on every segment before publication.

From first audit to production pipeline

Our approach to AI translation projects runs in four phases. Each phase is clearly scoped, delivers something concrete and allows for adjustments along the way.

Content audit and language strategy

We take stock of content types, languages, volumes, quality requirements per type and existing TM/TBX assets. We map where your content lives today (CMS, PIM, DMS, helpdesk) and where it needs to go.

Engine selection and proof of concept

We benchmark 2 to 3 NMT and LLM combinations on your own content (anonymised), measure BLEU/COMET scores and post-editing effort, and choose the stack with the best balance of cost, speed and quality.

Integration and MTPE workflow

We build API integrations with your CMS, PIM or helpdesk, set up TM/TBX pipelines, integrate with CAT tools (Trados, MemoQ, Phrase) and set up post-editing queues for your in-house or external linguists.

Monitoring and further development

Dashboards for turnaround time, machine cost, post-edit distance and QE trends. Periodic retraining of glossaries, expansion to new languages and tightening of QA rules based on post-editor feedback.

Not sure about a large project yet?

Test your idea first: a working prototype in 1 day

With OneDayBuild, we make your idea tangible in a single day for €1,150, so you know whether further development is worth the investment. Decide to go ahead with the full build? Then we deduct the cost in full.

View OneDayBuild →

Technology we use for AI translation software

The stack depends on content type, languages and compliance requirements. For pure e-commerce localisation we often choose a DeepL Pro API baseline with a Phrase TMS layer around it. For legal and medical content we shift to a setup with stronger terminology safeguards and a bigger role for LLMs in stylistic and context-aware translation. For live chat, sub-second latency is a hard requirement, so we choose different models than for batch document translation.

Where possible, all components run on EU infrastructure: DeepL is hosted in Germany, Microsoft Translator offers EU data residency, and we route LLM calls via Azure OpenAI Service in West Europe or a Claude deployment within the EU where available. For sectors with the strictest requirements (healthcare, finance, government) we build an on-premise or private-cloud variant with self-hosted models (NLLB, Marian, Mistral fine-tunes).

DeepL Pro API Google Translate API Microsoft Translator API OpenAI GPT-4o Claude (Anthropic) Whisper COMET QE BLEU / chrF TBX TMX XLIFF 2.1 Trados MemoQ Phrase TMS NLLB Marian NMT Hugging Face LangChain Python / FastAPI PostgreSQL

Translation data and privacy: what you need to arrange

Translation content contains personal data and trade secrets, and in legal and medical contexts even special category personal data. We build pipelines that comply with GDPR, NIS2 and sector-specific frameworks.

GDPR-compliant processing chain

For every external translation engine (DeepL, OpenAI, Google), we put the right data processing agreement and data residency setting in place. Where possible, personal data is pseudonymised before sending, and we switch off logging on the provider side by default wherever the service allows it.

EU hosting and data residency

DeepL is EU-hosted by origin, Microsoft Translator offers EU-only routing, and Azure OpenAI Service runs in West Europe. For clients with strict requirements, we build an on-premise or private cloud variant with self-hosted models, where no segment ever reaches an external API.

Medical and legal compliance

For MDR package leaflets and registered product information, we follow EMA templates and QRD guidelines. For legal translations, we set up a review workflow with sign-off capability and a complete audit trail per segment. No anonymous machine output without a responsible human.

Audit trail and explainability

Every translation is traceable: which model produced which segment, which glossary entries were active, what QE score it received, and who did the post-editing. In the event of compliance questions or disputes, we can reproduce the complete chain of decisions for each segment.

Real-world scenarios we build

AI translation software only delivers value when it fits a workable flow. These are examples of pipelines we can deliver for different sector profiles.

Webshop with 8,000 SKUs in five languages

An EU retailer with a Magento or Shopify shop wants product content in NL, EN, DE, FR and ES without manual export cycles. Pipeline: a PIM webhook on every SKU change, NMT with a brand glossary, automatic SEO meta mapping, post-editing only for segments with a QE score below the threshold, and publishing back to the shop within the hour. Delta detection ensures that only changed fields go through the engine again.

Legal contract management for a law firm

An international practice translates NDAs, MSAs and shareholder agreements between NL, EN and DE. Pipeline with TBX terminology lock on legal terms, DNT tags on party names and amounts, and mandatory review by a bilingual lawyer. Audit trail per clause for liability purposes, and integration with a document management system such as HighQ or iManage.

Medical device: package leaflets in 24 EU languages

An MDR-regulated manufacturer needs IFUs in all EU languages for every product release. Pipeline with EMA template validation, MedDRA terminology via TBX, back-translation per market, and human review by a medically trained linguist. Audit trail in line with notified body requirements, and automatic version comparison when the text changes.

