Building AI for Quebec: Bill 96 & Law 25

Most teams building software for the Quebec market know they have a "French requirement." Fewer realize they are subject to two distinct regimes that AI products hit head-on.

The first is Law 25 — Quebec's modernized private-sector privacy law (formerly Bill 64). It governs how you collect, use, and disclose personal information, and it speaks directly to automated processing: if a decision about a person is based exclusively on automated processing, you owe that person disclosure and a path to human review. That is a rule written for the exact thing an AI model does.

The second is Bill 96 (enacted as Law 14), which amended the Charter of the French Language. It is not just about a translate toggle. It requires that French be available, and of a quality at least equal to any other language, in the products and services you offer to consumers and to your own employees in Quebec.

A chatbot that gives sharper answers in English than in French is a Bill 96 problem. A risk-scoring model that declines a customer with no French-language explanation and no human-review path is both a Bill 96 and a Law 25 problem. The two laws overlap precisely where AI lives — automated decisions delivered through language.

This is not a roadmap item for next year. The teeth are in.

Law 25 phased in over three years. Since September 22, 2023, the Commission d'accès à l'information (CAI) can pursue penal fines up to C$25M or 4% of worldwide turnover, whichever is greater, plus administrative monetary penalties up to C$10M or 2% of turnover — see Osler's summary of the Law 25 enforcement scheme. The automated-decision disclosure obligation and the data-portability right came in with that final wave.

Bill 96's final provisions came into force on June 1, 2025: francization is now mandatory for businesses with 25+ employees, with new French requirements on trademarks, signage, and packaging — Gowling WLG walks through what landed.

The regulatory pressure is not theoretical to the market either. Across the Deloitte enterprise surveys, regulatory compliance has become the #1 barrier to generative-AI adoption, rising from 28% to 38% of leaders — see Deloitte's State of Generative AI in the Enterprise. For a Quebec-facing product, that barrier is named and dated.

Strip away the legalese and Law 25 asks an AI feature for four concrete things.

1. Disclose automated decisions. When a decision about an individual is based exclusively on automated processing, you must inform them — at or before the decision — and, on request, tell them the personal information used, the reasons, and the principal factors that led to the decision. Practically: your model needs to emit an auditable explanation, not just a score.

2. Offer human review. The person can submit observations and ask a human to review the decision. That means a routing path and a staffed queue, not a dead-end "contact us" form.

3. Consent and purpose limitation. Personal information fed to a model must have been collected for a purpose the person consented to. Training a model on customer data "because we had it" is the failure mode the CAI is built to catch.

4. Breach handling. Confidentiality incidents involving a risk of serious injury must be reported to the CAI and to affected individuals. AI raises the stakes here: ungoverned, unsanctioned "shadow AI" added roughly US$670K to the average breach, and 97% of organizations that suffered an AI-related breach lacked proper AI access controls — see the IBM Cost of a Data Breach 2025 report. A model with quiet access to personal data is a breach surface.

None of this is exotic. It is logging, an explanation layer, a human-in-the-loop queue, and access control — designed in, not retrofitted.

Here is where most AI products quietly break Bill 96: they treat French as a post-launch translation job. The model is tuned, evaluated, and red-teamed in English; French gets a string-table pass and ships. The result is a product that is materially worse in French — slower, less accurate, occasionally nonsensical — which is exactly the "quality at least equal" bar Bill 96 sets.

Bilingual-by-design means French is a first-class locale through the whole stack:

• Routing and content are locale-aware from the first request (a Quebec visitor should land in French by default, not opt in).
• Prompts, system messages, and guardrails exist in French, not just translated UI chrome. An LLM prompted in English and asked to "reply in French" degrades; prompt it natively.
• Evaluation runs in both languages. If you measure hallucination, refusal, and task success only in English, you have no evidence your French output meets the equal-quality bar. Given that even purpose-built legal-AI tools hallucinated on 17–34% of hard queries, "it's probably fine in French" is not a defensible position.
• French is Quebec-authored, following OQLF conventions — not a literal calque of the English. A calque reads as foreign and undercuts the equal-quality claim.

The payoff beyond compliance: Quebec AI adoption is real and accelerating. Statistics Canada reports AI use by Canadian businesses roughly doubled year-over-year to 12.2%, with professional services and finance well above 30%. A product that is genuinely good in French is competing for that market, not just avoiding a fine.

Put the two laws together and a build pattern falls out. None of it is heavy; it just has to exist before launch, not after the CAI calls.

• Locale at the edge. Detect Quebec geo and default to French in middleware. Persist the choice. Every downstream service receives an explicit locale.
• Native bilingual prompts and content. Maintain EN and FR prompt sets and string catalogs with enforced key parity — a missing French string is a compliance gap, so make it a build-time failure.
• Decision logging. For any automated decision about a person, log the inputs, the model version, the principal factors, and the output. This is your Law 25 explanation on demand and your audit trail.
• Human-review routing. A real queue with a service-level target, surfaced in both languages, so "ask a human to review" is a working path.
• Output guardrails and evaluation gates. Block or flag low-confidence outputs; run hallucination and task-success evals in both languages on every model change. Map the program to a recognized framework — the NIST AI Risk Management Framework and ISO/IEC 42001 give you an auditable structure regulators and clients recognize.
• Access control on data feeding the model. Given that 97% of AI-breach victims lacked it, scoped credentials and logging on the model's data access are table stakes.

This is the work behind Maverin's AI Security & Governance and Agentic Automation engagements — and it is why we run bilingual EN/FR by default rather than as an add-on.

Quebec compliance is easy to get wrong from a distance. The French-quality bar is cultural, not mechanical; the privacy regime has its own enforcement body and its own definitions. Building for it well usually means having people who read the OQLF guidance in the original and have shipped under Law 25 before.

That is a deliberate part of how Maverin is set up. We run a Montreal office alongside Ottawa, Toronto, and Calgary, and we staff senior, AI-literate, bilingual people onto the work — whether that is staff augmentation inside your team or a fixed-fee engagement to get a feature over the compliance line. There is also a structural angle worth naming for public-sector buyers: as an Indigenous-owned firm, Maverin counts toward the federal 5% mandatory minimum for Indigenous procurement.

The practical takeaway is simpler than the regulation: in Quebec, French and privacy are not features you finish a product and then add. They are constraints you design from. Teams that internalize that ship once; teams that don't ship twice — and the second build, under deadline and scrutiny, is the expensive one. Start with a scoped Discovery Assessment if you want a read on where your current AI surface stands against both laws.