Five years ago, connecting machine translation to a translator's editor was an optional extra that many professionals switched off. Today it is a standard feature, and clients increasingly ask whether their provider uses it. For translation agencies in the UK, the question is no longer whether machine translation belongs in the workflow, but how to use it responsibly.
The answer depends heavily on the software that hosts it. How machine suggestions appear, which engines are available and what happens to your data all shape the quality and safety of the result.
How MT fits into a CAT workflow
In a typical setup, the editor first looks for matches in the translation memory. If a high-quality match exists, it is offered first. If not, the editor asks a connected machine translation engine for a suggestion. The translator then edits that suggestion, or ignores it and translates from scratch. Approved results go into the memory, ready for reuse.
Memory always takes priority, because it contains human-approved translations specific to the client. Machine output fills gaps, never overrides approved content.
Engine quality varies by language and domain
No single engine is best for everything. Results differ by language pair, subject matter and even document type. An engine that performs well on English to French marketing copy may stumble on English to Polish legal text. When evaluating a cat tool for translation, check whether you can connect several engines and choose per project or per language.
Run your own comparisons. Take a few hundred segments of typical content, generate suggestions from each engine, and have experienced linguists rate them. Measure how much editing each engine's output needs. Real data beats vendor claims.
Privacy is the first question
UK clients, particularly in legal, financial and healthcare sectors, care deeply about where their text goes. Some engines retain submitted text or use it for training. Professional configurations use enterprise connections that do not store or reuse data, hosted in jurisdictions acceptable to the client.
Make sure engines can be disabled per client or per project. A confidential merger document should never reach an external engine just because a default setting was left on.
Post-editing levels
Clients need to understand what they are buying. Light post-editing produces understandable, accurate text without stylistic polish. Full post-editing aims for quality comparable to human translation. Translation without machine assistance remains the right choice for creative, sensitive or highly specialised content. A transparent provider agrees the level with the client in writing.
Measuring effort fairly
Good platforms record how much each machine suggestion was changed. This edit distance data shows which engines and language pairs genuinely save time and which only appear to. It also supports fair pricing for linguists, because heavily edited suggestions are effectively new translations and should be paid as such.
Terminology still matters
Machine engines do not automatically know a client's preferred terms. Some platforms apply termbase entries to machine output or let you train custom engines on approved memories. At minimum, terminology checks should flag where machine suggestions ignore approved terms, so post-editors can correct them consistently.
Training linguists to post-edit
Post-editing is a skill. Good translators sometimes over-edit, rewriting acceptable sentences for style, which wastes time on light post-editing jobs. Others under-edit, trusting fluent output that hides meaning errors. Clear guidelines, examples and feedback help linguists calibrate. Agencies that invest in this training get better results from the same engines.
What clients notice
Clients care about results: accuracy, consistency, turnaround and price. A well-run machine translation workflow can improve all four for suitable content. A poorly run one damages trust quickly, especially if a fluent but wrong sentence reaches a published document. Transparency about when and how machine translation is used protects the relationship.
Large language models in the editor
A newer development is the use of large language models alongside or instead of classic neural engines. Some platforms use them to rephrase suggestions, adjust tone, fix terminology or explain ambiguous source sentences. These features can be genuinely helpful, but the same questions apply: where is the data processed, is it stored, and how is output quality verified? Treat model-generated text exactly like any other machine suggestion, as a draft that a qualified linguist must review.
Certified work and machine assistance
For certified translations of official documents, many agencies avoid machine suggestions altogether. Personal data, the need for complete accuracy and the translator's personal certification make human translation the safer choice. Where machine assistance is used for internal drafts, the final certified version should always be checked line by line against the original by the certifying translator.
Questions to ask before you buy
Which engines are supported, and can you add your own? Is data retained by any engine? Can engines be restricted per client? Is edit distance recorded? Can terminology be enforced on machine output? Clear answers to these questions reveal far more than a feature checklist.
A balanced approach
Machine translation inside the editor is neither a threat nor a magic solution. It is a tool that works well for some content, poorly for other content and must always be supervised by qualified linguists. Choosing software that offers engine choice, strong privacy controls, edit tracking and terminology support gives UK agencies the flexibility to use it where it helps and switch it off where it does not.