# Where machine translation earns its place, and where it does not [Website page](https://verlineo.cc/insights/machine-translation-post-editing) ยท en This document reflects the public website. Starting rates are not a final quotation. Agreements for the individual project take precedence. We use machine translation daily. We also refuse it for roughly a third of the work that comes in. The line between those two is more specific than the industry usually admits. Publication date: 2026-07-14 Author: Anna Volkova Any bureau claiming it does not touch machine translation in 2026 is either misinformed or being careful with words. The useful question is not whether it is used, but where in the process it sits and who is accountable for what comes out. Our rule is straightforward: machine translation may produce a draft, never a delivery. What separates a good outcome from a poor one is almost entirely what happens after the draft exists. That sounds like a platitude until you price it. The draft is the cheap part of this trade and always was; the reading is the expensive part and always was. A technology that makes drafting nearly free does not make translation nearly free. It changes which half of the invoice is doing the work. ## Where it genuinely helps High-volume, repetitive, low-consequence text is where the technology is strongest. Internal knowledge bases, support tickets, product listings, historical archives being made searchable: material where the reader needs the gist quickly and an occasional infelicity costs nothing. It also helps on well-resourced European pairs with a mature translation memory behind them. When a client's memory already contains fifteen thousand approved segments, the engine is effectively completing patterns we established, and post-editing is fast and reliable. There is a third case that gets less attention: triage. A thousand pages of discovery material in a language nobody on the team reads is not a translation project yet. It is a sorting problem. A raw engine pass is the correct tool for deciding which forty pages are worth a translator's time, provided everyone understands that the output is an index and not a document. ## Where we decline it Anything that will be filed, signed, published or relied on. Regulatory submissions, contracts, patient-facing documentation, patent claims, marketing where register carries the value. The failure mode matters more than the failure rate. A machine translation engine is fluent when it is wrong, which is precisely the property you do not want in a document nobody will read against the source. A human mistranslation usually reads awkwardly at the point of error; a machine one reads perfectly well and quietly changes an obligation into a recommendation. Legal and pharmaceutical texts are also full of terms of art whose everyday meanings are strong attractors. Engines resolve ambiguity toward the common reading, which is the wrong direction almost every time in those domains. Two more properties matter in filings. Engines are inconsistent across a long document in a way a human is not: the same defined term can come back three ways in forty pages, because each sentence is translated as though it were the first. And they do not know what they do not know: there is no flag on the output saying this clause was hard. The uncertainty that a translator would have raised as a query simply does not survive into the draft. ## What post-editing actually is The industry's own vocabulary here is unhelpfully vague, so we are specific about it in quotations. Light post-editing means the output is made accurate and no more: no attention to register, no attempt at a natural sentence, no consistency pass. It is appropriate for material that will be read once by somebody who needs to know what it says. Full post-editing means the result is indistinguishable from a human translation, and the honest thing to say about it is that for demanding text it is not obviously cheaper. A post-editor has to read the source, read the draft, and hold both in mind: while resisting a fluent suggestion that is subtly wrong. On dense legal prose our editors frequently ask to discard the draft and translate, and we let them, at the rate we quoted. That last point is the one worth taking away. If a supplier offers full post-editing at a large discount on every text type, they are either not doing the second read or not charging for it, and you will find out which in the one document where it mattered. ## The questions nobody asks about the engine Where does the text go? A generic consumer endpoint may retain input for model improvement. For anything under an NDA that is not a technicality: it is disclosure to a third party, and it is the reason we run engines that contractually do not retain, and why some client material never leaves a translator's machine at all. What is it trained on? An engine tuned on a client's own approved memory behaves very differently from a general one. It is also the only version of this technology that gets better because of work you have already paid for, rather than in spite of it. And who signs the result? Certified and sworn translations carry a person's name and, in several jurisdictions, their legal liability. No engine can be a party to that. The signature is not a formality; it is the mechanism by which somebody is accountable, and it is the reason those documents are quoted as human work regardless of what the draft looked like. ## How we price the difference Post-edited machine translation is charged at a lower rate than full human translation, and we tell clients which one they are buying. What we do not do is quietly post-edit a draft and invoice it as translation: a practice that is more common in this industry than anyone would like to admit, and the reason some clients arrive here having lost confidence in their previous supplier. The quotation names the mode for each document rather than for the project, because a single tender pack routinely contains both: the technical annexes are a good candidate, the undertakings at the front are not. If post-editing is right for your material, we will propose it. If it is not, we will explain why in terms of your document rather than our process. > A machine translation engine is fluent when it is wrong, which is precisely the property you do not want in a document nobody will read against the source. ## Related pages - [Notes from the bureau.](https://verlineo.cc/knowledge/en/insights.md) - [Accountable for every line since 2024.](https://verlineo.cc/knowledge/en/about.md) [llms.txt](https://verlineo.cc/llms.txt)