No regulator publishes a model policy for your firm. Two Canadian documents give you a structure that will survive being questioned.
Key takeaways
No Canadian regulator publishes a model AI use policy for professional-services firms, and you should be suspicious of anyone who says otherwise. What does exist are two public documents worth building on: the federal Voluntary Code of Conduct on the Responsible Development and Management of Advanced Generative AI Systems, and the Office of the Privacy Commissioner’s principles for responsible, trustworthy and privacy-protective generative AI. Between them they supply a structure and a privacy backbone.
Be precise about what the code is. It sets out commitments for organisations that develop or manage advanced generative systems, and it names its signatories publicly — among them CIBC, TELUS, Mila, the Vector Institute and Cohere. A ten-partner firm using a commercial assistant is not a signatory and should not imply that it is. What transfers is the shape of the thinking, not the badge.
The code commits signatories to work towards six outcomes. Read as headings for your own policy, they map cleanly onto professional practice.
Accountability — the code describes organisations understanding their role and putting risk-management systems in place, and its measures table includes establishing policies, procedures and training so staff are familiar with their duties. In a firm: one named owner, a tool register, and training that happens.
Safety — risk assessment before deployment. In a firm: no tool touches client material until someone has written down what could go wrong and who would notice.
Fairness and equity — impacts assessed across development and deployment. In a firm: this is the clause that governs anything touching recruitment, evaluation or client selection, and the answer is almost always that a person decides.
Transparency — sufficient information published for informed decisions. In a firm: your position on telling clients, and on labelling AI-assisted work internally.
Human oversight and monitoring — use monitored after deployment. In a firm: the sign-off step, plus somebody actually reading the exception queue.
Validity and robustness — systems operating as intended and secure against attack. In a firm: verification of output, and information-security review of the vendor.
A defensible policy for a professional firm runs to about two pages and contains ten numbered clauses: 1. Purpose and scope (who it binds, including contractors); 2. Approved tools (a dated appendix, not prose); 3. Prohibited inputs (categories, in plain language); 4. Client confidentiality and consent (what the engagement letter says, and when you ask); 5. Verification and sign-off (a named reviewer per output type); 6. Records and retention (where output lives and for how long); 7. Privacy and vendors (contract terms, processing location); 8. Incidents (how to report without blame); 9. Training (who, what, how often); 10. Ownership and review (a name and a date).
Two clauses are commonly missing and both are load-bearing. Contractors and locums are inside the scope or they are a hole. And “review” needs a trigger list — a new tool, a regulator publication, an incident — because annual review alone means a policy that is eleven months stale on the day it matters.
Week one is discovery, not drafting: ask everyone what they already use, with an amnesty. Week two, the two partners decide the three hard questions — consumer tools yes or no, client-identifying information yes or no, and who signs off what. Week three, the policy is written around those decisions and circulated. Week four, it is trained and acknowledged in writing.
The order matters. Firms that draft first produce a document that describes a firm they do not have; the decisions then get made informally afterwards and the policy is quietly wrong from day one.
Privacy is the part with statutory teeth. Schedule 1 to PIPEDA requires, at clause 4.1.4, that organisations implement policies and practices to give effect to the principles — including procedures to protect personal information, procedures for complaints, training staff, and explaining the policies. At clause 4.1.3 it keeps the organisation responsible for personal information transferred to a third party for processing. An AI policy that ignores privacy is missing the only part of itself that is already mandatory. If nobody in the firm owns that, start with appointing a privacy officer and a real privacy policy.
Employment law is the other connection. A policy binds staff because it is part of the employment framework, which is why it belongs in the handbook — see workplace policies and the employee handbook. Ontario firms should also check whether they are caught by the written electronic monitoring policy requirement, because AI tools that log activity can turn a monitoring question into a live one.
Can we adopt a US firm’s policy?
Not without rewriting it. The privacy statute is different, professional conduct is provincial, and the employment framework is different. The structure may travel; the citations do not.
Should the policy name specific vendors?
Yes, in the appendix. A policy that describes categories of tool cannot be complied with, because staff cannot tell which category a new product falls into.
Is a voluntary code worth referencing at all?
As structure, yes — it is a published federal document with named signatories and a measures table. As a claim about your firm, no. Say what you have adopted, not what you resemble. Sector-specific detail belongs in a sector policy: see the CPA firm version and the law firm version.
A 30-minute call is enough to tell you whether AI pays for itself here.
