GenAI Content Production at Scale: Mondelez’s $40M Bet

Two numbers from a snacks company’s investor desk have reshaped a conversation that most brands are still having in the abstract. Mondelez has committed more than $40 million to a proprietary generative AI platform for marketing content, and it is targeting 30 to 50 percent lower content production costs. The platform is already running Chips Ahoy posts in the United States, Milka clips in Germany, and Oreo product pages on Amazon and Walmart from November. A holiday TV ad is slated for 2026, and a Super Bowl slot for 2027.

I have watched one of these waves land every decade since I bought my first Mac Plus. QuarkXPress put the print shop in a drawer. Flash put the broadcast studio in a browser. The ones that stayed were the ones that forced the organisation to change shape underneath them, not the ones that shipped a tool. Mondelez has done the forcing part in public, and that is why this case is worth reading closely rather than another earnings-call sentence about AI.

The question for a reader running a brand with real content obligations is not whether to copy Mondelez. It is which of their choices actually survives at mid-market scale, which ones are an artefact of a $37 billion revenue line, and what a WordPress and WooCommerce estate actually has to carry if the ambition is the same.

Mondelez’s content job was impossible before the platform

Mondelez International reported approximately $36 billion in 2023 net revenues (Mondelez press release, 26 September 2024). It carries roughly thirty-seven brands including Oreo, Cadbury, Milka, Toblerone, Chips Ahoy, Ritz, Lacta and Sour Patch Kids, and sells into more than 150 markets.

The content job that comes with that structure is not glamorous. Every brand needs localised campaign assets in every market in every language, refreshed every quarter, consistent with the brand’s codes and compliant with each market’s advertising rules. Before the platform, that work moved through a chain of local agencies and in-house teams, each one briefing, drafting, reviewing, approving, trafficking. The cost does not sit in any single invoice; it sits in the elapsed time between a brief and an aired asset, and in the number of adaptations that never got made because the budget ran out before Germany’s Milka got the version the UK’s Cadbury did.

That is the before-state. It is boring and expensive and it is the state most multi-brand businesses are in.

They chose a proprietary platform, and the choice is the story

On 26 September 2024 Mondelez announced a joint build with Accenture and Publicis Groupe. The split of roles is the detail worth noticing. Accenture built the digital core: real-time data integration, measurement, the responsible AI framework, employee training and adoption. Publicis built the creative foundation: the generative models themselves, the asset production workflow, the quality and brand-safety layer sitting on top of the output (Accenture newsroom, 26 September 2024).

Jon Halvorson, Mondelez’s Senior Vice President, Global Consumer Experiences and Digital Commerce, framed it this way in the press release: “Harnessing the power of gen AI will empower our people to play a proactive role in how our brands show up in the market.” Venky Rao, Accenture’s Americas and AI Lead for Mondelez, added that the point was to let marketers “tap into the power of data, AI and gen AI” against brands that are “some of the world’s most iconic.” Scott Hagedorn, Publicis Groupe’s Global Chief Solutions Architect, called it “part of our broader Power of One solution.” The quotes are polite press-release prose; the deal underneath them is not.

The decision not to buy an off-the-shelf tool is where I would spend time if I were sitting in a procurement review of this case. An off-the-shelf GenAI content tool is a feature of someone else’s roadmap. A proprietary platform built on real-time first-party data is a capability that can be governed, audited and tuned to a brand book. For a business with thirty-seven brand codes to protect, there is no version of the first option that holds up at scale. The $40 million is the price of owning the second.

The measured result lives in two places, and only one of them is public

Public, from Halvorson to Jessica DiNapoli at Reuters on 25 October 2025: the platform is targeting a 30 to 50 percent reduction in marketing content production costs. Chips Ahoy’s United States social feed and Milka’s German social feed are already running on it. Oreo’s product pages on Amazon and Walmart start using it from November. Lacta and Oreo roll out in Brazil over the coming months. Cadbury rolls out in the United Kingdom. A short television commercial will be ready to air for the 2026 holiday season, and the ambition stretches to the 2027 Super Bowl.

Not public: the brand-level creative effectiveness. Did the Chips Ahoy posts outperform the ones a human agency would have produced? Did the Milka clips carry the brand’s distinctive assets correctly on an eight-second schedule? Mondelez has not published a comparison yet, and I would read a cost figure without an effectiveness figure as half of a dashboard. The reason to care about that gap is specific. If GenAI reduces unit cost by 30 to 50 percent and reduces creative effectiveness by any amount at all, the economics are worse than before and the saving is a drawdown on the brand.

I say this because the industry commentary around AI-generated advertising has been consistent, published and uncomfortable. The “Blandemic” essay by Red Brick Road frames the risk precisely: AI trained on prevalent material produces averages of averages, and distinctiveness, which is the Byron Sharp school’s whole point about why brands grow, is the first thing to erode. Mondelez’s own line on human likenesses, where it has chosen not to use computerised faces in ads even when competitors do, reads to me like someone who has thought about which lines are worth not crossing. That restraint is the brand-safety dividend of owning the model rather than renting it.

