AI Tools
AI Translation for Client Work: The Sign-Off Nobody Assigns

☰ Table of Contents
Almost every guide to AI translation for client work argues about quality: which engine is more fluent, which one handles idioms, whether the output still sounds like a person wrote it. That argument is a decade old and mostly settled. Any current model will produce a readable German or Spanish paragraph from an English one.
What it will not do is take responsibility for the paragraph. That is still your job, and it is the job most freelancers forget to schedule. A translated proposal, landing page, packaging insert or support email carries names, prices, delivery promises and legal hedges. When one of those is wrong, the client does not blame the engine. They blame the invoice they paid.
So this is not an engine comparison. It is the missing step after the engine: a sign-off protocol you can run in about twenty minutes, built around one question that nobody in the workflow has been asked. Who checks the translated copy before it goes out, and against what?
Translation Is Rarely the Part That Breaks
We went through the failure reports we see most often on translated client deliverables, and the pattern is consistent. The language is usually fine. The breakage sits in the details that a translator would have flagged and a model has no reason to question:
- A price that changed currency formatting and now reads like a different number.
- A product name that got translated into a generic noun, so the client’s trademark disappeared mid-sentence.
- A date written 03/04 in one section and 4 March in another.
- A call-to-action button that says “Learn more” in English inside a fully translated page, because it was an image.
- A polite form of address that lands as either rude or unnaturally stiff depending on which market the reader is in.
None of those are translation errors in the narrow sense. They are handoff errors. The work was translated, then sent, and no one looked at the result as a business document rather than as text.
If you already run a versioning habit on client files, this slots in beside it: the same discipline that keeps two file versions straight also keeps two language versions straight. Our guide to client deliverable file naming covers the naming half of that problem.
Sort the Job Into Three Risk Tiers
The reason sign-off gets skipped is that it feels like it applies to everything, so it gets applied to nothing. Split the work first. Different tiers deserve genuinely different amounts of human attention, and pretending otherwise is how reviews collapse into a quick skim.
| Tier | What lands here | Who signs it off |
|---|---|---|
| Tier 3 — consequential Refund, reprint or liability |
Contracts, terms, pricing tables, safety and legal notices, medical or financial claims | A native speaker with subject knowledge, every time |
| Tier 2 — visible Embarrassment, lost trust |
Landing pages, product descriptions, proposals, onboarding emails, packaging copy | You, against a fixed checklist, plus a native read on high-value pages |
| Tier 1 — internal Almost nothing |
Scratch notes, meeting summaries for your own use, draft briefs to yourself | Nobody. Send it and move on |
Two things follow from that table. First, most of your daily translation volume is Tier 1 and needs no process at all. Second, the small percentage that is Tier 3 deserves a reviewer even when the client says the budget does not allow one — because the reviewer is cheaper than the correction.
If you are choosing engines in the first place, the same tier logic tells you where to spend: a free engine is fine for Tier 1, while Tier 3 work should use a tool whose output you have already stress-tested on your own material, not a demo sentence.
The Six-Line Brief That Replaces “Translate This”
Half of the bad output we see is a bad instruction, not a bad engine. “Translate this to Spanish” leaves the model to guess the market, the formality, and whether your brand name is a noun or a foreign word. Six lines fix it, and you can paste the same six lines into any tool or hand them to a human translator:
- Market and variant. Not “Spanish” but Mexican Spanish for retail buyers. Not “Portuguese” but European Portuguese, formal.
- Reader and relationship. Existing customer or cold prospect; peer or authority; do they expect “you” or a formal address.
- Do-not-translate list. Brand names, product names, feature names, certifications, internal codenames.
- Fixed vocabulary. The five to ten terms your company always uses, with the approved translation for each.
- Numbers and units. Keep source currency or convert; metric or imperial; date format the market expects.
- Output shape. Same structure, same headings, no added explanation, no translator notes in the middle of the copy.
The fixed vocabulary list is the highest-value line. Consistency is what makes translated copy feel professional, and consistency is exactly what a fresh model call cannot guarantee on its own: ask for the same paragraph twice and the key term may come back two different ways.
Lock the Things Machines Get Wrong Silently

