Is It Safe to Put Client Documents Into AI Translation Tools?

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for Is It Safe to Put Client Documents Into AI Translation Tools?
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for Is It Safe to Put Client Documents Into AI Translation Tools?

It can be safe, but only on a business-grade service whose terms say your text isn't kept or used for training: DeepL Pro, the translation built into Microsoft Office and Azure, or Google's Cloud Translation API. The free versions are the risk. DeepL's own privacy policy says its free translator must not be used for texts containing personal data of any kind.

Privacy is only half of "safe". A translation can be perfectly confidential and still wrong in a way that matters: a negation dropped from a warranty clause, a date read the wrong way round, a legal term turned into its everyday meaning. And some documents can't be machine-translated for official use at all, however good the output, because the body receiving them wants a certified translation by a qualified human translator.

Follow me on Instagram@sagnikteaches

What each translation tool does with your text

The names are the same across free and paid versions, which is exactly why staff get this wrong. Here is what the vendors' own pages say, as of September 2026:

Connect on LinkedInSagnik Bhattacharya
Tool and versionWhat happens to your textFit for client documents?
DeepL free translator (website and apps, not logged in to Pro)Content and translations kept for a limited period to train and improve DeepL's models; personal data not permittedNo
DeepL ProKept only as long as needed to translate, then deleted; not used to improve the serviceYes, with a staff rule to log in
Microsoft Translator apps, Translator for Bing, Microsoft Edge translation, Web TranslatorMicrosoft says these free end-user products record text and voice for service improvementNo
Translation features in Microsoft Office apps, and Azure TranslatorCovered by Microsoft's "No Trace" policy: not written to persistent storage and not used for trainingYes
Google Cloud Translation APIHeld briefly in memory to translate; Google says it doesn't use the content to train its translation featuresYes, but it's a developer service, not a website
Google Translate website and appNot covered by the Cloud Translation terms; falls under Google's general consumer termsNot for anything confidential
General chat assistants on consumer plansMay be used for training unless the switch in privacy settings is offOnly with training off and nothing identifying
ChatGPT Business, Claude Team, Gemini in WorkspaceDon't train on business content by defaultYes, under your normal data rules

One DeepL detail catches small teams: the Pro plans cap document uploads. The Individual plan includes 3 file translations a month, Team 20 and Business 100, according to DeepL's plan page. A business that hits the cap mid-month is the business whose staff quietly switch to the free site to finish the job, so size the plan to your real volume. Count last month's foreign-language attachments before choosing. If the garage below received eight (a few warranty letters, two service books, a supplier's recall notice), a three-file plan runs out in the second week, and the free site starts looking tempting on the ninth of the month.

Subscribe on YouTube@codingliquids

Scanned paper adds a step of its own. A photographed service book goes through text recognition before it's translated, and that step misreads handwriting more often than the translation misreads language: a handwritten 1 read as a 7 turns a service at 17,000 km into one at 77,000 km, and the translation will render the wrong number perfectly. Scan with the OneDrive or Google Drive mobile app rather than Microsoft Lens, which has been retired, and compare every number in the translated text against the image, not against the recognised text.

The same tool, two different rulebooks

The Microsoft split is the clearest example of why "we use Microsoft" isn't an answer. A receptionist who opens a customer's letter in Word and uses the Translate button on the Review tab is inside the No Trace policy. The same receptionist who opens the same letter as a web page in Edge and clicks the browser's translate prompt is using a consumer product that records text for improvement. Same company, same letter, same afternoon, different terms.

DeepL has the same trap in a different shape. The Pro service and the free site look almost identical. If someone's Pro session has expired, or they're on their phone at home, they may be pasting into the free version without noticing. The fix is boring and effective: bookmark the Pro login on every work device, remove the free apps from work phones, and say in writing which one is approved.

Three questions that settle it for any document

  1. What's in it? Names, addresses, health details, financial details, identity documents and anything under a confidentiality agreement are sensitive. A supplier's product leaflet isn't. Sort documents into those two piles first; the tutorial on classifying business data before using AI gives a simple scheme.
  2. What have you promised? Check your privacy notice and any contract with the client. If you told customers their details go only to named processors, an unlisted translation service is a gap to fix, not just a technicality.
  3. Which version will actually be used? Not which one you pay for. The one a tired member of staff will open at 4pm.

If the document is sensitive, only an approved business-grade tool is acceptable. If it's sensitive and also going to be relied on by someone else (a customer, an insurer, an official body), you need the checking step described below, or a human translator.