Live chat translation for SaaS support

A Dutch SaaS provider supports B2B customers in Germany, France and Spain. Implementation: a translation proxy for the helpdesk (Zendesk or Intercom) that translates incoming customer messages into Dutch and outgoing agent replies into the customer's language. Glossary with product and feature names, and a correction loop when the COMET score falls below the threshold.

Why Appfront for AI translation software

Pipeline thinking, not plugin thinking

We don't build a standalone translate button, but an end-to-end content flow from source CMS to target market. The translation engine is a component, not the product. That difference decides whether you're still having to step in manually a year from now.

Domain expertise per sector

E-commerce localisation calls for different choices than an MDR package leaflet or a law firm's contract. We make those choices explicit and tailor the architecture to your quality and compliance requirements, rather than offering a one-size-fits-all stack.

EU-first infrastructure

For customers in the EU, we arrange EU hosting, data residency, GDPR-compliant data processing agreements and, where needed, a private-cloud or on-premise variant as standard. No US cloud unless you explicitly choose it.

Frequently asked questions about AI translation software

What is the difference between NMT and LLM translation?
Neural machine translation (DeepL, Google, Microsoft) is optimised for sentence-level translation, and is fast and cost-effective. LLMs such as GPT-4o and Claude excel at stylistic translation, idioms, marketing tone and context spanning multiple sentences. For product catalogues and support cases, NMT is often sufficient; for marketing copy, legal context or cultural adaptation, we use LLMs. In practice, we combine both.
How is terminology kept consistent across thousands of products?
Through a centrally managed TBX (Term Base eXchange) terminology database. NMT engines such as DeepL Pro accept glossary uploads that enforce the same target term for a given source term every time. In LLM pipelines, we inject the relevant glossary entries into the prompt dynamically. A Translation Memory (TMX) also stores previously approved translations for reuse.
Can AI translation handle MDR package leaflets or medical records?
Yes, provided the pipeline architecture is right. For MDR- and EMA-regulated content, we use EMA templates, MedDRA terminology, mandatory review by a medically trained linguist, back-translation as a QA step and a full audit trail per segment. The machine does the heavy lifting; responsibility and final checks stay with a human.
What is COMET and why do you use it?
COMET (Crosslingual Optimized Metric for Evaluation of Translation) is a neural quality-estimation metric that predicts a quality score per segment without needing a reference translation. We use COMET to automatically determine which segments can be published straight away and which need to go to a post-editor. We sometimes use BLEU and chrF as a supplement for batch evaluation against a reference.
Does this work with Trados, MemoQ or Phrase?
Yes. CAT tools such as Trados Studio, MemoQ and Phrase TMS have mature APIs and support XLIFF, TMX and TBX. We build integrations in which segments automatically pass through the NMT or LLM pipeline, reappear in the CAT tool for post-editing, and are then published to your CMS, PIM or DMS.
What about data privacy with external translation engines?
For every external engine (DeepL, OpenAI, Microsoft, Google), we put the right data processing agreement in place, choose EU data residency where available and switch off provider-side logging where supported. For customers with strict requirements, we build an on-premise or private-cloud variant with self-hosted models such as NLLB or Marian, so that no segment ever leaves your own infrastructure.
Can AI translate live chat in real time?
Yes. With DeepL or Microsoft Translator, we achieve sub-second latency for sentence-level translation. The architecture: a translation proxy between customer and agent in the helpdesk, with glossary injection for product names, a correction loop for low QE scores and logging in the ticket history. It works with Zendesk, Intercom, Salesforce Service Cloud and custom helpdesk implementations.
What does an AI translation pipeline cost?
The investment depends on the number of languages, content volume, integrations and compliance requirements. An e-commerce pipeline with DeepL and a Shopify integration is a very different project from an MDR-compliant medical pipeline with self-hosted models and an audit trail. We usually start with a clearly scoped proof of concept to validate cost and quality before you commit to a larger project.
How long does it take for the pipeline to go live?
A first working pipeline for a clearly scoped content type is usually up and running within four to eight weeks. Expansion to more languages, content types or compliance requirements happens in iterations. We work in sprints, so you see tangible results along the way and can adjust the scope as you go.

Custom AI translation software for your EU markets?

Talk through your content flow with us. We'll map where AI translation adds the most value for your product catalogue, legal practice or medical pipeline — with no obligation.

Schedule a call

On applatenmaken.com, our platform on custom software development, you'll find a deeper dive into building translation software.

Edit content