AI translation in business: DeepL, ChatGPT or Gemini — which to choose and what not to leave to the machine

How does an AI translator differ from an old-style machine translator?

Classic machine translation worked sentence by sentence. An AI translator based on a large language model (a program that has learnt from billions of sentences and predicts the most sensible continuation of a text) looks at the whole paragraph, and often the whole document, so it understands context better and knows that an “offer” in a sales email is something different from an “offer” in a job advert.

The second difference is that you can talk to an AI-based translator. Write “translate into German, formal register, for a purchasing director, no anglicisms” and you will get text in that style. The old translators couldn’t do that. We show how to phrase such instructions so that they work every time in our guide to prompts for marketing, and in English tools of this kind are simply called “AI translators”.

The third difference can be a problem. AI can be overconfident, and when it doesn’t understand a passage, it doesn’t leave a gap but adds something that sounds good. So for important texts, always compare the output with the original. We explain where such fabrications come from in our piece on AI hallucinations, and for the wider context, meaning where translation sits among other uses of AI in business, see our guide to AI for marketing and PR.

Who saves the most time with an AI translator?

The biggest winners are companies where translation is everyday routine rather than a one-off project. A sales rep replies to an enquiry from Germany. Customer service receives a complaint in Czech. Marketing prepares an English version of a product page, and each of these people currently loses time copying text between windows or waiting until a translator has a moment.

• Sales correspondence and customer service: a quick reply in the customer’s language that keeps the company’s tone.
• Offers, product sheets and manuals: translating entire files without breaking the formatting.
• Internal communications in companies with teams abroad: announcements, procedures, training.
• Meetings and calls: live speech translation, for now with the caveats described below.
• Monitoring foreign media and competitors: understanding the gist of a text before deciding whether to commission a full translation.

DeepL, Google Translate, ChatGPT or Copilot?

The question “best LLM for translation” (i.e. which language model translates best) has no single answer, because quality changes with every model version and depends on the language pair. In a business, other things matter more. Does the tool stick to a term glossary? Does it translate files without breaking the layout? And what happens to the data?

ChatGPT as a translator works best where the text also needs reworking, i.e. shortening, changing the tone or adapting to the audience; this is essentially editorial work, which we describe under AI copywriting. Copilot as a translator is convenient for companies working in Microsoft 365, because it runs directly in Word and Outlook. DeepL was built for translation and has features chatbots lack: glossaries (term lists enforced in every translation) and translation memory.

Live conversation translation: what already works and what doesn’t yet

On 9 June 2026 Google announced Gemini 3.5 Live Translate, a speech-to-speech translation model which, according to the company, recognises more than 70 languages. It is coming to the Google Translate app on Android and iOS, and in Google Meet it is currently in a closed testing programme for selected companies.

Microsoft has an interpreter agent (Interpreter) in Teams that translates what meeting participants say in real time. According to Microsoft’s help page, a Copilot licence provides 20 hours of translation per user per month. Ten languages are supported, including English and German, as well as French and Spanish. Polish is not on the list. In practice, a call with a German partner in which the Polish participant speaks Polish cannot yet go through this feature.

How to introduce an AI translator in your company step by step

A rollout is not about buying licences. It is about agreeing who uses what, for which texts and who checks the result, and the sequence below protects you from chaos in which every employee uses a different free tool.

1. STEP 01List what the company translates

   For two weeks, record the type of text and language, the volume and the name of the person translating. It often turns out that most of the volume comes from a handful of recurring text types. Start with those.

2. STEP 02Divide texts into three risk groups

   Define a low group (internal emails, background research), a medium group (offers, website) and a high group (contracts, safety instructions, public statements), and set a different procedure for each.

3. STEP 03Build a glossary of terms

   Enter product and job titles and the phrases the company does not translate, then upload it as a glossary in DeepL or as a standing instruction in ChatGPT or Copilot.

4. STEP 04Choose one company tool

   Choose a plan that does not train models on your data and has an admin panel, because only then will you know who is using what. Block private accounts for company business, or at least clearly discourage them.

5. STEP 05Decide who checks medium- and high-risk texts

   It can be someone who knows the language or an external translator, as long as every published text records that person’s name, so that it is clear whom to contact if questions arise.

6. STEP 06After a month, measure the time and the number of corrections

   Count how many texts went through unchanged, how many needed correcting and where the AI went wrong most often, because that is what you will use to improve the glossary.

What can you hand over to AI, and what must a human check?

The line does not run between “important” and “unimportant” text. It runs between text in which an error is easy to spot and fix, and text in which an error has consequences. AI as a translator handles meaning well, but legal and cultural nuance less well.

How much does an AI translator cost and what drives the budget?

AI translation tools are cheap compared with the cost of an error, because the time spent on corrections costs more, and for public texts so does the work of someone who knows the market and will catch sentences that sound foreign. The table shows public prices (as at 24 September 2026; ChatGPT and Microsoft excluding VAT, DeepL including VAT as shown on its website).

Data security: what to paste in and what not to

The biggest risk is not translation quality but where the text ends up. An employee who pastes a customer contract into a free browser translator does not know how long that text will be stored or what it may be used for. DeepL states that Pro subscribers’ data is not used to train models and is deleted once the service has been performed, and OpenAI states that ChatGPT Business does not train on company data by default. Free accounts can be a different story.

The second layer is regulation. Since 2 February 2025, Article 4 of the AI Act has required companies using AI to ensure their staff have the relevant skills, and a short instruction on “what we don’t paste into the translator” is the cheapest way to meet this obligation. We write more about the risks in our article on the risks and ethics of AI in business, and the principles of talking to customers are covered in our piece on communication with customers.

When an AI translator pays off, and when it isn’t enough

An AI translator almost always pays off for large volumes of short, repetitive texts. As the only link in the chain for texts that build the company’s reputation abroad, it won’t do. It is time to get translation in order when employees use several different free tools, foreign customers ask what your emails mean, or product names sound different in each language every time.

When entering a new market, translation is only the beginning. What local media write about the company matters too, as do the answers search engines and AI models give about it in that language. We explain the difference between ranking in Google and visibility in AI answers in our article on how SEO differs from GEO and AEO, and the role of the media in what ChatGPT says about a company in our piece on PR AI.

AI translation in a company’s international strategy

Companies that use AI translators well treat them as part of a bigger system. They have a term glossary, one person responsible for tone in each language and a local partner who knows the media and customs. The machine speeds up the work. It won’t, however, explain what a buyer in Germany expects as opposed to one in the Czech Republic.

If you are growing sales abroad, start by checking how visible your company currently is in a given country: in the media, in search and in AI answers. We help with this as part of international PR. Recurring translations can also be linked to other processes, as we show using the example of automation with AI and n8n.

The AI translation mistakes that cost companies the most

Read next