Run through a fortnight's documents at an illustrative two-dentist practice and the three questions sort them quickly:

DocumentSensitive?Relied on by someone else?Route
New patient's medical history form, filled in by handYes: health dataYes: the dentist treats on itScanned, then Word Translate on the recognised text to understand it; the dentist confirms allergies and medicines with the patient in person
Referral letter from the patient's previous dentistYesYesDeepL Pro, redacted; clinical details confirmed with the referring practice before any treatment decision
Supplier's price list for a new composite materialNoNoAny approved tool
Safety data sheet for the same materialNo personal dataYes: staff handle the productAsk the supplier for the sheet in the practice's working language; don't rely on a machine version
Complaint email from a patient's relativeYesNot yetDeepL Pro to understand it; the reply is written in the practice's language and translated, with the step below

Only one of the five needed nothing more than a tool. The medical history form is the one to watch: a missed allergy is the kind of error no confidentiality setting can protect you from.

Three small businesses, three foreign-language documents

A funeral director. A family arranging a funeral for a relative who died abroad sends a death certificate and a covering letter from the overseas authority, neither in a language anyone on the team reads. The certificate contains the deceased's full name, date of birth, cause of death and parents' names: special-category data by most definitions. The team can use DeepL Pro to understand what the letter asks for, so they can advise the family the same day. But if the certificate has to be registered or presented to an official body, that body will usually require a certified translation, so the director books a qualified translator for that, and tells the family the cost and time up front.

A car repair garage. A customer who bought a car overseas brings in a service book and a warranty letter in another language, and wants to know whether a gearbox repair is covered. The garage owner translates the warranty letter in Word (No Trace) after removing the customer's name and address from a copy. The translation says repairs are covered "if serviced at authorised dealers". Before quoting, he checks that phrase with a back-translation (see below), because "authorised" versus "any qualified workshop" decides who pays a $2,400 repair.

An independent bookshop. The owner is negotiating to stock titles from an overseas small press and receives its trade terms: discount, returns window, payment days. Nothing personal, but the terms are commercially confidential under the press's email footer. She uses DeepL Pro for the full document and asks the press to confirm the two figures that matter (returns window and discount) in English in an email, so the deal doesn't rest on a machine reading of one clause.

The confirmation earned its keep. DeepL rendered the discount line as "40% discount on orders of 10 copies or more", which she read as ten books in any mix. The press's reply clarified that the 40% applies to ten or more copies of the same title, and mixed orders get 35%. The translation wasn't wrong, just faithful to an ambiguous original, and on a first order of 60 mixed books at an average list price of $18 the five-point difference is $54. Small, but it's the difference between the margin she planned for and the one she'd have got.

Redact first: a before and after

Even on an approved tool, the safest text is text without identifiers. Most of the time you need the meaning, not the names. Here is the garage example before and after redaction:

Before (as received):
"Mr A. N. Other, 14 Linden Street, vehicle VIN WVWZZZ1KZ8W123456,
registration AB-CD 1234. Warranty extension no. 55-20931 covers
gearbox and clutch until 31.03.2027 if serviced at authorised dealers."

After (what goes into the tool):
"[CUSTOMER], [ADDRESS], vehicle VIN [VIN], registration [REG].
Warranty extension no. [NUMBER] covers gearbox and clutch until
31.03.2027 if serviced at authorised dealers."

Keep dates and amounts in: they're what you need translated accurately, and on their own they don't identify anyone. Keep the original file untouched in the customer's record. The tutorial on anonymising client data before pasting it into AI covers faster ways to do this for longer documents.

When machine translation isn't acceptable, however private

  • Documents for an official body. Registrations, court papers, immigration or licensing applications and similar usually need a certified or sworn translation. Ask the receiving body what it accepts before you translate anything.
  • Anything you'll sign. A contract you agree to on the strength of a machine translation binds you to the original wording, not the translation. Get the key terms confirmed in writing or reviewed by a qualified translator.
  • Safety and health information going to customers. Aftercare instructions, allergen details, product safety warnings. An error here harms someone. The tutorial on checking AI translations before customers see them sets out a review routine.
  • Anything where one word changes the money. Warranty conditions, cancellation terms, penalty clauses.

Checking a translation in a language you can't read

Back-translation is the practical check for a small business: translate the output back into English with a different tool, compare, and look hard at every difference. Use an approved chat assistant for the second step, with a prompt like this:

Below is a text in its original language and an English translation. Do not
retranslate the whole thing. Instead:
1. List every number, date, amount and deadline in the original text
   and confirm whether the English matches exactly.
2. List any words that could have more than one meaning in this
   context (legal, technical or commercial) and give the alternatives.
3. Flag any negations (not, unless, except, only if) and confirm they
   survive in the English.
4. Tell me which sentences you are least confident about.

