Food companies know the difference between experimentation and production. Test kitchens can stomach an odd flavour or a recipe that never reaches the shops, but their factories cannot. A similar distinction runs through Nestlé’s approach to artificial intelligence (AI). The Swiss food giant wants staff to prod the technology widely but it is much less keen to let software that on occasion makes things up make decisions without human supervision.
In conversation with Economist Enterprise, Enrique Mora, Nestlé’s principal enterprise architect for AI, likens the company to “a small country”. It has around 270,000 employees, more than 2,000 brands and operations spanning raw-material procurement, manufacturing, logistics, marketing and sales. Amid the AI boom, a business of that size quickly faces a buffet of AI bolt-ons and pilots—and like most buffets, it tempts diners to pile their plates too high.
Nestlé’s answer is what Mr Mora calls “freedom in a frame”: room to experiment inside clear technological, ethical and operational boundaries. A frame with three struts emerges. The first separates the AI that staff experiment with from the AI that runs business processes, so that each can be judged by its own logic. The second treats the company’s 160 years of institutional knowledge as infrastructure to be made ready for machines. And the third is discipline about where AI goes and what it does: most workflows stay deterministic, humans keep accountability and shared platforms trump shiny point solutions.
Know where the test kitchen ends
In most workplaces, AI can feel ubiquitous and absent at once. Employees use it to draft emails, summarise meetings and get past the tyranny of the blank page, yet it rarely changes how a firm makes, ships or sells things. Finance chiefs see bills for tokens—the chunks of text models process and suppliers bill for—but struggle to see what they bought. The trouble is that much of this AI activity still sits in the margins of real work as companies often bolt the technology onto old processes.
One way Nestlé has overcome this challenge is by dividing AI into two categories—tools for “personal productivity” and applications for “functional AI”. This split can help firms survive the awkward middle stage of adoption by allowing each type to be managed according to its own logic, easing the concerns of finance chiefs and giving workers room to experiment at the frontier.
Personal AI at Nestlé began in 2023 with NesGPT, an in-house chatbot. Today more than 120,000 employees use AI each month, saving minutes on drafting and summarising and becoming increasingly AI-savvy. “The more exposed you are to large language models (LLMs), the greater understanding you have of them,” notes Mr Mora. Familiarising staff with these tools, Nestlé says, is a means to a greater end: to sharpen decisions made with AI’s help and prepare teams for work rebuilt around the technology. (Read more about how experimentation helps firms discover useful applications of AI in our series.)
Functional AI demands a fundamentally different approach, with a sharp focus on business architecture. Instead of hunting for isolated “use cases” department by department, transformation at Nestlé begins at the enterprise level. The firm is governed through nine sweeping, cross-functional end-to-end processes, such as “order-to-cash” and “ideate-to-scale”, designed to stop teams optimising in isolation. “If you optimise one part of the business but neglect another, you end up with something like a car that has a great engine, but wheels so bad that you can’t really drive,” says Mr Mora.
“The more exposed you are to LLMs, the greater understanding you have of them.”
Enrique Mora
Principal enterprise architect for AI, Nestlé
Principal enterprise architect for AINestlé
Use AI where it matters
One of the hardest parts of adopting AI is identifying the tasks the technology can take over and those that are better left to humans. Mr Mora breaks the problem into three parts. “Any process that you can imagine,” he says, “is a combination of talent, imagination and knowledge.”
Talent and imagination, he argues, remain human-led. LLMs can recombine existing data, but they lack true causal reasoning. “If I just read books, I can learn a lot of things,” says Mr Mora. That is not the same as grasping “the real causal effect” of what happens in a business by acting in it.
Knowledge is the dimension AI can most usefully extend, but only if the company prepares it properly. Nestlé possesses a century and a half of institutional knowledge, from formulae and packaging rules to consumer habits and local exceptions. Much of it already exists in what Nestlé describes as its digital backbone: databases that are governed by common systems and processes across markets and functions. The next step, made easier because of Nestlé’s trust in its digital foundation, is to make the knowledge “agentic-ready”. Mr Mora calls that a “context layer”: a business system that knows which documents, rules, exceptions and precedents matter, so that “[the LLMs] generate more grounded and relevant outputs” for humans to review and act upon.
