Nestlé's 35,000 Assets Need More Than a Brand Book

Nestlé's plan for more than 35,000 content assets a year shows where marketing is heading: production capacity is rising faster than brands' ability to define what should be produced. Brand guidelines, customer evidence, approved claims and human review now have to function as working inputs, not reference material.

A 35,000-asset content operation sounds like a production story. It is really a knowledge story. This week's launches kept pointing to the same constraint: once software can create, place and optimise work at scale, every fuzzy instruction a marketing team used to resolve informally becomes expensive.

Scale exposes the rules

Nestlé has appointed WPP to run an AI-powered content operation across Greater China that is designed to produce and deploy more than 35,000 pieces of content a year.¹ That number is impressive, but the interesting question is what has to be true before volume like that becomes useful. A brand can generate thousands of variations and still end up with thousands of slightly different ways to sound unlike itself.

Jet2's new Adobe partnership makes the same point from another direction. The airline plans to use agentic AI across a base of 10 million myJet2 customers, while Adobe Firefly Foundry and Brand Intelligence are intended to keep generated work tied to approved assets and brand standards.² At that scale, the production system has to include rules that used to sit outside production.

That matters because traditional brand governance was designed for a slower workflow. A team wrote a brief, an agency or internal creative team interpreted it, somebody reviewed the work, and mistakes were caught through a series of human handoffs. At 35,000 assets a year, relying on people to remember every approved phrase, product distinction, visual convention and forbidden claim is not a serious operating model.

The brand book therefore has to become more operational. Approved examples, current product facts and evidence for claims need to be easy to retrieve. Exceptions need to be explicit, and the people approving work need to agree on where variation is welcome and where consistency is non-negotiable. A static PDF can describe a brand, but it cannot by itself tell an agent what to do when a new audience, channel or performance signal creates a choice. That decision layer has to exist somewhere.

This is one reason the conversation around AI content marketing is moving away from pure generation. The useful systems are beginning to connect generation to brand knowledge, customer signals and feedback. The output is only the visible end of a much bigger decision process.

Instructions can spend money

The stakes rise again when the software is allowed to act, not merely draft. Omneky's Slack integration can take a campaign brief, research the brand and competitors, generate creative variants, gather approval, launch the campaign and shift budget towards better-performing work.³ A sentence written in a Slack thread can now sit much closer to live media spend than most teams are used to.

A vague brief was annoying when it produced a weak first draft. It becomes much more consequential when the same vagueness can influence audience selection, media placement and budget allocation. "Increase conversions" sounds like an objective, but it says nothing about the customer the brand wants, the messages it will not use, the trade-off it will accept between efficiency and reach, or the point at which a person must review the decision.

That is why human approval keeps appearing in products that are otherwise becoming more autonomous. Pipedrive's Nova can prepare calls, record meetings, draft follow-ups and suggest CRM changes, but those changes are presented for review before they are committed. Pipedrive's own research found that only 23% of respondents were comfortable with AI completing tasks without their review.⁴Approval is becoming part of system design rather than a final courtesy.

The same distinction matters in Instagram marketing AI, especially for small businesses where the owner may also be the final approver. Instagram AI content management can automate resizing, scheduling, first drafts or repetitive variations and save useful time. Deciding whether a post makes a claim the owner can stand behind, whether a customer image should be used, or whether a joke sounds like the business is still a judgement call. The best Instagram content strategy will not be the one with the highest automation percentage. It will be the one that makes routine work cheap while keeping meaningful decisions visible.

That is also a useful way to evaluate any Instagram AI content tool. Ask what the system needs to know before it creates anything, what it is allowed to change, and where the owner gets the final say. Those questions tell you far more about whether the tool will stay on brand than the number of posts it can generate in a minute.

Context before generation

Raspberry AI's expansion across fashion design, merchandising, wholesale, marketing and ecommerce starts from a useful premise: generation is stronger when it inherits context from the product and commercial decisions that came before it.⁵ Instead of treating marketing as a separate prompt at the end, the platform is designed to carry information forward from concept to commerce. That makes the creative output a continuation of the product story rather than an attempt to reconstruct it from scratch.

Typeface pushed this idea further with the next generation of its Orchestration Engine. Arc Graph is designed to turn brand guidelines, previous campaigns, assets and other knowledge into structured rules. Arc Fabric provides reusable agents and skills, while Arc Loop writes campaign learnings back into the system so they can influence what gets made next.⁶ That is a more useful model for AI in marketing than treating every campaign as a fresh conversation with a model that knows nothing about what happened before.