Some errors announce themselves; you read the sentence and it feels wrong. The dangerous ones are invisible, because they render as perfectly normal text in the target language. Lock these before you review anything else:
- Money. Currency symbols and decimal separators move. 1,500.00 in one locale is 1.500,00 in another, and a model may swap the convention halfway through a table.
- Dates. 03/04 means two different days depending on who reads it. Write the month out, or use an unambiguous format, in every language version.
- Numbers with units. Ranges, dimensions, tolerances and page counts drift when a unit system gets converted. Round-trip them and compare against the source.
- Names. Person names get localized (a William becomes a Guillermo), company names get inflected, and product names turn into descriptions.
- Negation. In many languages the negative particle attaches in a place a model can misplace, and a flipped negative in a safety line is the worst possible error.
- Placeholders and links. Variable tokens like {first_name}, URLs and tracking parameters must survive byte for byte.
Make the fix mechanical. Build a small do-not-touch sheet for each client — brand terms, the numbers that matter, the links — and answer one question per item: does it appear unchanged in the translated version? That is a five-minute comparison, and it catches the class of error that costs the most.
Read the Output for Facts, Not Grammar
Fluency is the one thing you can safely assume. It is also the thing reviewers spend their attention on, which is why the factual errors survive. Review in this order instead:
| Check | How to run it | What it catches |
|---|---|---|
| Non-translatables | Search the target text for each brand and product term on your list | Localized names, generic nouns replacing trademarks |
| Numbers and money | Compare source and target line by line, digits only, ignoring the words around them | Separator swaps, unit drift, rounded prices |
| Links and CTAs | Click every link; read every button label in context | Untranslated images, broken URLs, buttons that promise the wrong action |
| Negation and conditions | Read every sentence with “not”, “unless”, “only” in both versions | Flipped meaning in terms, safety notes, eligibility rules |
| Headings and order | Check that the structure matches, section for section | Merged or dropped sections, reordered steps |
| Placeholders | Search for the { } pattern and for any token your system injects | Variables translated into words, which silently breaks personalization |
Only after those six checks do you read for tone. If the text reads smoothly and the facts are right, the job is done. If it reads smoothly and a price is wrong, the smoothness was never the point.
Where Machine Output Fails Predictably
Certain sentence types break in the same ways every time, in every engine, which means you can look for them directly instead of reading with vague suspicion:
Idioms and figures of speech. They come back literal or replaced with an unrelated local idiom. Neither is wrong grammatically; both can be wrong commercially. If a phrase only works in English, rewrite it in English first and translate the plain meaning.
Marketing superlatives. Claim language is regulated differently by market. A word that is harmless in the US can be a compliance problem in Germany, and the model has no reason to soften it. Tone down before translating, not after.
Legal hedges. Words like “should”, “may” and “reasonable” carry specific weight in contract language. A translation that reads naturally can quietly upgrade a suggestion into an obligation. Tier 3 work belongs with a human for exactly this reason.
Interface strings. Short labels need context the model does not have. “Post” is a noun or a verb depending on the button, and a table cell labelled “Back” becomes the body part in at least one language. Give UI text a note about the element it sits in.
Formality systems. Languages with a formal and informal second person force a choice English never made. That choice is a brand decision, not a linguistic one, so it belongs in the brief.

There is a useful side effect to naming these categories: they become a pre-flight list. Before you translate a page, scan it for idioms, claims, hedges and UI labels, and fix them in the source. Translating cleaner input is cheaper than repairing output, and it improves the English too. The same pre-flight habit shows up in our checklist for AI browser agents, where the point is deciding what a tool is allowed to touch before it touches anything.
When a Human Reviewer Pays for Itself
The objection to a native reviewer is always cost, and the math is usually done wrong. People compare the reviewer’s hourly rate to zero, as if the alternative were free. The honest comparison is reviewer cost against the cost of the failure you are protecting against.
Run it on your own numbers. A two-page translated pricing sheet for a distributor in a new market: a native review of that length is typically an hour or two of work. The failure it prevents is a wrong price reaching a customer, which means a credit, a reprint, an apology and a conversation about competence with the person who signs your invoices. For a Tier 3 document, the reviewer is the cheap option almost every time.
The same logic tells you where not to spend. Paying for a native review of internal scratch notes is waste, and doing it teaches you to dread the process, which is how processes die. Save the reviewer for the documents where a mistake has a price tag attached, and be generous about using them there.
A Handoff You Can Repeat Next Month

Here is the version that survives a busy month, in the order it happens:
- Classify. Before touching a tool, write the tier at the top of the document. Tier 1 goes straight through; Tier 2 and 3 get the rest of this list.
- Clean the source. Fix idioms, soften unregulated claims, expand ambiguous dates, and mark every do-not-translate term in the source file itself.
- Write the brief. Six lines, saved in the file, reused next time for the same client and market.
- Translate in sections. Section-by-section output is easier to compare against the source than one long block, and easier to re-run when one part fails.
- Run the fact checks. Non-translatables, digits, links, negation, structure, placeholders. In that order, before you read for tone.
- Record the sign-off. One line in the delivery email: who read it, in which market variant, and which checklist it passed. That sentence is the whole protocol, and it is also what a client remembers when something goes wrong.
The instinct when volume rises is to skip the checks, not the translation. Resist it in exactly one place: the sentence that names who signed off. If you cannot name that person, the document is not ready to send, and it is cheaper to find that out in your own review than in the client’s inbox.
FAQ
Is machine translation good enough for client-facing work?
For Tier 2 pages and emails, usually yes, with a fact check and a brief. For contracts, pricing, safety notices and any regulated claim, no. The output can read perfectly and still carry the wrong obligation, and there is no way to see that from the language alone.
How long should a translation review take?
About twenty minutes for a two-page document if you use the six fact checks in order. It takes longer when you read for style first, because style reading hides the numbers you were meant to compare.
Should I tell the client that AI did the translation?
You do not have to lead with the tool, but you should not hide the process either. The useful disclosure is about accountability: who reviewed it, in which market variant, against which checklist. Clients care about that more than about the engine name.
Do I need a native speaker or is fluent enough?
For Tier 2, a fluent reviewer with market knowledge is generally fine. For Tier 3, use a native speaker who does that kind of document professionally. Fluency catches awkward phrasing; only familiarity with the local regulatory and commercial conventions catches the errors that matter.
How do I keep terminology consistent across months?
Keep a fixed vocabulary list per client and market, five to ten terms with approved translations, and attach it to every job. Consistency is what readers notice as quality, and it is the first thing a fresh translation loses.
What is the single most common mistake?
Translating, then sending. The tool gets the language right often enough that people stop looking, and the sign-off step disappears from the workflow entirely.

