A generative AI policy is a list of failures you have decided to prevent. Write the failures down and the clauses follow.
Key takeaways
A generative AI policy needs eight working clauses: approved tools, prohibited inputs, vendor data handling, output verification, copyright and ownership, disclosure, records, and incident reporting. Each exists to prevent a specific failure, and the fastest way to write a good one is to name the failure first and let the clause fall out of it.
What separates a generative AI policy from a general acceptable-use policy is that generative tools do three things ordinary software does not: they absorb whatever you type, they produce assertions that look sourced but are not, and they emit content whose ownership and provenance are unsettled. Those three facts drive most of the document.
Approved tools. Failure prevented: a member of staff pastes a client brief into whatever was top of the search results. Name the tools, name the tenant or plan, and say plainly that the free web version of an approved product is not the approved product.
Prohibited inputs. Failure prevented: confidential material leaves the firm invisibly. Write categories a person can apply without asking: client identifying information, unreleased campaign or deal material, anything received under an undertaking, anything about an employee. Where the confidentiality obligation is contractual, the consequences are contractual too — see what a breach of a confidentiality clause exposes you to and how an NDA is actually drafted.
Vendor data handling. Failure prevented: discovering after a client asks that you cannot answer where their data went. The clause should require that, before a tool is approved, someone reads the vendor’s published position and files it — Microsoft’s data, privacy and security document for Microsoft 365 Copilot is an example of the kind of document you should expect to exist. If a vendor has no equivalent, that is your answer.
Output verification. Failure prevented: a confident fabrication reaches a client. Generative systems produce fluent text regardless of whether the underlying fact exists. The clause names who checks, what they check against, and what evidence of the check is kept.
Copyright and ownership. Failure prevented: shipping content whose provenance you cannot stand behind, or promising a client exclusivity you do not have. For an agency this is the sharpest clause in the document: it governs generated imagery, generated copy, style imitation and whether a client’s brand assets may be used as an input at all.
Disclosure. Failure prevented: a client discovering the method from someone else. Decide the firm’s default position, and decide it before a client asks in a meeting.
Records. Failure prevented: the work exists only inside a vendor workspace. Outputs that form part of a deliverable are saved to the client file in a format you control.
Incidents. Failure prevented: silence. A route to report a mistake within one working day, without a disciplinary framing on first report.
Generative tools are extremely good at producing marketing sentences that assert an outcome — a percentage lift, a comparison, an environmental benefit. Those are representations. The Competition Bureau enforces the Competition Act’s deceptive-marketing provisions, and a claim does not become safer because a model wrote it. The workable rule is that any AI-drafted claim about performance, comparison or environmental benefit must be traced to substantiation held on file before it is published, and the policy says who holds that file.
The agency writes its policy backwards, starting with a list of eight things it does not want to happen. Two are about client data, three about output quality, two about ownership of generated assets, one about a claim in a campaign.
Each becomes a clause with a named owner: the operations lead owns tools and vendors, the creative director owns ownership and provenance, the account lead owns disclosure and claims. Nothing in the policy is owned by “the team”, which is the same as being owned by nobody.
The test they apply before publishing it: can a new hire, in their first week, read this and know what to do with a client’s unreleased product brief? If not, it goes back.
Two published Canadian sources give the policy something to point at. The federal Voluntary Code of Conduct on Advanced Generative AI Systems commits its signatories to outcomes including transparency, human oversight and monitoring, and validity and robustness, and its measures table includes establishing policies, procedures and training. The Office of the Privacy Commissioner’s generative AI principles set out expectations for legal authority and consent, appropriate purposes and the rest — and are addressed to organisations using these systems, not only to those building them. Where personal information is involved, clause 4.1.4 of Schedule 1 to PIPEDA already requires policies, procedures and staff training.
How long should it be?
Two pages plus an appendix. Length is inversely related to compliance.
Do we need a separate policy per tool?
No. One policy, one appendix listing tools, with per-tool notes where a tool has an unusual condition attached to its approval.
What if we are a regulated professional firm?
Then this is the general layer and your regulator’s rules sit on top of it. See the model policy skeleton for the structure, and your own body’s conduct rules for the detail.
A 30-minute call is enough to tell you whether AI pays for itself here.
