Agencies use AI at two points in the content pipeline: turning a brief into an outline and a first draft, and turning a finished piece into the eight other formats it should always have had. Everything in between — the angle, the facts, the sources and the sign-off — is where both the value and the liability live.
Key takeaways
A content programme has nine stages: brief, research, outline, draft, fact-check, edit, sign-off, publish, repurpose. AI is genuinely good at research synthesis, outlining, first drafting and repurposing. It is mediocre at editing to a house voice, and it must not be anywhere near sign-off.
The brief is the stage everyone underinvests in and the one that determines whether any of the rest works. A brief that names the reader, the single question the piece answers, the sources that are acceptable, and the claims that must not be made will produce a usable draft. A brief that says “write about AI in construction” produces exactly what you would expect.
One rule, applied without exception: every factual claim carries a source that a person opened. Not a plausible-looking citation, not a remembered statistic — a source someone fetched. Models produce confident, well-formed figures that do not exist, and they do it most often for precisely the kind of claim a marketer most wants: a market size, a percentage improvement, a survey result.
There is a legal edge to this, not just an editorial one. Where the claim concerns the performance, efficacy or length of life of a product, section 74.01 of the Competition Act requires it to be based on an adequate and proper test, and states that the proof lies on the person making the representation. An unsourced performance claim in a client’s blog post is not a copy problem; it is an evidentiary one.
Content programmes quote, excerpt and illustrate constantly, and the rules do not relax because a model assembled the paragraph. Our sister firm sets out how fair dealing works for business use in Canada, and separately the copyright risks of using stock photos. The practical control is a provenance column in the content calendar: for every asset and every quoted passage, where it came from and on what terms.
Decide before the first draft whether the client is getting an assignment of copyright or a licence, and say which in the statement of work. Section 13 of the Copyright Act makes the author the first owner and requires an assignment to be in writing signed by the owner of the right, so the default is not what most clients assume. The difference between licensing and assigning copyright is worth reading before the contract is drafted, particularly where freelancers wrote part of the work.
The newsletter that carries the content is governed by Canada’s anti-spam legislation regardless of how the content was written. Under section 6, a commercial electronic message may not be sent without consent, express or implied, and must identify the sender, provide contact information valid for at least sixty days, and include an unsubscribe mechanism. The CRTC publishes its guidance on the legislation, and the practical questions — implied versus express consent and what the unsubscribe mechanism must do — are covered by our sister firm.
It is worth being blunt about this, because agencies are being sold the opposite. Publishing more pages than you can check is how a content programme becomes a liability: unsourced claims accumulate, near-duplicate pages compete with each other, and the client’s site fills with material nobody at either organisation has read. Search engines treat scaled, unchecked content as a quality problem in its own right, and a client who discovers an invented statistic under their own logo will not be reassured that a tool wrote it.
The design consequence is simple: size the programme to your review capacity, not your generation capacity. If one editor can properly check eight pieces a month, the programme is eight pieces a month, and the AI is there to make those eight better rather than to make them eighty.
An agency producing a monthly content set for a professional-services client rebuilds the pipeline around the gate. Briefs are written by the strategist and include a permitted-source list. A model produces research synthesis and a first draft against the brief. The writer rewrites to house voice — not edits, rewrites, because editing a machine draft to sound human takes longer than rewriting it.
The fact-check step is a separate pass by a different person, working from a claims table the model extracts from the draft: claim, source, link, date checked. Anything without a source is cut, not softened. Sign-off is the account director, by name, in the calendar.
Repurposing happens only after sign-off, from the approved text, which is the safest AI task in the whole pipeline because the facts have already been settled.
Will AI-assisted content rank?
The useful framing is not the tool but the check: unchecked, near-duplicate pages at volume are a known quality problem, and no amount of production fixes that. Content that answers a real question, carries real sources and was read by a person before publication is the version that works.
Should we tell readers the piece was AI-assisted?
That is a client decision and belongs in the statement of work. What is not optional is that a named person stands behind the facts.
Where does the strategy work fit?
Ahead of all of it, and it stays human. The delivery-wide view sits across our Business & Professional Services hub.
A 30-minute call is enough to tell you whether AI pays for itself here.
