Machine translation & AI quality
How this is scored
What the machine translation does and how well it is controlled: the languages and language pairs covered, the engine (own neural or LLM model, or a choice of third-party engines such as DeepL, Google, Microsoft or an LLM), glossaries and formality applied during translation, document translation that keeps the layout (Word, PowerPoint, PDF, InDesign), custom or adaptive models trained on the customer's own translations, automatic quality estimation or scoring, and published evaluations of translation quality.
0 — No machine translation of any kind, and no way to connect an engine.
3 — Machine translation is offered or one engine can be connected; languages, glossary use and quality controls are basic or not described.
5 — Machine translation for the major European and world languages with the list published, glossaries applied during translation, a formality or tone setting, document translation that keeps formatting for common office formats, and either an own engine or a choice of several named engines.
8 — Custom or adaptive models that learn from the customer's translation memory or corrections, automatic quality estimation per segment with a threshold that routes weak segments to human review, LLM-based translation with style guides and context, engine selection per language pair stated, and a published quality evaluation with its method.
10 — Translation quality that an organisation can measure and steer: evaluations against named benchmarks or blind human review published with their method, quality estimation calibrated and explained, custom models with their training data and isolation documented, and every machine-translated segment traceable to the engine, model version and glossary that produced it.
The Skeptic
Twenty-plus engines are named — OpenAI, Anthropic Claude, Gemini via Vertex, DeepL, Amazon Bedrock — with Auto Select routing per language pair and content type, RAG prompts built from the customer's own translation memory and glossary, custom-trained engines, and a Language Quality Estimation using a fine-tuned XLM-R model or LLM assessment that routes work to experienced linguists or past human review, plus a published MQM score of 98-plus for its AI human translation against a stated industry benchmark. That covers most of the top tier. We found no public information on a published language list, document translation that keeps layout, or per-segment traceability to the engine, model version and glossary that produced a translation, so it stops one rung short. 3 9 10 14 17