Illustration: glass facade, building, architecture

Quebec Law 25 AI Compliance: A DPO’s Guide for Insurers in 2026

Quebec’s Law 25 requires insurers to tell a person when a decision about them rests exclusively on automated processing, explain the main factors on request, and let that person submit observations to a human who can review the decision. It also requires a privacy impact assessment before deploying AI that handles personal information, and before sending data outside Quebec.

Insurance runs on personal information: health histories, driving records, claims narratives, payment data. When an underwriting model, a claims-triage engine or a fraud score touches that data, Law 25 applies to the model as much as to the database behind it. The Commission d’accès à l’information (CAI) can now impose administrative penalties, and the AMF’s new AI guideline lands on the same teams in 2027. A Chief Privacy Officer or DPO who treats these as one program avoids doing the work twice.

Which Law 25 obligations actually reach your AI systems?

Law 25 amended Quebec’s Act respecting the protection of personal information in the private sector (CQLR c P-39.1) in three waves, according to BLG’s summary of the reform: the privacy officer and incident reporting duties in September 2022, most operational duties in September 2023, and data portability in September 2024. Four of those duties matter most for AI:

  • Automated decisions (section 12.1): notice, explanation on request, a right to correct the data, and a route to a human reviewer.
  • Privacy impact assessments (section 3.3): required for any project to acquire, develop or overhaul an information system that handles personal information.
  • Transfers outside Quebec (section 17): an assessment before personal information leaves the province, which covers most US-hosted AI APIs.
  • Governance: a designated person in charge of personal information protection, plus published governance policies.

The Act’s consolidated text on Légis Québec is the authoritative source for the exact wording; verify current text before you rely on any summary, including this one.

What does section 12.1 require when a model makes the call?

Section 12.1 applies when a decision is based exclusively on automated processing of personal information. The organization must inform the person concerned, and on request must disclose the personal information used, the reasons and principal factors that led to the decision, and the person’s right to have the information corrected. The person must also be able to submit observations to someone with authority to review the decision.

For an insurer, that means four things you can test today:

  1. Inventory every decision flow where no human reviews the output before it reaches the customer, such as instant quote declines or automated claim denials.
  2. Write the notice that tells customers an automated process made the decision.
  3. Make sure the model can produce the principal factors behind an individual outcome, not only global feature importance.
  4. Name the reviewer role that receives observations and can change the result.

A hypothetical example: a claims-triage model that auto-closes low-value glass claims is a section 12.1 decision. A model that only ranks claims for an adjuster, who then decides, is not exclusively automated, but only if the adjuster really reviews the case rather than rubber-stamping the score. Document that distinction, because the CAI will ask for evidence.

How should you run a privacy impact assessment on an AI project?

Section 3.3 requires the assessment to be proportionate to the sensitivity of the information, the purposes, the quantity and distribution of the data, and the medium. The person in charge of personal information protection must be consulted from the start of the project. The CAI publishes a guide to carrying out a privacy impact assessment (in French) that you can use as the template.

For AI, add four questions the standard template does not ask:

  • Training and fine-tuning data: which personal information, from which source, for which original purpose?
  • Vendor models: does the supplier retain prompts or use them to train, and where is it processed?
  • Output risk: can the model infer sensitive attributes such as health status from non-sensitive inputs?
  • Drift: who re-runs the assessment when the model, the data or the use changes?

Shadow AI makes this harder. An underwriter pasting a medical report into a public chatbot is an unassessed transfer of health data, so detecting Shadow AI is a precondition for an honest PIA register.

What do penalties and enforcement look like?

BLG’s summary reports two tiers: administrative monetary penalties of up to C$10 million or 2% of worldwide turnover, and penal fines of up to C$25 million or 4% of worldwide turnover, plus a private right of action for individuals. The summary notes that its table layout is hard to read, so verify the amounts against the current Act before quoting them to your board.

ObligationTrigger in an insurerEvidence the CAI would expect
Section 12.1 notice and reviewAutomated quote decline or claim denialDecision-flow inventory, notice text, reviewer log
Section 3.3 PIANew scoring model or claims chatbotDated PIA, privacy officer consultation record
Section 17 transfer assessmentUS-hosted LLM APIAssessment of the destination’s protection, contract terms
Governance policiesAny AI usePublished policy, role assignment

How does the AMF guideline change the picture for Quebec insurers?

