Machine Translation Post Editing When It Actually Saves Money

Machine translation post editing has quietly become the default way large volumes of content get translated, and almost nobody buying it has a clear picture of what they are actually paying for. The pitch is simple enough. A machine produces a draft in seconds, a qualified linguist cleans it up, and the client pays less than they would for translation from scratch. That is true often enough to be worth doing. It is also false often enough to have burned a lot of budgets, and the difference between the two outcomes is almost entirely decided before a single word is translated.

Two very different jobs share one name

The industry splits post editing into light and full, and the distinction matters more than the labels suggest. Light post editing aims for accurate, comprehensible text and accepts that it will read like a machine wrote it. It suits internal documentation, support tickets, product reviews and anything with a short shelf life where the reader only needs to understand.

Full post editing aims for output indistinguishable from human translation. Terminology is corrected against a glossary, register is adjusted, sentence structure is rebuilt where the machine produced something technically correct and stylistically dead. It is the right choice for anything customer facing.

Problems start when a buyer prices for light and expects full. The linguist works to the brief they were given, the client reads the delivery and sees stilted prose, and both sides conclude the other did not understand the job. Agreeing the level in writing before work begins prevents most post editing disputes we have ever seen.

The quality of the source decides the quality of the saving

Machine output is only as good as what you feed it. Clean, consistent, well structured source text in a language pair with abundant training data produces drafts that need light touching. Marketing copy full of wordplay, badly written source, inconsistent terminology or a low resource language pair produces drafts that take longer to repair than to retranslate.

This is the calculation buyers routinely get wrong. Post editing is not universally cheaper. Below a certain draft quality threshold the economics invert, and any experienced provider will tell you so after seeing a sample. If nobody offers to run a sample before quoting, that is worth noticing.

There is a standard, and it is worth citing

Post editing is not an informal practice. It has a published international standard, ISO 18587, which sets out the competences a post editor must hold and what full post editing has to deliver. Asking a supplier whether they work to it is a fast way to separate a structured process from someone improvising with a browser tab open.

The academic and industry background is summarised well in the Wikipedia entry on postediting, including the research on why editing effort does not scale linearly with error count. A draft with a handful of subtle errors can take longer to fix than one that is obviously broken, because the editor has to read everything twice to trust it.

The workflow matters more than the engine

Buyers tend to ask which engine a provider uses. It is close to the least important question. What actually determines cost and consistency is how the engine sits inside the wider environment: whether approved terminology is enforced, whether previously translated segments are reused, and whether corrections feed back into anything.

Machine translation post-editing saves money only when the raw output is above a quality floor, and below that floor it costs more than starting fresh. That threshold logic is the practical question buried under most artificial intelligence news, where adoption stories rarely mention where the break-even actually falls. Every deployment has a line like this, and finding it is the real work.

That last point is where most of the long term saving lives. If your post editors fix the same mistranslated product name in every file for a year, you are paying for the same work repeatedly. Handled properly, corrections land in a CAT tools translation environment where they are stored and reused, so each round of content starts closer to finished than the last.

Translation memory does the heaviest lifting here. It is dull, unglamorous infrastructure, and it is the reason two companies with identical content volumes can end up with wildly different annual bills.

What the tooling actually looks like in practice

None of this happens in a word processor. Post editors work inside dedicated environments that segment the text, display the machine draft alongside memory matches and glossary hits, lock formatting and track how much of each segment was changed. That last metric is what allows honest pricing, because it measures real effort rather than assumed effort.

For a straightforward overview of that software category and what each type does, this guide to CAT tools is a useful primer before you sit down with a supplier. Understanding the vocabulary shifts the conversation from a debate about rates to a conversation about workflow, which is where the savings really are.

Deciding whether it fits your content

A reasonable rule: the more repetitive, technical and high volume your content is, the better post editing performs. Product data, manuals, knowledge base articles and specifications are ideal candidates. Brand campaigns, regulated documents, medical and legal material and anything where a single misreading carries real consequences are not, whatever the volume.

Most organisations end up with a split. They post edit the long tail and commission human translation for the material that carries risk or represents the brand. That mixed model is unglamorous but it is where the numbers work, and it is a far more durable answer than picking one approach and applying it to everything you publish.