Skip to content
Quebec

The AMF Guideline on AI Use by Financial Institutions

The AMF AI guideline, adopted in March 2026 and effective 1 May 2027, sets Quebec's expectations for how insurers, financial services cooperatives, trust companies and deposit institutions govern artificial intelligence. It requires an AI inventory, risk ratings, bias controls, accountable senior management and clear disclosure to clients.

Insurance risk team mapping AI systems to meet the AMF AI guideline in Quebec

Key takeaways

  • The AMF adopted its Guideline for the Use of Artificial Intelligence on 30 March 2026.
  • It takes effect on 1 May 2027, the same day as OSFI Guideline E-23.
  • It covers Quebec-authorized insurers, cooperatives, trust companies and deposit institutions.
  • Every AI system should be inventoried, risk-rated and controlled in proportion to its risk.
  • Clients must be told when they interact with AI and be able to reach a human.

What the AMF AI guideline is and when it applies

The AMF AI guideline is the Guideline for the Use of Artificial Intelligence issued by Quebec's Autorité des marchés financiers (AMF). It sets out the AMF's expectations for how Quebec-chartered and authorized financial institutions govern, build, buy, deploy and monitor AI systems. The AMF adopted it by decision dated 30 March 2026 and published it in its Bulletin on 2 April 2026. According to the AMF Bulletin notice, it takes effect on 1 May 2027, following a public consultation held from 3 July to 7 November 2025.

Insurance risk team mapping AI systems to meet the AMF AI guideline in Quebec
The AMF expects institutions to inventory, risk-rate and monitor every AI system they use, with clear accountability at senior management level.

The guideline applies to authorized insurers, financial services cooperatives (including the Desjardins network), authorized trust companies and other authorized deposit institutions. It is issued under the AMF's guideline-making powers in each sector statute: s. 463 of the Insurers Act, s. 565.1 of the Act respecting financial services cooperatives, s. 42.2 of the Deposit Institutions and Deposit Protection Act, and s. 254 of the Trust Companies and Savings Companies Act.

Like other AMF guidelines, it is principles-based and not a regulation in the strict sense, but the AMF uses guidelines as the benchmark for its supervisory assessments of sound and prudent management. Institutions are expected to apply it proportionately to their nature, size, complexity and risk profile. It supplements the AMF's existing Model Risk Management Guideline rather than replacing it.

Who is covered, and who is not

The AMF AI guideline is a prudential instrument for regulated financial institutions. It does not directly bind securities dealers, advisers or investment fund managers, which are supervised under securities law and the Canadian Securities Administrators' guidance on AI; nor does it bind federally regulated banks and insurers, which answer to the Office of the Superintendent of Financial Institutions (OSFI). The table below summarizes the landscape as of November 2026.

InstitutionPrimary AI expectationsStatus
Quebec-authorized insurers, financial services cooperatives, trust companies, deposit institutionsAMF Guideline for the Use of Artificial Intelligence, with the Model Risk Management GuidelineAdopted 30 March 2026; effective 1 May 2027
Federally regulated banks and insurersOSFI Guideline E-23 (Model Risk Management), which covers AI and machine-learning modelsEffective 1 May 2027
Registered dealers, advisers and fund managersSecurities law and CSA guidance, including CSA Staff Notice 11-348Staff notice published 5 December 2024
All private-sector organizations in QuebecLaw 25 rules on automated decisions and privacy impact assessmentsIn force

Many groups operate on both sides of this line, for example a federally chartered bank with a Quebec-authorized insurance subsidiary. Those groups should map which entity uses which system and align the AMF and OSFI E-23 frameworks, which share an effective date. Our page on insurance regulatory work describes the supervisory context.

Governance expectations under the AMF AI guideline

Governance sits at the centre of the AMF AI guideline. The board of directors is expected to ensure that senior management promotes a corporate culture of responsible AI use, to receive regular information on AI trends and risks, and to have sufficient collective expertise, particularly where AI supports critical activities.

Senior management is expected to maintain adequate governance and controls across the AI lifecycle, keep staff knowledge current and set validation frequency in light of the technologies used. The guideline expects a designated senior manager to be accountable for all AI systems and owners to oversee each system throughout its lifecycle. In practice, that means a named executive owner, a documented framework, and a cross-functional committee bringing together risk, compliance, data, technology and business lines.