What this means inside a real WordPress and WooCommerce estate

A reader at a mid-market multi-brand or multi-market business is not going to spend $40 million. The question is what the Mondelez pattern looks like at one hundredth of the budget, on the stack most of the world’s commerce actually runs on.

The platform has four layers under it, and all four are buildable on WordPress multisite and WooCommerce without inventing anything new. Costs below are realistic for a 2,000 to 10,000 SKU store with three to five markets. They are what we would quote. They are not a promise.

The product and content data layer. Before anything else, the catalogue has to be modelled so that a brief for one brand in one market can inherit from a global template and override only where the market requires. In WooCommerce that is attribute design, variation structure and global attributes, plus a disciplined taxonomy for brand, product line and campaign. In WordPress multisite that is a shared content type and shared taxonomy, with per-site overrides for the fields that must be local. The remediation of an existing catalogue to this shape is the single biggest cost on any project of this kind. Realistic effort: three to six weeks for a 5,000 SKU estate, two to three sprints for the content model if the business has more than one brand.

The content generation layer. This is the smaller lift than most people expect. OpenAI, Anthropic or an open-source model with a brand-tuned prompt library handles the drafting. Where the work actually is: a retrieval layer that feeds the model the brand book, product specifications and the market’s regulatory constraints before it generates a word. Without retrieval you are back to generic output that drifts to the mean. The implementation shape I like on WordPress is a custom block that calls the model server-side with the retrieval context, stores the generated draft as a revision, and keeps the published state in human hands. One to two sprints to ship a working first version, plus the ongoing cost of the model, which belongs on the invoice as its own line.

The brand governance layer. This is the one that decides whether the output is on-brand or merely acceptable. It is not an AI problem; it is an editorial problem. Who approves the draft, who checks it against the brand code, who signs it off for a market, where the audit trail lives. WordPress has had this layer for twenty years. It is called editorial workflow and it survives contact with every new content format thrown at it. Expect two to three weeks to configure, then ongoing time from senior brand owners that does not go away. The platform frees them from drafting, not from deciding.

The measurement layer. This is where Mondelez has publicly been thinnest, and where a smaller business should be most careful. Instrument the pipeline so that each generated asset carries a utm, a content provenance tag and a brand-code identifier. Then feed engagement back into the generation layer, so the model learns which patterns worked on which market. One week to instrument, four to six weeks of real data before the loop produces anything useful. Without this, the generator is a drafting shortcut, not a system.

Total realistic effort to go from a conventional WooCommerce multi-brand estate to something that can run a GenAI content pipeline with the governance to stand behind the output: eight to twelve weeks of focused work, roughly two-thirds of it on the catalogue, taxonomy and workflow and only a third on the model. The ordering is not negotiable. A content model with no catalogue discipline produces faster drift to the mean, which is the opposite of why you started.

What is still unsolved, and the honest line to hold

The holiday 2026 television ad is the real test, not the Chips Ahoy social post. Short-form social carries forgiveness that broadcast does not. Thirty seconds in front of a national audience, with a specific brand’s distinctive assets intact and a creative idea the audience has not seen before, is the honest bar. Mondelez has signalled they will meet it; nobody has published evidence yet that any GenAI-first ad has met it at that scale.

Brand distinctiveness is the second unsolved problem. If the sector collectively adopts platforms that converge on the same training distribution, every brand’s output will drift towards every other brand’s output. That is a market-level externality that the Mondelez platform cannot solve on its own, no matter how carefully its retrieval layer is tuned. The agency industry’s own commentary on this is a credible read; it is the Byron Sharp school warning about the first casualty of volume being distinctiveness, and it is right.

The third unsolved problem is the one that matters most to a procurement reader: who is responsible when the model is wrong. Mondelez has an answer in Accenture’s responsible AI framework, which is at least a named thing. A smaller business running a smaller stack needs an equivalent before anything ships, and the equivalent is not a policy document. It is a workflow that catches the output before it reaches the public, and a named person at the top of it.

If I were sitting across from an ecommerce director reading this case over coffee, my recommendation would be short. Treat Mondelez’s $40 million as a signal, not a template. The signal is that the structural remediation, catalogue, taxonomy, workflow, measurement, is the real spend, and the AI is the smaller part sitting on top of it. Spend the first eight weeks on the structure. Spend the next four on the model and the retrieval. Keep senior brand owners in the loop at approval, not drafting. Hold the line on broadcast until the thirty-second test is honestly passed on your own content, not someone else’s deck. And if the alternative being pitched to you is an off-the-shelf tool that promises to do all of this without touching your catalogue, that is the pitch to walk away from. We build the version that keeps the brand code intact, because the brand code is the asset the budget is defending.