Machine translation can be good enough as a starting point for routine client-facing documents when a competent translator checks the meaning and the finished file. Do not release high-consequence instructions, contracts or sensitive care information on fluency alone. Choose the review level by what a wrong sentence could make someone do.
The useful question is whether this document, in this language direction, can reach the required standard at a sensible total cost. A polished draft is not the same as an approved translation. If nobody can compare it with the source, you have not established that it is ready for customers.
Decide what the reader might do with a wrong sentence
Start with the action the document supports. Will someone choose a colour, place an order, install a part, sign an agreement or follow care instructions? A small language error has different consequences in each case. Classify sections separately when a document contains several kinds of information.
| Document or section | Illustrative consequence of an error | Sensible starting route |
|---|---|---|
| Public company description | An awkward explanation or unsuitable tone | Machine draft with bilingual meaning and style review |
| Catalogue, quotation or delivery information | A buyer orders the wrong item or expects the wrong service | Machine draft only with full checks of meaning, terms and commercial details |
| Safety instructions or care information | A reader acts on a wrong limit or instruction | Specialist-led translation and review; agree any machine use first |
| Contracts and formal declarations | A party misunderstands an obligation | Qualified specialist route with the required professional checks |
These are recommended working routes, not legal classifications. For documents with legal, clinical or safety consequences, ask a suitably qualified adviser what review is required. Do not decide that a document is low risk just because it is short. A twelve-word warning can matter more than a twelve-page brochure.
As an illustrative comparison, a packaging supplier's sentence “Available in three colours” is a different task from “Not suitable for direct food contact”. A missing “not” in the second sentence changes the use a customer might make of the product. The supplier should identify that sentence for specialist factual and linguistic review before choosing a translation method.
Check permission separately from quality. A good translation does not authorise uploading confidential source material or ignoring a client's prohibition on machine translation. Resolve those questions before running the trial.
A packaging supplier tests one catalogue before committing
Consider an illustrative packaging supplier preparing a 2,400-word catalogue for a new customer audience. It contains product descriptions, minimum order quantities, lead times and storage instructions. The owner has an approved translation service and a reviewer competent in both languages and packaging terminology.
The owner first selects 300 words from different parts of the catalogue: one product description, a specification table, an order condition and a storage instruction. Picking only the welcome page would make the trial too easy. The aim is to discover the types of correction needed, not to award the software a general accuracy score.
The reviewer compares the draft with the source and records three problems. “Lead time starts after artwork approval” has lost its starting condition. “Pack of 25” has become “25 packs”. The same technical term has two translations. These are illustrative findings, not results from a named tool test.
The owner clarifies the source and supplies the approved terminology before processing the whole document. The 300-word trial does not approve the remaining 2,100 words. The full catalogue still goes through bilingual review, number checks and a check of the final layout.
A quick sum shows why fixing the source first pays. The sample was 300 of 2,400 words, one-eighth of the catalogue, and it produced three problems, each of a repeatable kind. If those patterns ran through the rest at the same rate, the reviewer would meet roughly 24 of them in the full draft: every "pack of" line, every lead-time condition, every use of the inconsistent term. Correcting the source wording and the glossary once removes most of them before they are generated, and turns the full review into a check rather than a rewrite.
The release standard is written before that review: no unresolved change to meaning, product identity, quantity, condition or instruction. Style corrections must also be finished to the agreed client standard. One significant error is enough to hold the affected document; it does not become acceptable because the other pages read well.
The supplier uses this illustrative budget to compare two routes. The assumed labour rate is $45 an hour throughout, solely to make the arithmetic visible. Real translator quotes may use a per-word, hourly or project fee.
| Work | Machine-assisted route | Human-led comparison |
|---|---|---|
| Source preparation and brief | 20 minutes | Included below |
| Generate and arrange draft | 10 minutes | Included below |
| Full bilingual review and corrections | 110 minutes | Included below |
| Layout and final checks | 40 minutes | Included below |
| Total labour | 3 hours: $135 | 4 hours: $180 |
| Assumed tool cost allocated to this job | $10 | $0 additional |
| Total planning cost | $145 | $180 |
The apparent difference is $35. One extra hour correcting a difficult machine draft raises its total to $190, above the human-led estimate. The $10 tool allowance is a budgeting assumption, not a vendor price. Add any specialist review needed to both routes so the comparison buys the same finished standard.