Marketing teams already possess much of the material these systems need. It lives in research decks, campaign reports, customer calls, product documentation, brand books, asset libraries, CRM notes and the judgement of experienced staff. The problem is that much of it is difficult for software to retrieve at the moment a decision is being made.

This is where "how to stay on brand with AI content" stops being a copywriting question. A model cannot consistently honour a rule it cannot access. It cannot learn from a campaign result that remains buried in a dashboard. It cannot distinguish an approved product claim from an old line somebody happened to leave in a slide deck. Brand consistency increasingly depends on the quality of the context surrounding generation.

That changes the work for marketing teams. Before asking how to automate Instagram content creation, it is worth asking whether the underlying materials are usable: current product facts, a clear audience definition, examples of strong posts, phrases the business actually uses, images it owns, and examples of what should never be published. A stronger model will not compensate for contradictory instructions.

Trust does not scale automatically

More output can create a second problem: audiences may not value synthetic abundance simply because it is cheap to produce. VML's Future Shopper research found that 58% of consumers surveyed cross-check AI-generated product recommendations before buying, 49% say AI-generated product imagery reduces trust in a brand, and 48% skip content they think was created using AI.⁷ Those figures do not say consumers reject every use of AI, but they do show that production efficiency and audience trust are different measures.

The Advertising Research Foundation found confidence in AI-generated outputs among marketers using them had risen from 82% in November 2025 to 93% in May 2026. Yet the underlying ARF study found only 52% had formal internal AI policies.⁸ Confidence can increase faster than the mechanisms for checking what a system did, why it did it and who approved the result.

That gap becomes more visible when synthetic material enters public advertising. California's SB 1050 requires clear disclosure when audio or video advertising uses synthetic performers, according to coverage included in this week's material.⁹ Whatever view a marketer takes of that rule, it turns provenance into an operational requirement. Teams need to know what was generated, which assets were used, whether consent applies, and what must be disclosed.

The practical implication is not that brands should avoid AI captions for Instagram business posts or generated variations. It is that cheap production raises the value of provenance and judgement. A real product photo, a customer quote with permission, a specific founder opinion or evidence from actual buyers becomes more valuable when plausible alternatives can be generated endlessly.

This is also why "human in the loop" is too vague to be useful on its own. A person clicking approve on 500 nearly identical assets is not meaningful oversight. Human review has value when the reviewer has enough context, enough authority and a clear reason for being in the process. The review point should sit where judgement changes the outcome.

Discovery without the click

The same need for structured brand knowledge is appearing on the customer side. OpenAI's Sponsored Agents test lets a person move from an ad into a clearly labelled conversation with a business-sponsored agent.¹⁰ Pinterest's Visual Search Ads can place a brand while a user is visually comparing what they might buy, and Pinterest says it handles more than 80 billion searches a month, with more than half carrying commercial intent.¹¹

At the same time, products such as PeakMetrics AI Perceptions are trying to measure how brands are represented across systems including ChatGPT, Gemini, Claude, Grok and Perplexity.¹² The old digital assumption was that discovery eventually produced a visit to a page a marketer controlled. AI assistants can now interpret the brand before that visit happens, and in some cases they may answer enough of the customer's questions that the visit becomes less important.

That makes consistency across sources more consequential. Product descriptions, reviews, third-party coverage, customer evidence and brand-owned pages can all contribute to how a system describes a company. A beautifully written campaign cannot fully repair contradictory product information or a weak body of external evidence.

For marketers, this widens the meaning of content strategy. The material has to persuade a person and be legible enough for software to understand accurately. That does not mean filling the web with pages written for bots. It means stating product facts clearly, keeping them current, earning credible references and making the brand's actual point of view recognisable wherever it appears.

The brands that benefit from conversational discovery will probably be the ones that can answer a simple test. If a customer asks an AI assistant what you sell, who it is for, why it is different and whether your claims are credible, the underlying evidence should lead to an answer you recognise. If it does not, producing another hundred campaign variants will not fix the missing information.

The brand book needs a working layer

Nestlé's 35,000 assets are useful because the number makes the constraint impossible to ignore. Once production scales that far, "stay on brand" cannot remain an instruction people interpret differently. The brand needs a working layer of approved facts, examples, assets, exclusions, evidence and decision rights that both people and software can use.

That does not reduce the role of the marketer. It makes the marketer's judgement easier to see. The repetitive work can move into systems, while people spend more time deciding what the brand believes, which customers matter, what evidence is good enough and where automation should stop.

The teams that do this well will not be the ones that automate every step first. They will be the ones whose knowledge is clear enough to survive automation without becoming generic. More capacity only helps when the brand gives that capacity something worth repeating. The 35,000th asset still has to sound like it came from the same company as the first.

Sources