Translating a contract with AI: where it stops

A company that sells abroad always ends up translating legal texts: terms of sale, distribution agreements, privacy policy, legal notices, employment contracts for a subsidiary. It is tempting to run them through the same tool as the website pages. For part of the work, that is a good idea. For the rest, it is where a mistake costs the most.

This article draws the line, from the point of view of the company commissioning the translation. It is not legal advice: that is precisely one of its messages. It extends our articles on website translation.

What AI does well on a legal text

Three uses deliver, provided you know what happens next.

Understanding a text you received. A partner sends a contract in German or French: a machine translation gives you the general meaning, the obligations, the deadlines and the penalties within minutes. It is a reading aid, not a basis for signing.

Preparing a first draft. Terms already approved in one language are machine-translated, with a brand glossary that fixes the company’s terms and a translation memory that reuses clauses already approved. The lawyer or translator starts from this draft rather than a blank page: their time goes to what requires their expertise.

Comparing two versions. When a clause changes in one language, an AI model spots where the other version no longer says the same thing. The finding still needs checking, but the search is done.

The best models are also improving on this ground. The annual machine translation evaluation, WMT25, tested a legal domain (English to Hindi): the most advanced models do better than the others there, and the authors stress the need for greater robustness in critical applications. The same conclusions note persistent weaknesses in domain terminology. That is exactly where a contract lives.

Why a legal text is not like other texts

A badly translated offer page loses a sale. A badly translated clause has an effect: a broader warranty than intended, a deadline that does not run, a penalty you cannot claim.

Legal words do not always have an equivalent. Each legal system has its own concepts. English contract law’s “consideration” has no exact equivalent in French law, and the French “fonds de commerce” has no English translation that carries the same rules. A word-for-word translation produces a correct sentence, but a concept the country’s law does not know, or knows with different effects.

The law sometimes imposes the language. In France, for example, the law of 4 August 1994 makes French mandatory in the offer, presentation, instructions for use and warranty of goods and services sold to consumers. The employment contract is the clearest example:

The same article lets a foreign employee ask for a translation into their own language, and provides that if the versions differ, only the version in the employee’s language can be relied on against them. A rough translation is therefore not a mere matter of style: it is the version that counts against the employer. Other countries have their own language rules; the lawyer from that country knows them.

A plausible sentence is not a correct sentence. A generative model writes fluent text, even when it is wrong. On a product page, a mistake is often visible. On a liability clause, only someone who knows the country’s law sees that the sentence no longer says what the original says.

Where AI stops: three cases that stay with professionals

The binding version. A contract in two languages usually states which one prevails in case of disagreement. That one is drafted or reviewed in full by a lawyer who practises the applicable law, whatever tool produced the first draft. The other versions remain translations, and the contract should say so.

Concepts with no equivalent. Adapting a clause to another country’s law is no longer translation, it is legal drafting. That is the lawyer’s job, helped by a specialised translator when the language is not their own.

Official documents. Company records, articles of association or a judgment to be filed with a foreign authority often require a translation by a certified or sworn translator. In France, Service-Public.fr explains, this is a court expert registered on the lists of the courts of appeal and the Court of Cassation. No tool replaces their signature.

How to split the work between the machine and professionals

The right split saves everyone time without moving the risk.

Three rules make this workflow reliable.

Start from a stable source text. Terms still under negotiation are not translated: every change restarts the review in every language.

Keep track of versions. The translation memory keeps approved clauses: at the next update, only the changed clauses go back out for translation and review. That is where the machine really brings the bill down.

Do not mix pages. A website’s terms and privacy policy are treated as pages that commit the company, reviewed in full in every language, as detailed in our comparison of automatic and reviewed translation. The forms, currency and price notices around them are a matter of localization, and product pages, with their mandatory notices, of translating product pages for export.

What happens to the contract when you translate it

A contract contains names, addresses, amounts and negotiated terms. Pasting it into an online translator sends it to a third party, on servers you do not choose, under terms that change from one plan to another.

The questions to ask the provider, and what the GDPR says, are in translation and GDPR: where your texts go. The general obligations of a company using AI are summarised in AI Act and GDPR: what an SME using AI must do.

What it costs, and where the savings are

A review by a lawyer or specialised translator costs more than machine translation, and that is normal: it is what carries the risk. The savings do not come from removing it, but from using its time better. A clean first draft, a maintained glossary and clauses already approved reduce the review work without lowering its standard. The orders of magnitude, per word and per page, are in website translation cost: per word, per page, per market, and the choice of first-draft tool in DeepL, Google or an LLM: which translates your trade best.

Measuring what the first draft is worth, without reading everything twice, is done on a sample and by error type: that is the subject of measuring translation quality.

Where to start

List the legal texts your company publishes or signs in another language, and sort them into three columns: what is for understanding (machine translation), what is published or signed (review by a lawyer from the country), what is filed with an authority (certified translator). The last two columns have their professional; the first, a tool chosen for what it does with the data.

The other texts for the market follow their own logic: support in translating customer support emails without hiring, the app in translating software, the blog in translating a whole blog: which articles first. Choosing the market is prepared with keywords abroad, opening an English-speaking market, which European market to open first and, further afield, selling in China or Southeast Asia.

To organise the translation of your pages and contractual texts between machine, translators and lawyers, tell us about your documents and languages: we set up the workflow and the tool, your professionals keep the sign-off.

Frequently asked questions

Sources

1. Findings of the WMT25 General Machine Translation Shared Task, November 2025 (new tab)aclanthology.org
2. French Ministry of Culture, guaranteeing the use of French (new tab)culture.gouv.fr
3. Service-Public.fr, translation by an approved translator (new tab)service-public.gouv.fr