Measure actual editing and layout time on the pilot. If corrections are extensive, commission a fresh translation rather than asking someone to rescue a poor draft indefinitely. For a fuller budget, see what AI and human translation really cost.
Read for meaning before you polish the tone
Use separate passes. First compare the source and translation for what each sentence says. Then improve style for the intended reader. Finally inspect the exported file. Mixing all three passes makes it easy to spend time on elegant wording while missing a changed order condition.
Ask the reviewer to mark each issue with its source location, the draft wording, the corrected meaning and its consequence. A note such as “page 4, order table: minimum quantity changed from cartons to individual items” is useful. “Translation quality seems mixed” gives the owner nothing to act on.
Here is a second illustrative example from an import-export business. The wording below explains the translated meaning in English so an owner can see the commercial difference; it is not presented as an actual target-language translation.
- Source: “Dispatch within ten working days after receipt of the deposit.”
- Faulty translated meaning: “Delivery within ten days of order.”
- Correction: Preserve dispatch, working days and the deposit as the starting event.
- Reason: The draft changes both the deadline and what the business promises to have completed.
Three small changes can create one substantial promise. Checking only the number ten would miss all of them. Build a list of words that change obligations: after, before, unless, only, approximately, up to, excludes and subject to. Have the reviewer check their function rather than simply searching for equivalent vocabulary.
Numbers need a check of their own, because conventions differ between languages. In an illustrative quotation, the source shows a unit price of 1,250.50 and a delivery date of 03/04. The draft converts the price to the target language's decimal convention in the table but leaves it untouched in the covering paragraph, so one document now shows the same price two ways. The date is worse. Readers used to day-first and month-first formats will pick different months, and no translation tool can know which one the author meant. Write dates with the month in words in the source ("3 April"), and ask the reviewer to confirm that every number appears in one consistent format throughout the file.
DeepL's published explanation of its Translation Quality Evaluation work makes a useful distinction: detecting ambiguity does not supply the missing context needed to resolve it. Human reviewers remain part of that process. Treat automated checks as prompts for investigation, not evidence that an uncertain sentence is correct.
A fuller pre-publication translation check helps once you have chosen the production route. The suitability decision comes first: do you have the right reviewer, source material and time to reach that standard?
Give the translator a brief that removes guesswork
A translation tool cannot resolve a business fact that the source leaves unclear. Before generating a draft, state who will read it, what they should do, the intended tone, the approved terminology and the parts that must remain unchanged. Include the full relevant passage around ambiguous terms.
For an illustrative spare-parts manufacturer, “seal” might mean a physical component, a sealing action or a mark on a document. Giving only a column of isolated words removes the context that distinguishes them. A brief should identify the part, its function and any approved customer term.
Use a prompt like this with an approved text assistant to inspect a public or fictional source before translation. Do not upload restricted source material merely to prepare the brief.
Check this source text for translation ambiguity. Do not translate it.
Audience: purchasing staff ordering spare parts.
Purpose: identify what is included in one order line.
Return: unclear phrase | possible meanings | question for the owner.
Do not invent missing product facts.
Source: "Seal kit, 6 pieces. Replace after service. Sold per set."
Known fact: one set contains six seals of the same size.Illustrative output: “Replace after service | replace after maintenance, or after a period in use | Does ‘service’ mean a maintenance visit or the part's operating life?” That is a useful question. Suppose the output also says, “Replace all six seals after every maintenance visit.” Delete that sentence: it has guessed the maintenance requirement.
The owner asks the technical lead, then changes the source to: “One set contains six identical seals. Follow the replacement interval in the approved maintenance instructions.” The translator now has clearer wording and a reference to check. If that reference is unavailable, keep the instruction on hold.
Keep a small terminology list with a definition, approved target term, source and review date. Do not accept a different term simply because it sounds more natural to a general-purpose assistant. For a repeatable process, see keeping terminology consistent across translations.
The final file can fail even after the words are right
Review the file the customer will receive. Text can be correct in a review document yet appear under the wrong heading, outside a table or cut off in a narrow box. Count sections and table rows. Check footnotes, captions, labels and warnings as well as paragraphs.
