Agencies use AI in two places: inside delivery, on research, first drafts, variants and reporting commentary; and inside their own economics, on scoping, utilisation and quality control. What does not change is who answers for a claim that reaches the public — under the Competition Act, the proof for a performance claim lies on the person making the representation.
Key takeaways
Research synthesis is the least glamorous and most reliable win: turning a pile of transcripts, reviews, support tickets and competitor pages into a structured summary a strategist argues with. First drafts of anything structured — briefs, outlines, ad variants, alt text, meta descriptions, changelogs — come next. Reporting commentary is third: the numbers come from the platform, the paragraph explaining what moved comes from a model, and an account director corrects it.
What does not survive contact with a client: strategy, media allocation decisions, the client relationship, and the final read before anything is published. Those are the things clients are actually paying an agency for, and they are also the things where a confident wrong answer costs the most.
Scoping is the highest-value internal use, because agencies systematically underestimate revisions. A model reading two years of your own job records can tell you what a job of this shape has actually cost, which is a better starting point than what the last proposal said. Utilisation and capacity questions belong here too, and the capacity guides in our Business & Professional Services hub cover them.
Quality control is the other one. A model can check a deliverable against a house checklist before a human sees it: every claim has a source line, every asset has a licence, every link resolves, the client’s banned words are absent. That is a mechanical pass, and mechanical passes are exactly what should be automated.
Canadian agencies underrate how directly the Competition Act reaches them. Section 74.01 makes it reviewable conduct for a person, for the purpose of promoting the supply or use of a product or promoting any business interest, by any means whatever, to make a representation to the public that is false or misleading in a material respect. It goes further for performance claims: a representation in the form of a statement, warranty or guarantee of the performance, efficacy or length of life of a product must be based on an adequate and proper test, the proof of which lies on the person making the representation.
Since the 2024 amendments the same section carries two environmental provisions: a claim about a product’s benefits for protecting or restoring the environment or mitigating climate change must be based on an adequate and proper test, and a claim about the benefits of a business or business activity for the environment must be based on adequate and proper substantiation — in both cases with the proof on the person making the representation. If your agency writes sustainability copy for clients, that is the sentence to put on the wall.
Then the trap that catches creative work specifically. Under section 74.03(5), the general impression conveyed by a representation as well as its literal meaning must be taken into account. A generated hero image that shows a product doing something it does not do is a problem even when every word of the body copy is true.
The exposure is real: section 74.1 allows a court to order a corporation to pay an administrative monetary penalty of the greater of $10,000,000 — $15,000,000 for each subsequent order — and three times the value of the benefit derived from the deceptive conduct, or three per cent of annual worldwide gross revenues where that value cannot reasonably be determined. Our sister firm covers how the misleading advertising provisions apply to an Ontario business and the narrower rules for comparative advertising.
Disclaimers do not fix a false general impression. What does fix the process is a file that travels with the deliverable. Every factual or performance claim in it gets four fields: the claim as published, who supplied it, what test or evidence it rests on, and the date it was verified. AI drafts that file from the deliverable in seconds. A named person signs it, and the client countersigns for claims the client supplied.
The countersignature matters. When a client says its product cuts costs by a third, the agency that publishes the claim is a person making a representation. Getting the client’s substantiation in writing before publication is cheap; getting it afterwards is a dispute.
Ad Standards administers the Canadian Code of Advertising Standards as the self-regulatory framework for advertising in Canada, with a published complaint procedure and preclearance for categories including alcoholic beverages, children’s advertising, cosmetics, food and non-alcoholic beverages, and health products. If your agency works in a precleared category, build that step into the schedule rather than discovering it at the end.
A twelve-person agency introduces one rule before it introduces any tool: no deliverable leaves without a substantiation file. For the first month the file is written by hand, which is unpleasant and instructive — it surfaces how many claims in a typical deliverable have no identified source at all.
Only then does AI enter, doing three things: extracting every claim from the draft into the file template, flagging the ones with no source, and drafting the first version of the reporting commentary. The account director fills the gaps and signs. Client-supplied claims go back to the client for countersignature before publication.
The agency measures one number for a quarter: unsourced claims found at final review. The point of the programme is to drive that to zero, and the AI is only there to make the mechanical part cheap enough to do every time.
Is it the agency or the client who is liable?
Section 74.01 speaks of a person making a representation to the public, which is not limited to the brand owner. Treat it as a shared risk, allocate it deliberately in the contract, and get advice on your own arrangements rather than assuming the client absorbs it.
Do we have to disclose AI use to clients?
That is a contract question rather than a statutory one, and it is better settled in the master services agreement than in an awkward email. Clients increasingly ask; a clear written position is an asset.
What about the in-house marketing team?
Different reader, different page — see marketing automation. This article is about an agency delivering to a book of clients.
A 30-minute call is enough to tell you whether AI pays for itself here.
