Correction, 2026-09-30: a source check changed four entries. Lidl's 1.7 million images were made by the public in a branded app, not produced as retail visuals; Novo Nordisk's 14% click lift could not be traced to a primary source and is removed; Monks' 97% is a cost cut versus legacy approaches and its CTR gains came with Performance Max targeting; Omneky's figures are vendor-reported. Klarna, Heinz and Kalshi were corrected earlier the same day.
Most AI advertising "case studies" are vendor decks with a hero metric and no context. The seven below cleared a simple bar: a named brand, a described intervention, and a number that was reported publicly rather than whispered in a sales call. They also split cleanly into two tooling classes worth keeping straight, generation (making the creative) and optimization (targeting and assembling it), because the numbers mean different things depending on which one produced them. Where a case is impressive but the metric is soft, we say so. The goal is not a highlight reel. It is a set of interventions you could actually copy, with the honest caveat attached.
TL;DR
- Generation wins show up as cost and speed: Kalshi's $2,000 primetime spot, Klarna's six-weeks-to-seven-days image cycle, Monks' reported 97% cost cut.
- Idea-led generation wins show up as reach: Heinz turned a model's bias into over 1.15 billion earned impressions (2022).
- Optimization wins show up as ROAS and CTR: Omneky's 6x ROAS, Novo Nordisk's 24% open-rate lift.
- Read the tooling class before you read the number. A ROAS lift from targeting is not evidence that generated creative works, and vice versa.
- What to copy is in each entry. The pattern: use generation for volume and cost, use optimization to find the winner inside that volume.
1. Klarna: the operational case (generation + workflow)
What they did: Used off-the-shelf generative tools (Midjourney, DALL-E, Firefly and others) to produce marketing visuals at scale, per Klarna's May 2024 release. Numbers: Over 1,000 AI-generated images in Q1 2024, image development cut from about six weeks to seven days, about $10 million in marketing savings on an annualised basis attributed to AI, including a $6 million reduction in image production costs, and a 25% cut in external supplier costs, all while running more campaigns. What to copy: Treat generation as a throughput system, not a hero-spot generator. The compounding win is the cycle time. When making the fiftieth variant is nearly free, you test more and find better. This is the creative volume thesis in numbers.
2. Heinz "AI Ketchup": the idea-as-AI case (generation)
What they did: In 2022, with the agency Rethink, asked DALL·E 2 to draw "ketchup", noticed it kept producing a Heinz-like bottle, and built the campaign around it ("Ketchup even means Heinz to AI"), inviting fans to suggest prompts. Numbers: Over 1.15 billion earned impressions, worth more than 25 times the media investment, and engagement 38% above past campaigns, per the award entries. What to copy: The strongest generation ideas are the ones that could not exist without the model. Heinz did not use AI to make a normal ad cheaper. It made an ad whose entire concept depends on a property of generative models. That is the highest-value use, and the hardest to reverse-engineer, because it starts from an insight about the tool.
3. Kalshi: the cost-floor case (generation)
What they did: Aired a 30-second NBA Finals spot in June 2025 made almost entirely with Google's Veo 3, via filmmaker PJ Accetturo. Numbers: About $2,000 in production cost and three days from idea to air (per Kalshi), 300 to 400 Veo generations to get 15 usable clips (per Accetturo, who called it "a 95% cost reduction vs traditional ads"), on track for 20 million impressions. What to copy: Match the register to the tooling. The ad is deliberately chaotic and absurd, a style where Veo's occasional weirdness reads as intentional. The cost number is only reproducible if you pick a concept that flatters generated footage instead of fighting it.
4. Monks + Hatch: the performance-creative case (generation)
What they did: Monks (then Media.Monks) used its Monks.Flow AI workflow with Google Gemini to produce and iterate ad creative for Hatch's Restore 2 sleep device. Numbers: per Monks and Think with Google, 50% fewer design and production hours and a 97% cut in costs versus legacy approaches; with Performance Max targeting, 31% better cost per purchase, 80% higher click-through rate and 46% more engaged site visitors. What to copy: The performance numbers were measured together with Performance Max's AI targeting, so credit is shared between the creative and the delivery; the hours and cost figures are the generation result. The move is to feed a high-volume variant pipeline into a proper test, then let the CTR and CPP data pick the survivor. It is the variant testing workflow run at production scale.
