Multilingual drafts. Already written in the reader's language, with your signature on them.
A German buyer's request, the technical sheet the distributor wants in English, the terms of sale to send to France. Each time someone stops, opens a translator and puts the text together by hand. This piece delivers the draft already written in the right language, and whoever signs it rereads it and sends it.
The draft arrives ready. Emails, sheets, price lists and instructions are prepared in the recipient's language, with the structure you use and your data inside.
The terms that matter aren't translated at random. Product names, legal wordings, Incoterms and payment terms sit in a glossary decided once, and every draft respects it.
No line goes out without a reread. Whoever signs reads and sends. Whatever commits the business with a price or a confirmation never goes out on its own.
This is the page for a single piece of the system. The other pieces, and the criterion for choosing which to start from, are on the services page.
What changes in practice
Drafting disappears from the day and the reread remains, which is the work only someone who knows the customer can do. On technical content the measured difference is large: twelve professional translators who started from a machine draft instead of a blank page produced on average 74% more words per hour.
The figure comes from a test published by Plitt and Masselot in the Prague Bulletin of Mathematical Linguistics, on 144,648 words of real content and four target languages. The individual gain ranged from 20% to 131%, so the number that counts is yours, not the average.
Time isn't lost translating. It is lost starting from scratch every time a request arrives in the wrong language.
The second change shows after a few weeks. Communications to the same market start to resemble each other: the same name for the same product, the same wording for the same condition, the same tone. Today that consistency depends on who was in the office that day.
Readers buy in their own language, and often only in that one
76% of consumers prefer to buy products with information in their own language and 40% never buy from a website in another language: this is measured by the Can’t Read, Won’t Buy survey by CSA Research, on 8,709 consumers in 29 countries.
In the same survey 66% say they use a machine translator anyway when the information isn't available in their language. The same goes for a buyer who receives a sheet only in Italian: they have a machine translate it, and your description reaches its destination in words you didn't choose.
It is a sample of private consumers in 29 countries, not of purchasing departments, and that has to be said. The direction, though, is the same one seen in every reply to a foreign importer: whoever receives a text in their own language replies, whoever receives a text to decipher puts it off.
For mandatory information the language is decided by the law of the country
In food it isn't a matter of courtesy. Article 15 of Regulation (EU) 1169/2011 puts it this way: «mandatory food information shall appear in a language easily understood by the consumers of the Member States where a food is marketed».
The same article adds that each Member State may impose on its own territory one or more official languages of the Union. It means the language of the label depends not on who sells but on the market where the product ends up, and it changes country by country: the full picture is in the guide to labelling by destination country.
That is why legal wordings are never rephrased. They go into the glossary in the exact form required, and the system copies them instead of translating them. Translating a mandatory wording well still means getting it wrong.
The glossary is decided once and applies to every document
During analysis a short list is written down: the trade names of the products, the legal names of the food, the Incoterms, the payment terms, the units of measure and the closing formulas you use. They are the words that mustn't change between an email and a technical sheet.
That list is imposed on every draft, in every language. If your product has a certain name on foreign markets, no draft translates it, and if a payment term is written in one form only, it always comes out in that form. The technical sheet in English and the covering email stop saying two different things.
The rest of the text, on the other hand, adapts to the reader. The same information is written differently for a German purchasing department and for a Spanish distributor, and that is the work the system does well when the boundary of fixed terms is clear.
What the system does, step by step
Every outgoing document follows the same route, whatever the target language, and every step leaves a log entry saying what happened and when.
| Step | What happens | What you get |
|---|---|---|
| Languageof the reader | The recipient's language is recognised from the incoming request or from the contact record, not asked each time. |
Nobody has to remember which language to write to that customer in, not even someone dealing with them for the first time. |
| Contentfrom your data | The draft is built on your real data, that is price list, availability, terms and texts already approved, not on what the model imagines. |
The numbers inside the email are yours, and there is nothing to recheck by hand line by line. |
| Glossaryimposed terms | The fixed terms are applied after drafting, so no linguistic choice can overwrite them. |
The product name and the legal wordings always come out identical, in every language and every document. |
| Rereadby a person | The draft reaches whoever signs ready, with what the system had to decide on its own highlighted. |
You read it and send it in a minute, instead of rewriting from scratch in half an hour. |
| Sendingand archive | The text sent is kept next to the contact, with the language and version used. |
Next time you start from what you have already said, and the customer doesn't get two different versions. |
The reread, which is the part that makes the rest sendable
Machine translation is now the norm even among professionals: in the European Language Industry Survey 2026, with 1,058 participants in 45 countries, 63% of independent translators work on texts already machine translated.
In the same survey only 23% of those professionals rate the quality of what comes out of the machines as high or very high, against 40% the year before. Those who use it every day rate it worse than those who try it, and that is why the reread isn't a precaution to drop once the system has matured.
Revising a machine-translated text has a dedicated standard, ISO 18587:2017, which describes the process and the skills of whoever carries it out. An industry that has written a standard for the person who rereads is saying something precise about how replaceable that person is.
The two things a machine gets wrong without noticing are the same everywhere: a technical term rendered with a plausible synonym, and a tone that in that language sounds like a different degree of distance. Neither shows when rereading in Italian, and both are spotted in three seconds by someone who knows the customer.
