Add text to get the translation that 22 AI models agree on
What businesses buy after they already have an AI translation
In the four weeks to November 4, a hospitality company sent us staff training that an AI tool had already translated into Haitian Creole and asked us to edit it. A health company sent patient handouts that already existed in Spanish and asked for a review with a certificate of accuracy. An app publisher sent interface text that machine translation had already covered in 16 languages and asked for native review.
The conclusion I draw is that AI translation changed the translator's job in these requests, from writing a text to checking a text that already exists. Three translation requests from business leads between October 8 and November 4, 2026 asked for exactly that, and two of the three placed an order.
Nimdzi's 2026 report describes the same movement across the industry. Buyers who sent content once marked not for machine translation straight through raw AI output met quality failures and returned to providers for expert-in-the-loop validation.
This article walks through the three requests, what a check involves, and the question they raise for Tomedes about human review by default.
Key takeaways
- Three translation requests from Tomedes business leads between October 8 and November 4, 2026 asked for a check of a translation the business already had, and two of the three ordered.
- The three requests were an edit of AI-translated translations into a lower-resource language, a review of existing handouts with a certificate of accuracy, and native review of machine-translated app text in 16 languages.
- The checks sit where an error costs more or is harder to see, and three requests are too few to call that a measurement.
- Tomedes' guidance is that a translation you only need to read can stay as AI output, and a translation others will rely on should get AI post-editing.
- Tomedes launched its AI translation track without human review by default and said it would add one if buyers on that track consistently wanted it, so these requests are an early signal, and the evidence Tomedes named is engagement data from the track.
What do businesses ask for when they already have an AI translation?
They ask for a check of the text they have. Of all the translation requests from business leads between October 8 and November 4, 2026, three asked for an edit of AI output, a review with a certificate of accuracy, or native review of machine-translated text, and two of those three ordered.
The first was a hospitality company's staff training, covering posters, an employee guide, and orientation slides, which an AI tool had already translated into Haitian Creole. The company asked for the translations to be edited, and it ordered.
The second was a health company's patient handouts, already available in Spanish, which the company wanted reviewed with a certificate of accuracy. It ordered the review.
The third was an app publisher's interface text, which machine translation had already covered in 16 languages. The publisher asked for native review and has not ordered yet.
What I take from the three requests is that the AI translation was where each project started.
Why do the checks cluster where they do?
In these three requests, the checks sit where an error costs more or is harder to see, which covers staff training, patient information, a lower-resource language, and interface text across 16 languages. Three requests are too few to measure anything, and I am cautious about reading more into them.
Research and reporting support the instinct on the language point. A November 2026 article in The Conversation, written by a professional who reviews AI translations, describes hidden errors such as a wrong tense when translating Chinese into English, and warns that the problem could be more serious for less common languages.
A 2025 study in npj Digital Medicine found that ChatGPT-4o alone performed worst for Armenian and Somali when translating patient discharge instructions. Haitian Creole was one of the three requests, and a test would be needed before claiming it behaves like those languages.
A fluent AI translation gives a reader without the language no way to see the error. That is the practical reason a business with training or patient material in a smaller language asks a person to check it.
What does a check of an AI translation involve?
A check corrects meaning errors against the source, fixes terminology and tone for the audience, and leaves a record of who did it. Tomedes' AI post-editing assigns a domain-matched human post-editor under ISO 18587:2017, and that post-editor is accountable for the final text.
The three requests asked for three different checks. The table shows what each business had, what it asked for, and whether it ordered.
| What the business had | What it asked for | Ordered, as of November 5 |
|---|---|---|
| Staff training already translated into Haitian Creole with AI translation | An edit of the AI translations | Yes |
| Patient handouts already available in Spanish | A review with a certificate of accuracy | Yes |
| App text machine-translated into 16 languages | Native review in each language | Not yet |
An edit treats the AI text as the base and corrects it. A review compares an existing translation with the source, and native review puts a speaker of each target language in front of the text.
For a business choosing between them, how Tomedes combines AI translation with human review is covered in a separate guide.
Do all AI translations need a human check?
No, because the right check depends on who will rely on the text. Tomedes' guidance is that if you only need to read a translation, a free AI translator is enough, and if clients or colleagues will rely on it, AI post-editing is the better route.
Reading copies, drafts, and internal triage usually do not need a check. Training, patient information, contracts, published text, and anything a person signs for usually do.
For highly sensitive content, we recommend full human translation by a vetted translator under a signed NDA, which is a different route from checking an AI translation.
Should Tomedes add human review to AI translation by default?
Not yet, because Tomedes said it would add human review by default only if buyers on the AI translation track consistently wanted it, and these requests do not show which tool made the translations.
Tomedes launched that track without a human review step in the default workflow. The lead messages do not say which tool produced the AI translations, so these requests cannot tell us how buyers on the track behave.
These requests are an early signal about how businesses treat AI translation after the fact, and engagement data from the track itself would be a separate dataset.
Where I land is that the default stays as it is until the track's own data says otherwise, and that these requests are a reason to look at that data sooner.
What should a business do with an AI translation it already has?
Sort the text by who will rely on it, check the parts where an error costs something, and ask the reviewer to name who signs for the result. Start with anything that reaches patients, staff, customers, or a court, and leave internal reading copies as they are.
- List what the translation will be used for and who will rely on it.
- Mark the parts where an error would reach patients, staff, customers, or an authority.
- Ask for an edit, a review, or native review of those parts, depending on how the translation was made.
- Ask who reviews, who is accountable, and whether the review is documented.
Tomedes' human translation service and its AI post-editing handle these checks, and a business can ask for either.
What I recommend is that a business with an AI translation in hand decide who relies on it before it decides who checks it. Another four weeks of leads will show whether three requests were a blip or a pattern.
FAQs
Q: Can a translator edit an AI translation instead of redoing it?
A: Yes, because an edit starts from the AI text and corrects meaning, terminology, and tone against the source. One business asked for exactly this for AI-translated staff training in Haitian Creole and placed an order. If the AI text is too inaccurate, a fresh translation can be the better route.
Q: Can an existing translation be reviewed and given a certificate of accuracy?
A: Yes, businesses ask for this. One health company requested a review of existing Spanish patient handouts with a certificate of accuracy and placed an order. Whether a certificate is appropriate depends on who will receive the document, so confirm that with the receiving body first.
Q: Do machine-translated app strings need native review?
A: It depends on how many people will see them and what an error costs, because machine translation can be fluent and still wrong for a product. One app publisher asked for native review of strings machine-translated into 16 languages, which means a reviewer for each language.
Written by Ofer Tirosh
Connect on LinkedInOfer founded Tomedes in 2007 and now leads the company's push to combine two decades of human translation expertise with AI. He writes about where the language industry is actually heading and the shifts nobody's ready for, the bets that paid off, and the ones that didn't.
Share:
Post your Comment