# A Practical Guide to AI Translation for Business We manage content in four languages. Two years ago, professional translation for our website ran €3,200/month. Now it's €640/month — same quality, same languages, five times faster turnaround. AI translation with targeted human review changed the economics completely. Here's how to implement it without the quality disasters that give machine translation a bad name. ## The Quality Threshold Crossed Modern AI translation models — particularly those from DeepL, Google, and large language models — produce output that's significantly better than what we had even three years ago. The improvements aren't incremental; they're transformational. For standard business content, AI translation now achieves what linguists call "publishable quality" roughly 80% of the time. That means four out of five paragraphs need no human editing. The remaining 20% typically needs light editing — word choice refinements, cultural adaptations, and terminology alignment. This is fundamentally different from the Google Translate of 2015 that produced grammatically correct but awkward-sounding translations. ## Where AI Translation Excels **Product descriptions.** Structured, factual content with consistent terminology translates exceptionally well. A product page with specifications, features, and standard descriptions often needs zero human editing. **Support documentation.** Help articles, FAQs, and troubleshooting guides follow patterns that AI handles reliably. The language is straightforward, the structure is predictable, and the terminology is consistent. **Blog content.** Informational blog posts translate well because they use standard vocabulary and follow conventional article structure. Opinion pieces and humor are harder. **UI text.** Buttons, labels, error messages, and navigation elements are short, context-clear, and highly patternable. AI translations for UI text reach 90%+ accuracy. ## Where AI Translation Struggles **Marketing copy.** Wordplay, cultural references, emotional appeals, and brand-specific language confuse AI models. A tagline like "Think Different" loses its punch when translated literally. **Legal content.** Contracts, terms of service, and compliance documents require precision that AI can't guarantee. A mistranslation in a legal document creates liability. **Highly technical content.** Industry-specific jargon, acronyms, and domain terminology can trip up AI models that haven't been trained on your specific vocabulary. **Culturally sensitive content.** Content that references local customs, idioms, or social norms needs cultural adaptation, not just linguistic translation. ## The Three-Tier Translation Strategy Not all content needs the same translation quality. Implement three tiers: ### Tier 1: AI Only (50-60% of content) Internal documentation, draft communications, content syndication for SEO purposes, and low-stakes informational pages. AI translates; no human review. Cost: approximately €0.01/word. ### Tier 2: AI + Light Review (30-40% of content) Blog posts, product descriptions, support articles, and standard business pages. AI translates; a native speaker does a quick review focusing on accuracy and naturalness. Cost: approximately €0.03/word (including reviewer time). ### Tier 3: Human Translation (5-10% of content) Marketing campaigns, legal documents, executive communications, and brand-critical content. Professional translators create original translations, using AI output as a reference. Cost: approximately €0.10-0.15/word. ## Setting Up AI Translation in Your Workflow ### Step 1: Build a Translation Memory Before you start, compile a glossary of terms, brand names, and product names that should never be translated or should be translated in specific ways. This glossary feeds the AI model, dramatically improving consistency. For example: "Lucky Desk" should never be translated. "Workspace" should be translated as "Werkruimte" in Dutch, not "Werkplek" or "Ruimte." ### Step 2: Establish Quality Metrics Measure translation quality using these metrics: - **Fluency:** Does the translated text read naturally in the target language? (Scale: 1-5) - **Accuracy:** Does the translation convey the same meaning as the source? (Scale: 1-5) - **Terminology:** Are brand terms and industry jargon translated consistently? (Yes/No per term) - **Post-edit distance:** What percentage of the AI output was modified by reviewers? (Target: under 20%) ### Step 3: Choose the Right AI Translation Service **For maximum quality:** DeepL consistently outperforms alternatives for European languages. Google Translate is strong for Asian languages. **For platform integration:** Many business platforms include built-in AI translation that maintains context across your content library. This is preferred because translations happen within the system — no copy-paste, no format loss. **For specialized content:** Large language models (GPT-4, Claude) can be prompted with glossaries and style guides for domain-specific translations. More expensive per word but higher quality for specialized content. ### Step 4: Implement a Review Workflow For Tier 2 content, set up a review workflow: 1. Content is created in the source language 2. AI translation runs automatically 3. Translated content enters a review queue 4. Native speaker reviews and edits (typically 5-10 minutes per page) 5. Approved content is published The key metric: time per page in review. If your reviewer consistently spends more than 15 minutes per page, the AI quality needs improvement (better glossary, different model, or reclassify the content as Tier 3). ## Cost Impact: A Real Example For a business with 200 pages of content in 4 languages: **Before AI translation:** - 200 pages × 3 target languages × 500 words × €0.12/word = €36,000/year - Turnaround: 2-3 weeks per batch **After AI translation with tiered approach:** - Tier 1 (120 pages): 120 × 3 × 500 × €0.01 = €1,800/year - Tier 2 (70 pages): 70 × 3 × 500 × €0.03 = €3,150/year - Tier 3 (10 pages): 10 × 3 × 500 × €0.12 = €1,800/year - Total: €6,750/year - Turnaround: 1-2 days for most content Annual savings: €29,250 — an 81% reduction. ## The Privacy Consideration When using cloud-based AI translation, your content passes through third-party servers. For most business content, this is acceptable. For sensitive data — financial reports, personnel documents, trade secrets — consider self-hosted translation models. Self-hosted options like LibreTranslate or Argos Translate run entirely on your infrastructure. Quality is lower than cloud services, but data never leaves your network. The best approach: use cloud AI for non-sensitive content and self-hosted or human translation for anything confidential. ## Getting Started Pick one content type — your blog or product descriptions — and translate it with AI into your primary secondary language. Have a native speaker review it. Measure the post-edit distance. If it's under 25%, you're ready to scale. The technology is past the point of experimentation. It's a production tool with real cost savings and genuine quality. The question isn't whether to use AI translation — it's how to implement it thoughtfully.