What goes out on its own and what waits for a person
Outside working hours a first-contact reply can go out on its own, in the language of whoever wrote, on the information you have already approved and nothing else. During working hours you decide, channel by channel, with a switch.
Whatever commits the business never goes out on its own: prices, quotes, discounts, order confirmations and contractual terms wait for a person in any case, whatever time the request arrives.
When the reply goes out automatically it says so in one line. From 2 August 2026 Article 50 of the AI Act applies, and the reader must know they are talking to a system before talking to it, not after. How we say it is on the AI transparency page, and the reasoning about the human constraint on the page about the principles we build with.
The same thing changes name with the trade
The mechanism is identical; the document that is rewritten by hand today isn't. It is worth looking at your own case, because that is where you see how much it costs every week.
| Sector | The document rewritten today | Where you see it |
|---|---|---|
| Food and agricultureand export | The reply to a request from a foreign distributor, with price list, technical sheets and terms, which today waits for the person who knows English. |
|
| Hospitalityaccommodation and events | Quote requests arriving in three different languages, and arrival instructions that are rewritten the same every time. |
|
| Restaurantsand bars | The menu and allergens for foreign guests, which change every season and have to be redone in every language. |
The service of a dining room where the phone rings unanswered |
What this piece doesn't do
It isn't a sworn or certified translation service. Where a sworn translation, a legalisation or the signature of a sworn translator is needed, that step stays with a qualified professional, and the system stops before it.
It doesn't take responsibility for what ends up on the label. It prepares the draft and holds the agreed wordings fixed, while the compliance of the product placed on the market stays with whoever places it.
It doesn't decide whom to write to and doesn't handle incoming requests. The first reply to whoever knocks is a different problem, handled by Inbox AI; the documents that come in and have to be entered into the management software are handled by automatic document entry.
Questions and answers
How many languages can it work in?
In the languages you actually sell in, and it is best to start with two or three. Every language added brings a glossary to fill in and someone who can read the result.
Opening eight on day one means having eight languages nobody checks. They are added when the customer exists, not when the idea of the customer exists.
Who guarantees the translation is right?
A person's reread, which remains mandatory before sending. In the European industry survey ELIS 2026, with 1,058 participants in 45 countries, only 23% of independent professionals rate the quality of machine translation as high or very high, against 40% the year before.
Revising a machine-translated text has a dedicated standard, ISO 18587:2017. The system removes the drafting work, not the responsibility of whoever signs.
Is it also for the label and for mandatory documents?
It prepares the draft and keeps the glossary aligned, but the responsibility stays with whoever places the product on the market. Regulation (EU) 1169/2011, Article 15, requires mandatory food information to appear in a language easily understood by consumers in the country where the food is marketed, and allows Member States to impose one or more official languages.
Legal wordings are never freely rephrased: they go into the glossary in their exact form and stay that way.
Does the recipient notice that a system wrote the draft?
If a reply goes out automatically outside hours, it says so in one line. From 2 August 2026 Article 50 of the AI Act applies, which requires a person to know they are interacting with a system before interacting with it, not after.
Drafts that a person rereads and sends are communications from the person who signs them, like a letter written with a spell checker in front of you.
How do you measure whether it's working?
With the preparation time of a multilingual communication, measured first on a real batch and compared again afterwards. The second number is the share of drafts that a person rewrites instead of just correcting: if it grows, the glossary is incomplete or the language was opened too early.
The third, where it makes sense, is the time between a foreign buyer's request and the reply going out.
Notes on sources
- The language requirement for mandatory food information is in Article 15 of Regulation (EU) 1169/2011, consolidated text on EUR-Lex. It is a rule, so it has no sample: it applies to every food placed on the Union market.
- The average gain of 74% in words per hour, with individual variation from 20% to 131%, comes from Mirko Plitt and François Masselot, A Productivity Test of Statistical Machine Translation Post-Editing in a Typical Localisation Context, The Prague Bulletin of Mathematical Linguistics 93, 2010, pages 7-16. It is a test on twelve professional translators, on 144,648 words of software content from a single company, with 2010 technology: we cite it for the order of magnitude and for the variation between people, not as a promise about your case.
- The 76%, 40% and 66% are from CSA Research, Can’t Read, Won’t Buy. The sample is 8,709 private consumers in 29 countries, so it measures retail buying behaviour and not purchasing departments: the direction is useful, the number doesn't carry over as it is to a negotiation between businesses.
- The 63% of independent translators working on pre-translated texts and the 23% who rate machine quality as high are in the European Language Industry Survey 2026 report, pages 31-33. The respondents are 1,058 professionals and companies in the language industry in 45 countries, that is people who sell translation: it is an interested sample, and we say so.
- The ISO 18587:2017 standard defines the process and skills for full human post-editing of a machine-translated text.
- This page doesn't report results obtained for a client, because this piece hasn't yet been delivered to a client. The tests cited are functional checks carried out in testing.
The other pieces in this group
Requests and contacts lost between channels and languagesFifteen minutes, with your case in front of us.
How many requests in a foreign language arrive in a month, and how long before the replies go out? If the answer is that nobody knows, that is already the most useful information, and it can be counted in thirty days. In fifteen minutes on the phone we look at it together and tell you where it makes sense to start, even if we don't end up working together.
You get Mattia Esposito, who then builds the system: there's no salesperson in between. If you'd rather measure on your own before talking, the Diagnostico (in Italian) is twenty questions and five minutes.