A good early example is Nestlé’s “ideate-to-scale” process. A packaging designer working on a product for Spain can use AI to generate visual concepts, but the model is not left to improvise from the internet’s idea of a food label. It draws on approved master data and local rules, so the proposal reflects the consumer information requirements that apply in that market. AI speeds up the sketching, and qualified Nestlé staff supply the judgment and then carry out regulatory assessments and validation, as well as taking the responsibility for final approval.
Some knowledge, though, resists capture. Michael Polanyi, a philosopher, observed that humans know more than they can usually tell. Corporate knowledge is similar and some of it will remain stubbornly tacit, from the taste of an experienced product developer to the feel of a market and the reason a technically correct design still looks wrong. No context layer is likely to replace how humans use their knowledge but it could narrow the distance between scattered corporate know-how and its practical use.
“Any process is a combination of talent, imagination and knowledge. AI helps us retrieve knowledge; people bring the talent and imagination.”
Enrique Mora
Principal enterprise architect for AI, Nestlé
Principal enterprise architect for AINestlé
Have a break, have a KitKat, have a guardrail
A mad rush to deploy AI has gripped the corporate world. Nestlé's instinct to take measured steps, formed in a sector where a wrong label can do real harm, can be instructive. Instead of pursuing a future where AI agents run the company with little to no human intervention, its vision is one of “process automation powered by AI and overseen by humans”.
Between 80% and 90% of all workflows at the company, says Mr Mora, “remain deterministic, with only the rest powered by gen-AI”. The strongest reason for this approach, in which AI performs only tasks it is “demonstrably good at”, is that “with food, you cannot take any risks”. Mr Mora describes how every important loop in Nestlé’s processes ends with a human checking the output—and taking responsibility.
Governance gives that caution teeth. When ChatGPT launched in late 2022, Nestlé created an AI committee that brought together specialists in responsible AI, lawyers, compliance teams and security experts. It blocked high-risk applications of the technology, such as using it to screen CVs, which can introduce or reproduce the bias of some LLMs and breach Nestlé’s ethical principles. Sensitive work may be assisted by AI, Nestlé decided, but decision-making and accountability remain squarely in human hands.
Cost is another reason for discipline. If a basic computer script can run a process, says Mr Mora, reaching for an LLM is pure extravagance: “no tokenmaxing,” he says, invoking a techies’ term for conspicuous token consumption. The same logic governs the choice of technologies. He describes the temptation of “the extra 5%”—a niche tool marginally better than the company’s existing platform. A small firm can indulge that temptation but at Nestlé’s scale, that choice becomes a tax on coherence. He calls it “chaotic” if ten of Nestlé’s markets were to buy ten slightly different tools, requiring the company to pay again and again for procurement, legal review, security approval, integration and maintenance. Such fragmentation would also make the company less connected. Data would flow less easily and AI agents would struggle to reach the context they need.
Nestlé’s answer is a platform-first strategy, favouring capabilities that fit and strengthen the shared enterprise stack over one-off tools. Some features arrive more slowly than they would from a specialist vendor—but when they do arrive, Nestlé has avoided a whole load of duplicate effort. The food giant is also wary that at its scale, not every AI decision can be taken centrally. Mr Mora’s team sets the direction and the guardrails, but regional and local teams lead the development of AI applications that best suit them. In a company Mr Mora compares to a small country, AI must spread without splintering the state.
“Any process is a combination of talent, imagination and knowledge. AI helps us retrieve knowledge; people bring the talent and imagination.”
Enrique Mora
Principal enterprise architect for AI, Nestlé
Principal enterprise architect for AINestlé
Building AI-ready organisations
Much of corporate AI adoption oscillates between two errors: unbounded enthusiasm, which produces demos and token bills, and reflexive caution, which often produces nothing at all. Nestlé’s frame is an attempt to avoid both—to let the test kitchen run riot while the factory keeps to the recipe. The payoff remains hard to measure and a software startup may sensibly tolerate more tools and more failure than a foodmaker can. But Nestlé’s wager is that as AI use matures, advantage will accrue to companies whose knowledge, processes and governance are already in order. In cooking as in computing, mise en place comes first.
Building AI-ready organisations, the first instalment of our Enterprise AI in action series, also includes an article describing what Erik Brynjolfsson, a leading economist of the digital age, thinks it will take for a firm to gain from AI, as well as an article on how IKEA, the world’s largest furniture retailer, approaches the technology through the same principles that underpin its products and an article on how LVMH, the world’s largest luxury group, is using AI to strengthen the human core that makes it distinctive.