On April 7, 2026, the Autorité des marchés financiers published its final Guideline for the Use of Artificial Intelligence, effective May 1, 2027, per Blakes’ analysis. It covers authorized insurers, financial services cooperatives, trust companies and deposit institutions in Quebec. It expects a centralized inventory of AI systems, a risk rating for each system, explainability tied to that rating, human review of outcomes for high-risk systems, and a senior manager accountable for AI. Customers must be told they are dealing with AI and be able to request human help promptly.

The overlap with section 12.1 is the opportunity. One inventory, with a risk rating and a privacy flag per system, feeds the AMF register, the PIA queue and the section 12.1 decision-flow review. Building three spreadsheets for three regulators is how evidence goes stale.

What are the Canada and US implications?

Quebec insurers operate inside a wider stack, and several layers apply at once.

  • Quebec: the CAI enforces Law 25, and the AMF supervises insurers’ conduct and, from May 1, 2027, their AI use.
  • Federal: Canada’s proposed Artificial Intelligence and Data Act died with Bill C-27 when Parliament was prorogued in January 2025, so no federal AI statute fills the gap. Federally regulated insurers should also check OSFI’s Guideline E-23 on model risk, which covers AI models; verify the current text and effective date.
  • United States: an insurer selling across the border faces state rules such as Colorado’s SB21-169 on unfair discrimination in insurance, and the NAIC’s model bulletin on AI use by insurers as adopted state by state; verify each state’s status.
  • Standards: ISO/IEC 42001 gives an auditable management system, and the NIST AI RMF gives the Govern, Map, Measure and Manage vocabulary that US supervisors and Canadian regulators both recognize.

Mapping each AI system once to these requirements, instead of per regulator, is the practical way to keep a Canadian PIA and a US state filing consistent.

What should a DPO do this week?

  1. List every AI system in underwriting, pricing, claims, fraud and customer service, including vendor-embedded models.
  2. Flag which ones make or drive decisions with no human review, and draft the section 12.1 notice for each.
  3. Check that each has a dated PIA, and queue the ones that do not.
  4. Confirm where each vendor processes data, and run the section 17 assessment for anything outside Quebec.
  5. Add a risk rating to the same register so it satisfies the AMF guideline before May 1, 2027.

Key takeaways

  1. Law 25 treats a model like any other personal-information system: section 12.1 requires notice, explanation and human review for exclusively automated decisions.
  2. Run the privacy impact assessment before deployment, and again before any data goes to a vendor model hosted outside Quebec.
  3. Keep one AI register with a risk rating and a privacy flag per system, so Law 25, the AMF guideline (effective May 1, 2027) and US state rules draw on the same evidence.

Next step: See how Fruggr tracks AI regulations and the evidence behind each system on the AI regulations and compliance page.

FAQ

Does Quebec Law 25 apply to AI systems?

Yes. Law 25 applies to any use of personal information, including training data, prompts and model outputs. Section 12.1 adds specific duties for decisions made exclusively by automated processing, and section 3.3 requires a privacy impact assessment for projects to acquire, develop or overhaul information systems.

What must an insurer tell a customer about an automated decision?

The insurer must inform the person that the decision was based exclusively on automated processing. On request, it must disclose the personal information used, the reasons and principal factors behind the decision, and the right to have the information corrected, and it must let the person submit observations to someone who can review the decision.

Is a privacy impact assessment mandatory for AI projects in Quebec?

Section 3.3 requires one for any project to acquire, develop or overhaul an information system involving personal information, and section 17 requires an assessment before communicating personal information outside Quebec. The depth must be proportionate to the sensitivity of the data. Verify current text with Légis Québec.

When does the AMF artificial intelligence guideline take effect?

According to Blakes, the AMF published the final guideline on April 7, 2026, and it takes effect on May 1, 2027 for authorized insurers, financial services cooperatives, trust companies and deposit institutions operating in Quebec.

This article was drafted with AI assistance and reviewed by our editorial team before publishing.