Part III, The Craft · Chapter 17

The Dark Art and the Machine

AI in marketing promotions: what it can do for ideas, and where it stops

A film star for every corner shop

At Diwali, India gives gifts, and Cadbury Celebrations, a box of assorted chocolates, is built for it. In 2021 the brand had a problem that was not really its own. The small neighbourhood shops that carry its boxes had been through two years of the pandemic. Shoppers had got used to delivery, and the big online platforms were taking the festive attention. Cadbury needed the corner shops to have a good Diwali, because Cadbury’s Diwali ran through them.

The answer, from Ogilvy and Wavemaker for Mondelez, was an ad starring the Bollywood star Shah Rukh Khan, telling people to buy their Diwali gifts from local shops. Then they made the ad name the shops. Machine learning regenerated his face and voice so that he could say the names of individual stores, mapped by postcode across four categories: fashion, footwear, general stores and electronics. Each version was served on YouTube and Facebook to people near the shops it named, so a viewer saw the star recommend a shop down their own street. Shop owners could make their own version on a website and share it. About 130,000 versions were made, featuring around 2,000 stores, and they were viewed 94 million times. In 2023 the work won the Grand Prix for Creative Effectiveness at Cannes, and the award entry credits it with Cadbury’s best Diwali sales. We have not seen the underlying data, and chapter 15 is clear about what an award entry can and cannot prove.

Look at where the machine sits in that story. The idea, help the corner shop and the corner shop will help you, is the four-part rule from chapter 14 done at national scale: please the retailer, support the brand, encourage the shopper, reward the consumer. Nobody needed a computer to have it. What the computer did was make 130,000 versions of it, each one specific enough to feel personal, which nobody could have afforded to shoot one at a time. The person had the idea. The machine did the part that only a machine could do.

That is most of what this chapter has to say, and it matters, because clients now arrive with a different question. They want great ideas, they want the dark art, and they think AI can help. They are right that it can. The useful work is knowing which part.

Can AI come up with promotional ideas?

Yes, quickly and in volume, and on average they are not bad. In 2023 four researchers at Wharton, Karan Girotra, Lennart Meincke, Christian Terwiesch and Karl Ulrich, compared product ideas from MBA students with ideas from GPT-4. The brief was a product for college students that would sell for $50 or less. The students had produced 200 ideas. The model produced 200 in about fifteen minutes. In a blind survey, consumers were more likely to say they would buy the machine’s ideas: an average purchase probability of 47 per cent, against 40 per cent for the students’. When the model was shown examples of good ideas first, its average rose to 49 per cent.

A second study adds the catch. In 2024 Anil Doshi and Oliver Hauser asked 300 people to write very short stories, some with no help, some with one idea from an AI and some with a choice of up to five. With the most help, the stories were judged 8.1 per cent more novel and 9 per cent more useful. The writers who gained most were the ones who scored lowest on a test of creative thinking; the most creative writers gained little. And the stories written with an AI idea were 10.7 per cent more similar to each other than the stories written without.

Put those two findings next to the promotion aisle. AI raises the floor: a team that would have produced a weak list produces a decent one, fast. It also narrows the field: everyone who asks the same machine a similar question gets neighbouring answers. For a lot of marketing that may not matter much. For promotions it matters a great deal, because a promotion lives or dies at the shelf next to other promotions, and chapter 13 showed that the fame job depends on being the one idea people retell. If your competitor asked the same model for “a Diwali promotion for a chocolate gift box”, you have both got the same list. The value is in what you do next.

Where AI helps a promotion, and where it does not

The clearest public evidence on that question comes from a 2023 experiment by researchers from Harvard Business School and other universities with 758 consultants at Boston Consulting Group. On tasks the researchers had chosen to sit inside what AI could do well, consultants with GPT-4 completed 12.2 per cent more tasks, 25.1 per cent faster, and produced work judged more than 40 per cent better than a control group. On a task chosen to sit just outside it, a business problem built so that the model’s answer would be wrong unless the person checked the evidence, consultants using AI were 19 percentage points less likely to reach the right answer than those working without it.

The researchers called the boundary a jagged frontier. The problem is not that AI is good at some things and bad at others; every tool is. It is that the line is invisible, and the work outside it looks just as confident as the work inside. The consultants who did worse were not lazy. They trusted a good-looking answer in the one place it was wrong.

In our experience the frontier runs through a promotion brief roughly like this. Inside it, the machine is a fast, tireless junior:

Range. Fifty mechanics for a brief, sorted by job, in a minute. Twenty ways to say the offer on a wobbler in under ten words. The three-second test from chapter 1, run on every one.

Drafts. A first draft of the terms against a checklist, a first pass of the one-page brief, a summary of every promotion in the category last year from published pages.

Arithmetic. Redemption at three scenarios, the cost if it runs hot, the break-even on a cashback, the sum that chapter 7 says should be done before anyone falls in love with a number.

Argument. The pre-mortem from chapter 18, done tirelessly: every way this could fail in each of the three directions.

Outside it, and this is where the dark art lives:

The buyer. What this category manager wants this month, which in chapter 14 was a Harley-Davidson. That knowledge is in a relationship, not on the internet.

The give. Where the room is in a rule, an insurer’s appetite or a retailer’s calendar. Chapter 7 called finding it the third answer to an impossible brief. The machine has read the rules; it has not watched them bend, legitimately, a hundred times.

The question nobody asked. Whether to run the promotion at all. Chapter 15 told the story of a client asked what would happen if they ran their monthly promotion every other month. That question was not on the brief, so a machine answering the brief does not ask it unless someone tells it to.