Boards should treat this as part of their broader oversight of technology risk. Our insight on board oversight of AI risk suggests how to structure reporting so directors can discharge that role.

Inventory, risk rating and lifecycle controls

The operational core of the guideline is a risk-based approach applied system by system:

  • Identification and inventory. Institutions should regularly identify all models and AI systems. Those posing a non-negligible risk belong in a centralized register, such as the model inventory, with key information kept up to date.
  • Risk rating. Each AI system needs a risk rating, reviewed regularly; provisional ratings are acceptable at the planning stage. Factors include the underlying model risk rating, quantitative impacts (operational, financial, security) and qualitative factors such as the system's degree of autonomy and compliance risk.
  • Proportionate controls. The rating drives the intensity of validation, documentation, approval, monitoring and review.
  • Lifecycle. Expectations cover the justification for using AI and periodic reassessment of alternatives, selection (including explainability and cybersecurity), validation, approval, deployment, monitoring and decommissioning.
  • Data. Data quality must be managed in training and in use, across primary and secondary, public and private, synthetic, structured and unstructured data, with the origin of secondary data documented.
  • Third parties. Validation should cover third-party components, including platforms, foundation models and data sets.
Practical point

The inventory is where most institutions will start, and where gaps appear first. AI is embedded in vendor products, from fraud tools to claims triage and call-centre software. Ask procurement and IT to identify AI functionality in existing contracts, not only in-house models, and update contract templates so vendors must disclose AI components, data sources and material changes.

Fairness, explainability and transparency to clients

The guideline links AI risk management to the fair treatment of clients, which is a long-standing AMF supervisory priority. Its main client-facing expectations are:

  • Bias. For each system, identify prohibited factors and proxy variables, communicate them to stakeholders, correct problems promptly and document the fixes. Where a system affects decisions about clients, identify and monitor groups at risk of bias.
  • Explainability. Set explainability requirements when selecting a system, calibrated to its purpose, autonomy, regulatory context and impact on clients, and assess before deployment whether its outputs can be explained to those concerned.
  • Disclosure. Tell clients when they are interacting with an AI system, whether by chat, audio or video, and let them reach a human representative on request. Clearly label AI-generated content. Where a system makes or recommends a decision about a client, explain it clearly.
  • Human oversight. Consider a human-in-the-loop constraint for high-risk systems, particularly where evaluation data is incomplete.

These expectations overlap with s. 12.1 of Quebec's private-sector privacy act, introduced by Law 25, which requires notice when a decision is based exclusively on automated processing of personal information. The AMF expectations go further: they apply to AI-assisted decisions and to client interactions, not only to fully automated decisions. See our insight on automated decisions under Law 25.

Generative AI and foundation models

The guideline is technology-neutral. It does not create a separate regime for generative AI, but its definitions cover foundation models, it treats hallucination as a risk that should trigger validation work, and its content-labelling expectation applies to generated text, images and audio presented to clients. Institutions deploying assistants for advisers, claims handlers or call-centre staff should therefore rate those tools, test them for accuracy and bias, control what data they access and decide when outputs must be reviewed by a person. Internal acceptable-use policies for employees remain necessary; our insight on generative AI policies for employees covers that layer.

Vendor contracts under the AMF AI guideline

Few institutions build all of their AI in-house. Credit scoring, fraud detection, underwriting support, document extraction and conversational tools are usually licensed, and many core platforms now embed AI features that are switched on by default. Because the AMF AI guideline applies to any use of AI systems by the institution, outsourcing a system does not outsource accountability. Contracts should give the institution what it needs to rate, validate and monitor the system:

  • disclosure of AI components, the models and foundation models used, and material changes to them;
  • information on training data sources, known limitations and performance metrics;
  • cooperation with validation, bias testing and audits, including by the AMF;
  • controls on the vendor's use of the institution's data, including for further model training;
  • incident notification, service levels and exit assistance that allow the system to be replaced.

These terms sit alongside existing outsourcing and information-security expectations. Our insight on AI vendor contract clauses sets out drafting options.