[original text]
[English translation]

An illustrative response for the warranty letter might read:

1. Date 31.03.2027 = 31 March 2027. Matches.
2. "Vertragswerkstatt" is translated as "authorised dealer". It more
   precisely means a workshop under contract with the manufacturer,
   which can include authorised service partners that don't sell cars.
3. "nur bei" (only if) is preserved as "if". Consider "only if".
4. Least confident: the sentence on "Verschleissteile" (wear parts),
   which may exclude clutch components.

That response found two things that matter to the repair bill: "only if" is stronger than "if", and the clutch may be excluded as a wear part. Neither is a reason to trust the machine less; both are reasons to ask the warranty provider one precise question before quoting. Note the date check too: 31.03.2027 is unambiguous, but a date like 03.04 means 3 April in most of the world and March 4th in some systems, and a translation won't always flag it.

Replying to a client in their own language

The same rules apply when the text goes the other way, with one difference: the customer, not you, is now the person who can't check the machine. The garage's reply about the warranty shows the safer pattern. The owner's draft:

Before (as first written):
"Good news, your gearbox should be covered as long as the car was
serviced properly. We'll sort the paperwork."

After (rewritten for translation):
"The warranty letter says the gearbox is covered only if the car was
serviced at a workshop under contract with the manufacturer. Please
send us the stamped pages of your service book. We will then ask the
warranty provider to confirm cover before we start the repair."

The first version is friendly and almost untranslatable: "should be", "properly" and "sort the paperwork" each have several possible renderings, and one of them amounts to a promise. The second uses short sentences, one idea each, and repeats the key condition in plain words. Translate it with the approved tool, back-translate it as above, and add one line at the end in both languages: "This message was translated with software. If anything is unclear, please reply and we'll explain." That line costs nothing and turns a possible misunderstanding into a question.

If someone has already used the free version

It usually comes to light by accident. Say the florist next door to the garage asks how the owner translated that warranty letter so quickly, and it turns out the apprentice had been pasting customer letters, names and all, into a free translation website for months. Nobody meant any harm, and nothing may come of it, but it needs handling rather than ignoring.

  1. Find out what went in. Ask the person to list, as best they can, which documents and roughly how many. Look for identity numbers, health details or financial details, which raise the stakes.
  2. Clear what you can. If they used an account, delete the saved history. On a free website without an account there may be nothing to delete, and you should assume the text was kept for the period the vendor's policy states.
  3. Decide whether it's a reportable incident. Depending on what was shared and the data-protection law that applies to you, a disclosure to an unapproved processor may need recording or reporting. This is the point to ask your data-protection adviser rather than guess.
  4. Fix the cause, not the person. The apprentice used the free site because it was quick and nobody had said otherwise. Install the approved tool, remove the free apps, and pin the rule below where people work.

Record what happened and what you changed. If a customer ever asks, a dated note showing you found it, assessed it and fixed it is far better than silence.

A staff rule you can pin up

Translating customer or client documents
- Use only: DeepL Pro (log in first) or Translate in Word/Outlook.
- Never use: free DeepL, Google Translate website or app, browser
  translate buttons, or personal chatbot accounts.
- Remove names, addresses, ID numbers and account numbers first
  when you only need the meaning.
- Official documents, contracts and customer safety information:
  ask [name] before translating. These may need a human translator.
- Check every number, date and "only if / unless" in the result.

To see whether the rule is being followed, ask for one line in the customer record whenever a document is translated, for example "Translated 12/10: Word Translate, redacted copy, numbers and 'only if' checked, warranty provider asked about wear parts." After a month, pick five records with foreign-language documents. If the line is missing from three of them, the rule isn't in anyone's habits yet, and the usual cause is that the approved tool takes longer to open than the free one.

That's the whole policy for most small businesses. If translation is a bigger part of your work, the comparison of AI and human translation costs helps you decide which documents are worth paying a professional for.

Further reads

Sources: DeepL privacy policy and Pro data security page; DeepL Pro plan page (file translation limits); Google Cloud Translation data usage FAQ; Microsoft Translator No Trace page.

Need a safe translation setup for client documents?

On a 1:1 call we'll look at which documents you translate, which tools your staff already reach for, and set up an approved route with redaction and checking steps that fits your volume.

Book a 1:1 call with me