6 AI shifts that marketers need to act on

01AI content at scale

Everyone has seen an AI video. Scale is something else. It is a repeatable studio: a plain-language brief, a second-by-second shot list, generation, fixes in post, and a named person who signs off every ad. The models change every month, so the moat is not the model or the production budget. It is the brand system you build around them: real pack shots, brand rules, and a QA checklist of everything that has ever gone wrong.

For a household brand we built two 30-second ads in hours, a job that would normally mean two shoots. What made them hold up was not the model. It was using real pack frames as references, a tight shot list, and fixing small errors in post instead of re-rendering. Every failure became a new line in the checklist. BCG describes the same pattern across CPG: content volume is exploding while unit costs collapse, and the value sits in the guardrails and know-how a brand keeps for itself.

Monday: pick one always-on brand and write its brand system down for AI. That document is the asset, not the prompt. (How we run it: set up your own AI ad studio.)

02Compose for every format

For years, CPGs shoved cropped TV ads into YouTube and TikTok and hoped they would work. One master key visual, shot once because shoots were expensive, then cut down to square and vertical. The vertical crop usually lost half the people in the frame.

In the talk I showed the same picnic key visual twice. Cropped the old way, the 9:16 version lost two of the three friends. Composed natively for each format, everyone stayed in, and every version looked like it was made for that screen, because it was. The idea and the judgement still come from people. AI just removes the compromise. (Key visuals: from the photography era to the AI era.)

Monday: list what in your process exists only because shoots were expensive. The master KV, hero-plus-cutdowns and the annual shoot calendar are good places to start.

03Building persistent assets

A campaign's characters and sets used to live for one shoot and then disappear. Now every character, location and running joke that earns a response is an asset you own and can keep refining, so its relevance compounds instead of resetting with each campaign.

The Great Snack Run is a small example. The kids, the kitchen and the little brother's running gag are now reusable, so a second episode costs a fraction of the first and each one makes the characters stronger. We have done the same with a Fanta Halloween character universe and with Sam, a virtual influencer.

Monday: list the characters, places and running jokes your brand already has, and build reference sheets for them that AI can work from.

04Shelf moves to agents

The shelf has moved from the store, to the ecommerce search page, and now to the AI agent answering the shopper's question. The aggregators are being aggregated. In Singapore, Meta's Muse now sits on top of Shopee, Lazada, Grab and others. Agents don't scroll past your end-cap or your banner. They give two or three names. If ChatGPT doesn't recommend your brand, you're not on the shelf.

GEO, generative engine optimisation, is getting your brand into those answers, and it is the cheapest win in marketing right now. I got my own SaaS, IDstudio, to rank number 1 on ChatGPT for AI interior design in two months. It took clear, specific, consistent content about what the product does and who it is for, not a media budget. The engines don't even agree with each other: one BrightEdge analysis found ChatGPT and Google's AI answers recommended different brands for most queries, which means the shelf is still up for grabs. For challenger brands, this is the first time in years that budget size isn't what decides who gets seen.

Monday: ask ChatGPT, Gemini and Perplexity the ten questions your shoppers actually ask in your category, like the best biscuit for a kid's lunchbox. Note who gets named. That is your new share of shelf.

05Synthetic (AI) testing in every workflow

When did your team last approve a pack because it was proven, and not because it looked nice? Most pack and shopper calls run on taste and habit, because proper testing was too slow and too expensive for anything but the big launches.

For years we've obsessed over first-party data. Now it's time to do something with it: build synthetic consumers and shoppers from it. Start small with the basic differences, then scale to 1,000, then 10,000. In a Shelf Test demo on a fictional cookie brand, 1,000 synthetic shoppers showed attention on the pack claim rising from 1.3% to 4.5% after a redesign. They noticed the product 37% of the time at the far end of the aisle against 29% at the entrance. A shelf talker added 14.6 points while a digital shelf strip added 0.4. It is precise to about plus or minus 3 points, so treat it as a screen rather than a verdict: kill the weak options fast, then spend real research money on the two that survive. (How the synthetic shopper test works and how accurate predictive eye tracking is.) Built into the workflow, it beats the marketing director's judgement call every time.

Monday: take the last pack or POSM decision your team approved on instinct and test it against one alternative. You can try a single image in the free Attention tool today.

06AI analytics for deeper insights

AI finds the patterns we never had the time, or the lack of bias, to look for. It can sort through billions of data points and surface correlations nobody would have known where to look for.

Picture a biscuit brand that connects its loyalty data, retailer sales and campaign results, then asks AI one open question: where are we leaving money on the table? It might find that households who buy the brand for lunchboxes in term time drop off in the school holidays, and that the drop is sharpest in families who also buy fresh fruit. Nobody would have built that cut on a dashboard, but it points to a clear holiday occasion and a pack to test for it. I've seen real versions of this, and the hard part was never the AI. It was getting the data tagged and connected, which is why I keep writing about the semantic layer.

Monday: pick one question your team stopped asking because the data was too messy, and give it to AI with the raw data.

07The real shift: Taste Velocity

The thread through all six is the same. When making, testing and analysing become fast and cheap, the bottleneck is no longer budget, production or time. It is creativity, scale, judgement and the willingness to adopt: knowing what good looks like, and deciding fast enough to act on it. I call that Taste Velocity, and I think it is the number marketing teams should be working on now.

The Great Snack Run took an evening to make, but getting it good took around 900 generations over months of learning: which take to keep, when the physics looked fake, when the joke didn't land. The AI did not supply the taste. That part is still the job.

So the question is no longer "can we make it?" but "do we know what good looks like, and can we decide fast enough?" That is exciting, and a little unsettling. Thank you to the Mondelēz team for having me, and for the questions and the pushback.

This is the work I do now with CPG and consumer brands as an AI and digital consultant for CPG and FMCG brands: setting up the studio, the testing and the data so the team can run it themselves. If you want to work out where your team should start, talk to me.