How the AMF is likely to supervise compliance

The AMF has framed the guideline as part of its prudential supervision of sound and prudent management practices. Supervisory reviews of model risk management typically start with the framework documents, the inventory and a sample of models, then test whether ratings drive the controls that the framework promises. Institutions should expect the same pattern for AI: requests for the governance framework and board reporting, the AI inventory and rating methodology, validation reports for higher-risk systems, evidence of bias testing, and samples of client disclosures and complaint handling.

The AMF AI guideline also connects to market-conduct expectations. If an AI system leads to unfair outcomes for clients, the issue can be pursued as a fair treatment problem and not only as a prudential deficiency. Documenting design choices, testing and remediation is therefore as important as the controls themselves. Institutions that already comply with OSFI E-23 at group level should map the AMF expectations onto that framework rather than building a parallel one, adding the client-transparency elements that are specific to the AMF.

Preparing for 1 May 2027: a roadmap

  1. Appoint the accountable executive and confirm board reporting lines and committee mandates.
  2. Build or extend the inventory to cover in-house, vendor and embedded AI, with owners and use cases.
  3. Adopt a risk-rating methodology aligned with the existing model risk framework, and rate each system.
  4. Close control gaps for high-rated systems first: validation, documentation, monitoring and human oversight.
  5. Update client-facing practices: AI interaction notices, routes to a human, labelling and decision explanations.
  6. Review vendor contracts for audit, transparency, data and change-notification rights.
  7. Test fairness on decision-making systems and document the prohibited factors and proxies.
  8. Train staff and directors, and document the programme for AMF supervisory review.

Illustration: a hypothetical Quebec property and casualty insurer uses a vendor model to flag suspicious claims and a chatbot for first notice of loss. Under the AMF AI guideline, both go into the inventory. The fraud model, which affects client outcomes, likely receives a high rating, requiring bias testing for proxies such as postal code, independent validation and documented human review before a claim is denied. The chatbot must tell customers they are dealing with AI and let them reach an agent.

How TRW & Co helps

We advise insurers, cooperatives, trust companies and fintech partners on AI governance in Canada, including the AMF AI guideline, OSFI E-23 and privacy-law obligations. We help design governance frameworks and board reporting, draft AI policies and risk-rating methodologies, review vendor contracts and client disclosures, and prepare institutions for supervisory reviews. This article is general information, not legal advice. To discuss your AI programme, contact us.

Frequently asked questions

When does the AMF AI guideline take effect?

The AMF adopted the Guideline for the Use of Artificial Intelligence by decision dated 30 March 2026 and published it in its Bulletin on 2 April 2026. It takes effect on 1 May 2027, which gives institutions roughly thirteen months to build the required governance, inventory and controls.

Does the AMF AI guideline apply to banks?

Not to federally regulated banks, which are supervised by OSFI and subject to its Guideline E-23 on model risk management. The AMF guideline applies to authorized insurers, financial services cooperatives, authorized trust companies and other authorized deposit institutions under Quebec law.

Is the guideline legally binding?

AMF guidelines are issued under the sector statutes and set out expectations for sound and prudent management rather than prescriptive rules. They are nonetheless the reference point for AMF supervision, and failure to meet them can lead to supervisory findings and, where statutory duties are breached, enforcement measures.

Does the guideline cover vendor AI tools?

Yes. It applies to any use of AI systems by the institution, and validation should cover third-party components such as platforms, foundation models and data. Institutions should identify AI embedded in vendor products, rate those systems and secure the contractual rights needed to oversee them.

What must clients be told?

Clients should be informed when they are interacting with an AI system, be able to reach a human representative on request, see clear labels on AI-generated content, and receive a clear explanation when an AI system makes or recommends a decision about them. Law 25 adds separate notice duties for exclusively automated decisions.

Sources

  1. AMF, Guideline for the Use of Artificial Intelligence (page updated 24 April 2026)
  2. AMF Bulletin, vol. 23, no. 13 (2 April 2026): Decision 2026-PDG-0016 and guideline text
  3. AMF, AI in the financial sector
  4. OSFI, Guideline E-23: Model Risk Management (effective 1 May 2027)

General information only, not legal advice. References to statutes, courts and authorities describe published law and practice at the date shown.