Decide whether you need understanding or a publishable reply
Reading an incoming message to work out which colleague should handle it is different from sending a delivery commitment. Decide which job you are automating. An internal translation should remain beside the original, clearly labelled as machine-generated. A customer-facing reply needs a defined check before it leaves your business.
The European Commission’s notice about its own machine translation says accuracy varies across texts and language pairs. That is a useful limitation to recognise, not a benchmark for another provider. Treat each language pair and message type as something to evaluate, even when the English output reads well.
Start with one repeatable task, such as translating routine order-status drafts. Keep disputed terms, safety instructions and unfamiliar specialist language on a route with appropriate language and subject expertise.
Illustrative example: dispatch is not delivery
Imagine a UK spare-parts supplier replying to a French-speaking customer. This is a fictional workflow, not a Fewzen client result. The approved English message says: “We expect to dispatch 12 units on 4 December, subject to stock confirmation. Delivery takes another three working days.”
A translation that promises arrival on 4 December has changed the offer. So has one that removes the stock condition or changes the quantity. A friendly tone does not compensate for those errors. The reviewer needs to compare meaning, including the distinction between an expectation and a guarantee.
Prepare the English reply from the actual order record first. Then translate that approved text as a separate step. Do not ask one broad prompt to interpret the enquiry, decide stock availability, negotiate a date and translate the resulting promise.
Specify the language and agree the important terms
Record the requested target language and relevant regional variant. Do not infer a customer’s preference from their name. Check the exact language options in the service you choose: Microsoft’s Translator documentation, for example, says its pt code defaults to Brazilian Portuguese, while pt-pt identifies Portuguese for Portugal.
Build a small, reviewed terminology list for product names and recurring business terms. Include the intended meaning and an example sentence. A distinction such as dispatch versus delivery needs contextual review; a list of preferred words cannot establish whether the whole sentence makes the right promise.
Provider controls differ. Google Cloud Translation Advanced supports glossaries, but its documentation notes case-sensitive matching by default and ignored stopwords. Microsoft limits its dynamic dictionary to uses such as proper and product names, requires English on one side of the language pair and an explicit source language. These are documented features checked on 22 November 2026, not interchangeable guarantees.
If you use a general-purpose AI assistant, supplying a terminology list is an instruction to evaluate. Do not assume it provides the same behaviour as a translation API’s glossary feature.
Give the reviewer a short, concrete checklist
Show the original, the approved reply and its translation together. Highlight the facts that must survive. Use unambiguous dates in the source and explicitly named currencies where relevant. Do not let the translation step convert units, currency or time zones unless that is a separately defined requirement.
Software can flag missing product codes or changed quantities. Comparing raw digits alone is insufficient because number formatting varies by language. Keep a structured record of the intended values and have the reviewer confirm how they are expressed.
The person checking the target text must understand it well enough to assess meaning. A colleague who only reads English cannot establish accuracy by looking at a fluent translation back into English.
| Check | What must remain true |
|---|---|
| Order facts | 12 units; the correct product and order references. |
| Timing | 4 December refers to expected dispatch, not arrival. |
| Conditions | Stock confirmation is still required. |
| Meaning | Three working days has not become three calendar days. |
| Authority | No new discount, guarantee or offer has been added. |
Test the complete hand-off before allowing automatic sends
Create a small pilot set from permissioned, suitably redacted examples. Include short replies, mixed languages, ambiguous dates, negative statements, corrections and unfamiliar product names. Ask a competent reviewer to classify errors by consequence: awkward wording is different from reversing a cancellation or removing a condition.
Measure review time, material meaning errors and messages returned for clarification. Compare the complete effort with your current process, including terminology maintenance and access to reviewers. Low translation usage charges do not establish a useful saving if every output takes longer to repair.
Keep the final send separate from translation during the pilot. Store the exact source version and reviewed output together in your approved business system. An edited source invalidates the earlier translation approval. Decide who handles a message when the language is uncertain or the reviewer is unavailable; an accumulating queue needs an owner.
Start with one language pair and one kind of reply
Choose a routine message, agree its facts and identify someone who can assess the target language. Run it through your existing business-approved translation tool before commissioning an integration. Only send the information needed for that task, using an account and data-handling arrangement your business has approved.
Review the pilot’s difficult examples together. If the team cannot reliably check the result, keep AI translation as an aid to understanding and use qualified language support for consequential replies. Expand only when you can explain what is checked, who owns exceptions and why the process is worth its operating cost.