A sales rep gets a request in German and pastes it into a free browser translator. Descriptions of 800 products for the German market have been waiting for a year, because the translation agency's quote was a number nobody wanted to sign off. And the machine's operating manual was "translated" by someone in the office who learned the language at school.
This post lets you sort the texts your company translates into three buckets: machine translation only, machine translation with review, and human translator only. Afterwards you will know where machine translation saves real money and where one error costs more than a year of subscriptions.
Three buckets of text
The most common mistake we see: a company asks "is the translator good enough" instead of "what happens if this text contains an error". The same translation quality is fine for one document and unacceptable for another.
Bucket A — translating to understand. Incoming emails from customers and suppliers, technical documentation from a component manufacturer, competitors' offers, correspondence with a carrier. You read these, not your customer. A machine translator without review is enough. An error costs at most one follow-up question.
Bucket B — translating with review. Outgoing emails, product descriptions in your shop and on marketplaces, data sheets, sales offers, job ads for foreign workers. The translator produces a first draft, and someone who knows both the language and the product reads it. We typically see review taking 30–50% of the time a translation from scratch would take.
Bucket C — human only, or full review by a translator. Contracts, general terms of sale, complaint responses with legal risk, operating manuals and safety warnings, documents for courts or authorities. If an authority or a business partner requires a certified translation (in Poland, by a sworn translator, tłumacz przysięgły), no tool replaces that.
Machine manuals are a separate topic. The EU Machinery Directive requires instructions in the official language of the country where the machine is used (source: Directive 2006/42/EC, Annex I 1.7.4, EUR-Lex). From 20 January 2027 it will be replaced by Regulation 2023/1230 (source: EUR-Lex), so if you build machines, check with a technical translator what changes in your documentation.
Glossary and data terms — this is where the work is
A machine translator does not know your product range. Without a term list, the Polish word for "guide rail" comes out as "Führung" one time and "Schiene" the next, and the name of your product line gets translated literally. Your customer then sees three different names for one thing.
A glossary is a table: source term, approved translation, note. For a manufacturer or distributor it usually has 150–400 entries. Building it takes 2–4 days of work by someone who knows the product and the language, spread over 2–3 weeks. The sources are your existing catalogues, offers and correspondence with your largest customers.
The second issue is data. Free versions of translators and AI chat tools are consumer products. Paid versions usually come with business terms. For example, DeepL states that in its Pro version texts are deleted after the service is performed and are not used to train models (source: DeepL privacy policy). Pro plans cost from about EUR 7.50 to EUR 50 per user per month on annual billing, and the number of glossaries depends on the plan (source: DeepL Pro pricing, as of October 2026).
Do not buy accounts for everyone. Usually 2–5 accounts are enough, for the people who actually send texts abroad. Write down what may be pasted into which tool in a short AI usage policy, because a machine translator is, in practice, an AI tool too.
Two export examples
When it worked. In one of our projects, a 22-person Polish manufacturer of furniture fittings running Comarch Optima (a popular Polish ERP) wanted to bring its catalogue to the German market. A translation agency quoted about PLN 38,000 (approx. EUR 8,900) and 10 weeks for 1,200 product descriptions and 40 data sheets. Instead, we built a 280-term glossary, an export of descriptions from Optima to CSV, machine translation with the glossary, and an import back into the foreign-language description field.
Review was done by the sales rep who handles German customers, at 4–6 minutes per description. The whole thing took 5 weeks and cost about PLN 9,000 (approx. EUR 2,100) to set up, plus subscriptions for 3 people. We still sent the data sheets with safety parameters to a technical translator for review, for about PLN 3,000 (approx. EUR 700). We covered descriptions generated with AI in more detail in our post on AI for product descriptions.
The lesson: we would have started the glossary a week earlier. The first 200 descriptions had to be corrected twice, because the product series names were added to the glossary only after review.
When it cost more. At another client, a 35-person manufacturer of packaging machines, a 140-page operating manual for a customer in Austria was machine-translated and sent without review. The customer challenged the documentation at acceptance: inconsistent names for control elements and several ambiguously worded safety warnings. Review by a technical translator cost about PLN 11,000 (approx. EUR 2,600) and took 3 weeks. During that time the customer withheld the final instalment, 30% of the machine's value.
The lesson: a manual is bucket C. The machine translator could produce the first draft, but the budget and timeline for a technical translator's review should have been in the machine's cost estimate from the start, not added after a complaint.
When it does not make sense
Machine translation does not always pay off. These are the situations where we recommend something else.
- You translate fewer than a few pages a month. Setting up a glossary and a process will not pay back. Give 1–2 people a paid account and stop there.
- You have no one to review. Bucket B without someone who knows the language turns into bucket A sent to the customer. A regular freelance translator paid for post-editing machine output, rather than translating from scratch, will then be cheaper.
- One market makes up most of your sales. If 60% of your revenue comes from Germany, consider hiring a fluent German speaker part-time. They will do the review, and they will also answer the phone.
- Your texts live in PDFs and graphic files. Translators break catalogue layouts, and rebuilding them can take longer than the review itself. Keep the content in your ERP, your shop or Word first, then translate.
- Marketing for your main market. Slogans, the homepage and campaigns need someone who knows the market, not just the language. A translator gives you correct sentences that convince nobody.
What actually breaks: terminology drifts when nobody updates the glossary, and product names get translated literally. On top of that, staff go back to free tools because the paid account "belongs to someone else".
What to do this week
- List which texts left the company in a foreign language over the last 3 months: type, language, number of pages, who translated them.
- Assign each type to bucket A, B or C. When in doubt, choose the higher bucket.
- Ask sales and customer service which tool they use today and whether they paste customer data into it.
- Collect the 50 terms that appear most often in your offers and write down their approved translations from recent catalogues. That is the start of your glossary.
- Check whether the cost estimates for machines or equipment going abroad include a line for a translator to review the manual.
Do not ask whether the translator is good enough. Ask what an error in a given text costs, and decide on that basis: machine only, machine with review, or a translator.