Where AI belongs in a creative system

Generative tools have collapsed the cost of producing a video, and if creative is a cost line then the objection is right, it is heading towards free. The cost of making an ad has fallen, and the cost of running one has not, because the auction prices every impression on whether the person it reaches responds to the ad, and nothing about how the ad was produced changes that test. So the useful question is where AI belongs in a system whose job is making ads work, and the answer is specific, four places where it multiplies decisions already made, and one open question held on evidence.

Four places it belongs

Analysis across the portfolio. Twenty to thirty concepts live, each with a stop rate, hold rate and watched rate per week, and the question every Monday is which entry points, which faces and which openings are earning attention. That is pattern-finding a model does well, because the data is observed behaviour rather than opinion.

Pretesting at speed. Every ad scored against the brand's written standard before spend attaches, codes present and early, claims in bounds, legible with sound off, safe zones clear, failures returned as actions. A panel cannot do that twenty times a week and a system can, and What pretesting can tell you, and what only the feed can sets out why the attention question still goes to the feed.

Versioning and adaptation of proven work. The most underused asset in most accounts is their own footage. An execution that earned its watched rate can be given a new opening, a new length, a new format or on-screen text in another language, and each is a real test of what carries the idea. This is the remix move in the weekly loop.

Briefing and research. Category entry points mapped, scripts and storyboards drafted, briefs generated from last week's evidence, research synthesised. Each of the four multiplies a decision a strategist has already made, and none of them decides what deserves multiplying. A hundred variations of an idea nobody validated is the same bet placed a hundred times, faster.

What it does not change

The test. An ad earns attention in the first second or two from a person who owes it nothing, and Meta's auction ranks every ad on bid, estimated action rate and quality, so an ad people pass costs more to run for as long as it runs. Feed audiences classify what they see by texture before any message lands, and today synthetic people still read as off. Kantar's LIFT ROI work across 56,984 creatives, commissioned by Meta and modelled, indexed ads with a single human making eye contact 81% higher on effectiveness than ads without the feature, and a distinctive atmosphere 67% higher. YPulse's survey of young consumers, cited by Meta, had 84% agreeing they like it when content from brands is not perfect. In categories built on a demonstrable result, skin, food, a claim demonstrated on a face that does not exist has cancelled itself. Real people in real environments are the moat for now, and the position is held on evidence, not conviction.

Four conditions, each moving

Disclosure regulation, and what a label does to feed behaviour. Platforms and regulators are moving towards labelling AI-generated content, and whether a label reads as a neutral notice or a skip trigger is being settled in the data now, market by market.

Platform treatment of labelled synthetic content in delivery. The platforms decide what their systems do with it, rank it neutrally or hold it back, and they are invested in generation tools and in feed quality at once.

Audience response in the brand's category. A generated kitchen behind a real presenter is a different question from a generated face making a skincare claim.

The evidence on watched rates. Synthetic against real, like for like, in the account's own stop, hold and watched rate, tested in bounded cells inside the portfolio rather than argued in meetings. This is the condition that decides. When the conditions move, the position moves. Until then the test is applied blind to production method, an asset clears the floor and earns its watched rate or it does not.

The automation view

The counter-position holds that generative production turns the creative problem into a cost problem, and it is right about something real. A portfolio of twenty to thirty distinct ads was unaffordable for most brands a decade ago, and anything that lowers the cost of a variation lowers the cost of a test. Where it stops is at the constraint that binds on a skip medium, which was never the cost of making the ad. It is the attention of the person it reaches, and the auction charges for its absence at full rate whether the ad cost a crew or a prompt. Media creates the opportunity to see your ads. Creative determines whether they work. AI has changed how fast a brand can act on both sentences, it has not changed either of them.

What to do with this

Write the policy in one paragraph you can defend to a board next quarter. AI is used wherever it multiplies a decision already made, analysis across the portfolio, pretesting every ad against the written floor, versioning proven footage, briefing and research. Fully synthetic people stay out of the brand's paid advertising until the account's own evidence says otherwise. Then build the evidence. Run synthetic variants in bounded cells against real ones on the same media plan, put the stop, hold and watched rate readings in the weekly review beside everything else, and set a date to revisit the four conditions. What it takes to run a creative system is the natural next read, it covers the people and the hours the tools do not remove.

Questions

Should we be using AI-generated people in our ads?

Not yet in the work that has to earn attention, and the reason is evidence rather than principle. Feed audiences classify what they see by texture in the first second or two, synthetic people still read as off to them, and in categories built on a demonstrable result a generated face undermines the claim it is making. Where AI already clears the bar, backgrounds, variations off real footage, localisation, product swaps, use follows.

Hold the position as a test rather than a rule. Run synthetic variants in bounded cells against real ones on the same media plan, read the stop, hold and watched rate, and move when your own data moves.

Will AI replace our creators and our agency?

It replaces parts of the production bill and almost none of the judgement. The decisions that determine whether a portfolio works, which hypotheses to back, which situations and faces to cast, what to pause, remix and replace, what the brand standard says, are made by people, and the people in the frame are still real for the reasons above. What changes is the ratio. Fewer hours go into making the fourth version of a proven execution and more into deciding what deserves a version at all.

What should our AI policy for creative say?

Three lines. Where it is used, wherever it multiplies a decision already made, analysis, pretesting, versioning and adaptation of proven work, briefing and research. Where it is not, fully synthetic people in the brand's paid advertising, until the evidence says otherwise. And how the evidence is gathered, synthetic against real in bounded cells on the same media plan, with the four conditions, labelling rules, platform delivery treatment, category response and watched rates, reviewed on a set date rather than when someone raises it.

Sources

• Kantar LIFT ROI studies, US, Canada, Germany, France and Thailand, 2021 to 2023, 56,984 creatives and 1,295 campaigns, as presented in Meta, "Demystifying attention", 2025. Single human with eye contact and distinctive atmosphere, each against creatives without the feature. Commissioned by Meta, which sells the media, modelled, Kantar sells the studies.
• YPulse, "The Cult of Ugly", 2019, as cited in Meta, "Powering Pet Culture", 2024. 84% of young consumers agree they like it when content from brands is not perfect. Survey, carried in Meta sales material.
• Meta Business Help Centre, "About ad auctions". Each ad ranked on bid, estimated action rate and ad quality. Platform documentation, Meta sells the media.