Treadstone Associates
Article · Client reporting

AI translation for Quebec client work

Machine translation moves the first draft. The Charter decides which version the client is entitled to, and a bilingual professional decides whether it is right.

Treadstone Associates · Updated 2026

Key takeaways

  • • Translate with a tool, review with a person who practises in French — the review is the deliverable, not the draft.
  • • The Charter of the French language gives a client of an enterprise operating in Québec the right to be informed and served in French.
  • • Build a firm term list before you translate anything; terminology, not grammar, is where machine output fails a professional reader.
  • • Client documents carry personal information, so where the translation engine runs is a privacy question before it is a language question.

The short answer

You handle French client work with AI the same way you handle any other draft: a machine translation engine produces the first version, and a person who actually practises in French reads it against the source before it leaves the firm. The tool removes the typing; it does not remove the professional judgment, and it certainly does not remove the obligation to serve a Québec client in French in the first place.

What changes for Québec work specifically is that the French version is not a courtesy. Under the Charter of the French language, as the Office québécois de la langue française explains on its page on the language of commerce and business, a client of any enterprise carrying on activities in Québec has the right to be served and informed in French — before, during and after the transaction — and that page ties the right to sections 5 and 50.2 of the Charter. Translation quality is therefore a compliance question, not only a service question.

What the Charter asks of a firm, in practice

Three things matter to a professional-services firm. First, the right to be served in French applies to the enterprise, whatever its type — the OQLF page names retail, restaurants, financial institutions and service enterprises together. A firm in Toronto with Montréal clients is not outside the frame simply because its office is elsewhere.

Second, information supplied with a product must be in French; the same OQLF page attaches that rule to section 51 of the Charter. For a practice, the analogue is everything that travels with the deliverable — the covering letter, the appendices, the instructions for use of a model or template.

Third, if you belong to a professional order in Québec, French competence is a condition of the licence itself. The Government of Québec sets out the legal obligation to know French to become a member of a professional order, which it links to section 35 of the Charter. That is the honest reason a firm cannot outsource its French to a model: the person signing is required to have the language, and the file will eventually be discussed with someone in it. The province’s summary of the modernization of the Charter of the French language is worth reading once by whoever owns quality control in your firm.

What the engine is actually doing

A neural machine translation service is a sequence model, not a bilingual reader. Microsoft describes Azure Translator as a cloud-based neural machine translation service that supports text and document translation across its supported languages, and its own documentation is the right place to check which features and languages are in scope before you standardise on it. What no vendor page promises is judgment. The engine will happily render a defined term two different ways in the same report, because from its point of view both are acceptable French.

That is why the useful unit of work is not the document but the term list. Before a firm translates anything, it fixes the French rendering of the fifty or so terms that recur in its deliverables — the engagement types, the report headings, the standard caveats, the names of the schedules. Once that list exists, machine output becomes correctable at speed instead of debatable line by line.

Worked example: a report for a Montréal client

A practice with four Québec clients issues an English report and has historically sent French versions two weeks late, produced by whoever was free. The rebuild has four steps.

Step one, the term list: forty-two terms, agreed once by the two members who practise in French, stored as a glossary the translation service can consume. Step two, the source: the English document is rewritten so that the same concept always uses the same phrase, because inconsistent English is the largest single cause of inconsistent French. Step three, the machine pass, run per section rather than per document, so that a reviewer can compare paragraph to paragraph. Step four, the review, done by a French-practising member with the source beside the output.

The measurable change is not that the French appears faster, though it does. It is that the reviewer now spends her time on the eight paragraphs that carry professional meaning instead of retyping the thirty that carry none. And because the term list is fixed, next quarter’s report starts from an agreed vocabulary rather than from an argument.

The privacy question, before the language question

A client document sent to a translation service leaves your firm. Under Schedule 1 of the Personal Information Protection and Electronic Documents Act, clause 4.1.3, an organisation stays responsible for personal information transferred to a third party for processing and must use contractual or other means to provide a comparable level of protection. Translation is processing. The contract with the vendor is where that obligation is met, and it is settled before the pilot, not after.

For Québec clients there is a second regulator. The Commission d’accès à l’information du Québec oversees the Act respecting the protection of personal information in the private sector, and its own English page flags that the private-sector statute has been amended and points businesses to its guidance. Read that before routing Québec client files through any hosted engine. The Office of the Privacy Commissioner’s principles for responsible, trustworthy and privacy-protective generative AI apply to the firm using the tool, not only to the company that built it.

Where the human stays

Nobody signs a translation. A person signs the report, the letter or the opinion the translation carries, and the standard that applies to that person does not soften because the first pass was mechanical — see our sister firm’s note on the professional standard of care. Two rules keep that honest. Anything that creates an obligation — an engagement letter, a scope change, a limitation of liability — is drafted in French by a person, not translated into it. Everything else is machine-drafted and human-reviewed, and the review is recorded on the file.

If your interest is the intake and scheduling side of a bilingual practice rather than the deliverable, that sits on the practice-owner page instead.

Questions we get asked

Can we send the English version first and the French later?
Treat that as the wrong default. The OQLF describes the right to be informed and served in French as applying before, during and after the transaction, which is difficult to satisfy with a version that arrives a fortnight afterwards. Build the French into the delivery date rather than after it.

Does the Charter apply to my firm if we are not in Québec?
The OQLF page frames the obligation around enterprises carrying on activities in Québec, not around where the head office sits. If you are unsure whether your arrangement counts, that is a question for counsel and for the OQLF, not for a blog post.

Is machine translation good enough for a technical report?
For a first draft under review by a French-practising member, generally yes. As a finished deliverable, no — and the failure mode is not comic mistranslation but quiet inconsistency in defined terms, which a client’s own advisers will notice before you do.

See where AI pays off first in your firm.

A 30-minute call is enough to tell you whether AI pays for itself here.