AI in financial services in Canada has moved from experiments to regulated production. Banks, credit unions, lenders, insurers and wealth managers use it to read documents, triage alerts, answer staff questions and prepare credit files. At the same time, regulators have set clear expectations: OSFI’s model risk guideline E-23 explicitly covers AI from May 1, 2027, and Québec’s AMF has published its own AI guideline with the same effective date.
So the question for a Canadian financial institution is not whether AI is allowed. It is which workflows to start with, how to govern the models, and how to get a first system into production without creating a model risk problem. This guide covers all three.
How are Canadian financial institutions using AI?
70%
The use cases that reach production share a pattern. They are high-volume, text-heavy, and they support a decision a qualified person still makes.
AML alert investigation
Transaction monitoring systems produce far more alerts than investigators can comfortably work. An AI assistant can gather the customer profile, transaction history, prior alerts and adverse media for each alert, summarize the activity, and draft the investigation narrative in your template. The investigator reviews, decides and signs. Reporting decisions under Canada’s anti-money laundering rules stay with people; the AI cuts the time spent assembling the file.
Commercial lending and credit files
Commercial credit analysts spend hours spreading financial statements, reading covenants and assembling credit memos. Document processing extracts figures from financial statements, tax returns and appraisals into your spreading template, with a confidence score per value and a link back to the source page. The analyst checks the numbers and writes the recommendation. Credit adjudication itself stays with your credit policy and your people.
Client onboarding and KYC
Onboarding files combine identity documents, corporate records, beneficial ownership information and source-of-funds evidence. AI can classify documents, extract the required fields, check them against what the client declared and list what is missing, so the onboarding team chases gaps on day one rather than week three.
Complaints handling
Complaints arrive by phone, email, branch and social channels. AI can classify each complaint by product and issue, detect ones that need escalation, track deadlines and draft acknowledgements, making it easier to meet the complaint-handling procedures regulators expect.
Staff knowledge assistants
Branch staff, advisors and contact centre agents need answers from hundreds of policies, product guides and procedures. An assistant that answers from your approved documents, cites the source and respects access permissions reduces escalations and inconsistent answers.
Forecasting and analytics
Cash and liquidity forecasting, deposit behaviour, delinquency trends and branch staffing all benefit from forecasting and analytics models built on your own data. These are classic quantitative models and fall squarely within model risk management.
What do OSFI and the AMF expect from AI in financial services?
| Guidance | Who it applies to | Status | What it means for AI |
|---|---|---|---|
| OSFI E-23, Model Risk Management | Federally regulated financial institutions | Final September 2025; effective May 1, 2027 | AI and ML models are in scope. Inventory, risk rating, validation, monitoring and governance across the lifecycle, proportionate to risk |
| OSFI B-10, Third-Party Risk Management | FRFIs | In effect since May 1, 2024 | AI and cloud vendors are third parties; the institution keeps accountability |
| OSFI B-13, Technology and Cyber Risk Management | FRFIs | In effect since January 1, 2024 | Security, resilience and governance of the technology AI runs on |
| OSFI E-21, Operational Risk and Resilience | FRFIs | Final August 2024; full adherence expected by September 1, 2026 | AI-dependent critical operations need resilience planning |
| AMF Guideline on the use of AI | Authorized insurers, financial services cooperatives, trust companies and deposit institutions in Québec | Final April 2026; effective May 1, 2027 | Board and senior management accountability, central AI inventory with risk ratings, bias monitoring, client explanations, disclosure and access to a person |
Privacy law sits underneath all of this: PIPEDA federally and Law 25 in Québec, with privacy impact assessments, consent rules, transparency for automated decisions and controls on transfers outside Québec. Our guide to Law 25 and PIPEDA AI compliance covers those obligations.
What E-23 means in practice
E-23 does not ban any technique. It expects you to know where your models are, how risky each one is and whether it still works. For an AI system, that translates into concrete engineering work:
- A model inventory entry for each AI component, including third-party models used through an API, with owner, purpose, data and risk rating.