For a third illustrative business example, a laboratory sends a sample-submission guide with two columns: “ambient transport” and “chilled transport”. After layout changes, a temperature range sits visually between them. Every translated word may be accurate, but the customer can apply the instruction to the wrong sample category.
The repair is to restore the heading and cell relationship, then ask the laboratory's authorised technical reviewer to verify the complete table. Comparing a list of extracted sentences would not catch this. Use the approved source for the temperature requirement; do not ask an AI assistant to choose a plausible range.
A home-care provider has another reason to check the whole file. In an illustrative welcome pack, the contact hours are correct, but the sentence explaining what to do outside those hours disappears at a page break. Count the full instructions against the source and confirm their order. Any care or escalation guidance needs review by someone qualified to approve it.
Do not use back-translation as the only release test. Translating the result back into the original language can raise useful questions, but a plausible return sentence does not prove the target text is suitable for its audience. A second machine's agreement is also not a substitute for a competent source-to-target review.
Keep evidence for a yes, and a clear route for a no
Record the decision in a short release note. For the packaging catalogue, a filled-in illustrative record might read as follows.
- File: catalogue_v4, 2,400 source words, approved language direction.
- Route: machine draft followed by full bilingual review and layout inspection.
- References: product list revision 8, glossary revision 3, approved order conditions.
- Corrections: quantity unit, lead-time condition and inconsistent terminology resolved.
- Release checks: all tables, warnings and order conditions compared with the source; no unresolved meaning queries.
- Approval: reviewer and product owner recorded against the final exported file.
Keep the source version with that record. If the owner changes the minimum order after approval, identify and review every affected occurrence in the translation. Do not assume an earlier sign-off covers a newly edited document. Repeated documents still need checks for changed prices, dates, quantities and instructions.
Track whether the route keeps earning its place. Record the reviewer's correction time per 1,000 source words on each job. For the illustrative supplier, the first catalogue took about 46 minutes per 1,000 words (110 minutes for 2,400 words). If the next one, with the glossary in place, falls to around 30, the preparation is working. If it climbs, or if customers start asking questions that trace back to wording, such as two buyers querying how many items a pack contains, go back to the sample stage before the next job rather than pressing on.
If the sample trial reveals repeated meaning errors, if the source is unclear, or if nobody suitable can review the target language, the answer is “not through this workflow yet”. Ask a translator to quote for the full job or a defined review. Reducing the scope to the pages customers actually need may be more sensible than lowering the standard.
When the sample is workable, the full review is affordable and the release record is complete, machine translation can be a useful part of delivery. Use the post-editing workflow to organise that work. Approve the finished document on the evidence in front of you, not on how quickly its first draft appeared.
Choosing a translation route for client documents
Can I judge quality from the translator's brand alone?
No. A brand name cannot tell you whether a particular document preserves your meaning. Ask for a sample using your subject matter, intended audience and language direction. Have a competent reviewer assess it against the source. Keep the corrected sample as a reference for later assignments, rather than treating a vendor ranking as acceptance of every future document.
Should I send the machine draft to the client for approval?
Agree that role before sending it. A client may be able to approve product facts without being able to assess the translation. If a draft is shared, label it clearly as unfinished and explain what feedback you need. Do not treat a client's silence or general approval as evidence that a qualified language review has happened.
What if the client needs a certified translation?
Ask the receiving party for its exact requirements before choosing a workflow. Certification, declarations and translator qualifications can depend on the purpose of the document. Confirm those requirements with a suitably qualified translator or adviser. A fluent AI output, or a general human review, does not by itself establish that the requested certification requirements have been met.
Further reads
- Does Your Translation Business Need an AI Usage Policy? — Set rules for permitted tools and client instructions.
- How Small Translation Agencies Build an AI-Assisted Workflow — See how an agency organises the wider delivery process.
- Is It Safe to Put Client Documents Into AI Translation Tools? — Check confidentiality before uploading source documents.
- How to Offer Multilingual Customer Support With AI Translation — Apply language checks to customer support conversations.
- How to Train Staff to Use AI in a Small Business — A three-layer training plan with learning outcomes, a 75-minute session agenda, a copyable error-spotting exercise and a language-school example.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: DeepL AI Labs, How we are building Translation Quality Evaluation for DeepL, reviewed 27 September 2026. Examples and budgets are illustrative; the acceptance rules are practical recommendations, not vendor accuracy claims.