5. Novo Nordisk: the copy-generation case (generation)
What they did: In 2022, used Phrasee's AI language platform to generate, test and adapt US email messaging to healthcare providers. Numbers: a 24% higher open rate (Econsultancy). A 14% click lift circulates on aggregator sites; we could not find it at a primary source, so we leave it out. What to copy: Generation is not only visual. Language models tuned on brand voice and tested against real opens can beat a copywriter's best guess on the specific, unglamorous surface of a subject line. The caveat: this is optimization dressed as generation, the win comes from testing many machine-written options, not from one brilliant line. Budget for the volume, not the single output.
6. Omneky + New Sapience: the full-funnel case (generation + optimization)
What they did: Used Omneky's AI image generation and AI creative insights to run a crowdfunding campaign for New Sapience, mid-April to early October 2023. Numbers (vendor-reported, from Omneky's case study): over $460,000 raised through Omneky ads toward over $1.18 million total, 6x ROAS, about $160 cost per investment, and over 52,000 link clicks. What to copy: This is the combined play: generate the creative variants and let insight into what performs route spend to the winners. The 6x ROAS is a full-stack result, so credit is shared between the creative and the targeting. That shared credit is exactly why you should keep the two tooling classes separate in your own reporting.
7. Lidl France + Marcel/Bria: the participation case (generation)
What they did: The agency Marcel (Publicis Groupe) and Bria built Lidlize, a public app that turned any object a person photographed into a Lidl-branded product. Numbers: users created over 1.7 million images in three weeks, peaking at over 1,000 a minute (PYMNTS); the work won a D&AD Wood Pencil in Creator Content. What to copy: The generation was done by the public, not the brand. That is the participation play: give people a toy built on your brand's look and let them make the volume. There is no ROAS attached, so do not read it as proof that the creative converts; read it as reach and engagement.
The pattern across all seven
Two tooling classes, two kinds of number.
Generation cases report cost, speed, and reach. Klarna's cycle time, Kalshi's $2,000, Monks' 97% cut, Lidl's 1.7 million, Heinz's 1.15 billion impressions. These prove the same thing from different angles: making the creative got radically cheaper and faster, which lets you make and test far more of it.
Optimization cases report ROAS, CTR, and conversion. Omneky's 6x, Novo Nordisk's open rates. These prove that assembling and targeting the creative got smarter.
The trap is attributing an optimization number to your generated creative, or a generation cost saving to your targeting stack. Keep them in separate columns. The full win, the one Omneky and Monks each caught a piece of, is stacking them: generate the volume, then optimize inside it.
What this means for brand teams
The reproducible move in every generation case is the same. Lower the cost of a variant toward zero so you can afford to make dozens, pick a register that flatters the tooling, and run a real test to find the winner. That is not a vendor pitch, it is what Klarna, Kalshi, Monks, and Heinz each did in their own register. The cost collapse behind all of it is the same one reshaping agency pricing.
If you want to run the generation half of that stack across every major image and video model on one canvas, then hand the survivors to your optimization tools, start on 8frame and clone a variant workflow. For the campaigns that won on craft rather than numbers, see the best AI-made ads of 2026.
FAQ
Which AI advertising case study has the strongest numbers? Klarna's is the most complete first-party report: over 1,000 AI images in Q1 2024, image development cut from six weeks to seven days, and $6 million less spent on image production.
Do these case studies prove AI creative converts better? Not on their own. Where performance numbers exist, as with Monks and Omneky, they were measured together with AI targeting, so credit is shared between the creative and the delivery.
Which cases are vendor-reported? Omneky's and Monks' figures come from the companies' own case studies; Klarna's from its press release; Heinz's from award entries. Read each with that in mind.
Sources
- Klarna: press release, 2024-05-28
- Heinz: Adweek, 2022-08-01 and One Club ADC Awards entry
- Kalshi: The Verge, 2025-06-12 and Business Insider, 2025-06-13
- Monks + Hatch: Monks case study and Think with Google, 2024-09-25
- Novo Nordisk + Phrasee: Econsultancy case study
- Omneky + New Sapience: Omneky case study
- Lidl France: PYMNTS, 2025-06-03 and D&AD 2025
All checked 2026-09-30.