- Documented design and data, including which documents the model reads, how prompts are built and what it is not allowed to do.
- Validation before use, on a test set drawn from your own cases, reviewed by someone independent of the build team for higher-risk uses.
- Ongoing monitoring of accuracy, drift, overrides and complaints, with thresholds that trigger review.
- Change control, so a new model version or prompt is tested against the same set before it goes live.
None of this is unusual for teams that build production AI properly. An evaluation suite and full logging are what make it cheap to comply.
How to design AI for a regulated institution
A few principles make the difference between a system that passes model risk review and one that stalls in it.
- Keep people accountable for decisions. AI prepares, summarizes, extracts and recommends. A named role decides, especially on credit, fraud, AML reporting and anything adverse to a client.
- Cite the source. Every summary or answer links back to the document, page or transaction it came from, so reviewers can check it in seconds.
- Host in Canada, in your control. Build in your own cloud account in Canadian regions, with private model endpoints that don’t train on your data. Our article on Canadian data residency for AI explains the architecture.
- Log everything. Inputs, retrieved documents, model version, output, reviewer and final decision, retained per your records schedule.
- Design for model change. Models improve every few months. An abstraction layer and a standing test set let you switch providers or versions with evidence, not hope.
- Scope permissions tightly. The system gets read access to exactly the data the workflow needs, and write access only where a person approves the action.
Our security and data handling practices follow these defaults, and we provide the documentation model risk and privacy teams need: design notes, test results, data flows and a privacy impact assessment.
What changes when AI faces clients directly?
Internal tools are the easier start, but most institutions eventually want AI in front of clients: a chat assistant on the website, help in the mobile app, or faster answers in the contact centre. The bar rises in four ways.
- Disclosure. The AMF guideline expects clients to be told when they are interacting with AI, and to get timely access to a person when they ask. That is a sensible default everywhere in Canada.
- Explanations. If AI contributes to a decision about a client, such as a fee reversal or a product eligibility check, the client should get a simple, clear explanation. Under Law 25, a decision based exclusively on automated processing also triggers notice and review rights.
- Scope control. A client assistant must stay inside approved topics, refuse to give personalized financial advice unless it is designed and supervised for that, and hand off to a person on complaints, vulnerability signals or anything it is unsure about.
- Fairness testing. Test outcomes across client groups before launch and monitor them afterwards, so a pattern of worse service for some clients is caught early.
Start with read-only answers from approved content, measure accuracy and hand-off rates, and add transactions only once the evidence supports it.
A realistic path to the first production system
| Phase | Weeks | What happens | Decision at the end |
|---|---|---|---|
| Discover | 1–2 | Choose one workflow, map it with the team that does it, gather a few hundred historical cases, agree on metrics and risk rating | Right use case, right risk tier? |
| Prototype | 3–5 | Working version on real historical data, measured against the test set; first model inventory entry | Accurate enough to save time safely? |
| Build | 6–11 | Integrations with core, CRM or case management systems, approval steps, logging, validation package, privacy impact assessment | Ready for validation sign-off and pilot? |
| Launch and improve | 12+ | One team first, monitoring dashboard, monthly reporting, then expansion | Expand, adjust or stop? |
Higher-risk uses take longer, mostly because independent validation takes time. Plan for that in the schedule rather than treating it as a surprise.
Where to start
If you are choosing a first project, favour one that is internal, document-heavy and measured today, with a person making the final call. AML investigation narratives, commercial loan spreading and complaint classification are common starting points because the time savings are visible within weeks and the model risk is manageable.
We build AI systems for banks, credit unions, lenders and wealth managers, with Canadian hosting, audit logs and model documentation designed in from the first sprint. To see what a first project could look like for your institution, try our free AI blueprint, or book a call with a senior engineer.