Taste. Which of the fifty ideas is the one people will retell, and which is merely the one that scores well. The Wharton study measured purchase intent in a survey. It did not measure whether anyone would talk about the product the next day.

What our own AI taught us

We have run the experiment on ourselves. In 2025 we built an AI, which we called Trudy, to read promotion briefs the way we do, and we trained its judgment on the first edition of this book. Brands and colleagues sent it briefs and it wrote back. Before writing this chapter we went back through its archive: about forty readings, across appliances, wine, beer, snacks, coffee, a telco and a card company. We had remembered it as a machine for saying no. That is not what the archive shows, and what it does show is more useful.

It hardly ever said no. In about forty readings it never once told anyone not to run a promotion. What it did was take the first edition literally. A campaign outside the “optimal four to eight weeks” was marked down, whatever the product. A three-month window “kills urgency”. And whatever the brief, the cure was more. Of the 29 written evaluations it sent in the middle of 2025, 25 told the brand it needed a trade incentive and 15 suggested adding a game or a digital layer, including to one entry that was already automatic and could not have been easier. Almost every promotion sat “closer to forgettable than iconic”, and the way out was always another layer: more story, more prizes, more plays.

That is the Doshi and Hauser finding at the scale of one tool. The machine gave every brand the same notes, because the same notes were in the book it had read. It was still the best reader we had: tested against two general-purpose AI tools on the same brief, it was brilliant by comparison, about 85 per cent of the way there.

The bigger gap was the question. In none of those evaluations did it ask whether the promotion should run at all, or run less often, or what would happen with nothing. The clearest case came in 2026, when it reviewed a client’s year of planned promotions and recommended adding one at the end of the financial year. The client’s answer was that they stay out of that period on purpose: so many suppliers and retailers are running activity then that nobody wants to invest into the noise. Anyone who had worked with that client would have known. The machine could not know, and nothing in the brief invited it to ask. As one of us wrote at the time: “It is those nuances that AI just cannot handle. You and I both know it instinctively but the challenge is teaching a machine about those elements.”

There was one exception, and it is the most useful thing in the archive. In 2026 we built Trudy a workshop mode, with several agents told to argue with each other, one of them cast as a heretic whose job was to disagree with the brief. On an appliance brief, the heretic argued that the answer was not a promotion at all, and an always-on offer that refused to move with the sales calendar made the final shortlist. The machine asked the unasked question once, when we built it a seat whose job was to ask it.

So the lesson is not that AI cannot ask. It is that it answers the brief it is given, in the posture of whatever it has read, and it will only question the brief if someone designs that in. That is why this edition reads the way it does, from NO to HOW, and why Trudy changed with it: it now asks which job a promotion is for before it reads anything, returns versions that work rather than a verdict, and is not allowed to say no. It still does not have the brilliant ideas. That part was always ours.

How to use AI on a promotion brief

None of this is an argument against using it. We use it every day. It is an argument for using it on the right side of the frontier, and for keeping a person on the other side. The practice we have settled on is short.

Seven Rules For Ai On A Brief

Give it the brief before you ask for ideas: the job, the person in the aisle, the venue, the budget and the number to beat. Ask for many, then throw most away; the first ten answers are the ones your competitors got too. Ask it for the version that works, not for a score; a machine asked to judge will judge. Give it a contrarian’s job: ask it to argue that you should not run the promotion at all, then run the pre-mortem with it. Check every fact, threshold and rule it gives you against the source; it drafts, it does not advise. Keep personal information out of public AI tools; Australia’s privacy regulator says so in its first recommendation on the subject. And take whatever it gives you into a room with someone who has run a promotion, before anyone else sees it.

The last rule is the one that matters most, and the Cadbury story shows why. The machine made 130,000 ads. A person decided the corner shop was the story.

AI on the other side of the counter

The machine is not only on our side of the promotion. Two changes are already reaching the engine and the shopper, and both belong in the brief.

The first is fraud. Promotions that ask for a photo of a receipt have always attracted a few fakes. What has changed is the tool. AppZen, a company that checks expense claims for businesses, reported in June 2026 that of the receipts its system caught as fake, the share generated by AI went from none in March 2025 to 70.8 per cent by the middle of May 2026. Those are workplace expense claims, not promotion claims, and the count is small: 1,471 generated receipts from 745 employees at 174 companies. But a promotion that pays cash against a picture of a receipt is checking the same kind of document, and the picture is now easy to make. That moves the engine row on the Kill Sheet. Ask how the platform validates a purchase, and prefer proofs a generator cannot invent: a unique code under the cap, a retailer’s loyalty record, a match against the retailer’s own transaction data.

The second is the shopper’s own machine. In September 2025 OpenAI began letting people in the United States buy some products without leaving a ChatGPT conversation, using a protocol it built with Stripe and released for any merchant to adopt. When an agent does the shopping, the three-second sum from chapter 1 is run by something with no dopamine. The Gambler has no seat. The Accountant gets a calculator, and reads every word of the terms. We do not know yet how far that goes in the aisle. We do know which promotions it favours: a reward stated plainly, conditions anyone can read, and a claim that does not depend on the shopper giving up halfway. That is the same promotion this book has been arguing for since chapter 4, read by a reader who never gets tired.

Where this leaves the brief

Clients are right to want AI in the room. It makes more ideas than any team, faster, and it never gets bored of the arithmetic. What it does not do is know the buyer, find the give in a rule, ask whether the promotion is needed, or tell the idea people will retell from the one that merely scores well. Those are the dark art, and they are still done by people who have seen enough promotions to know where to look.

So the working answer is a pairing. Let the machine widen the list and do the sums. Let a person choose, and ask the question that was not on the brief. The next and last chapter puts the tools for both in one place.