How L’Oréal, Mondelēz, and Nestlé Are Building Real-World AI Systems

ByDivyanshu Rawat

Hey everyone! If you’ve been following my recent updates, you know I’ve spent the last few years building full-stack web applications using the MERN stack (MongoDB, Express, React, Node) out here in Mohali, Punjab. Lately, though, I’ve been diving deep into Agentic AI—exploring how autonomous agents, tool integration, and domain-specific models go way beyond basic API wrappers and conversational bots.

While skimming through industry news this week, I stumbled upon a series of updates from global giants like L’Oréal, Mondelēz, Nestlé, and Haleon. They aren’t just using AI to write marketing copy—they’re running digital simulations to formulate shampoos, tweak Oreo recipes, and design eco-friendly packaging.

As a web developer learning how AI agents make decisions and run tool calls in complex workflows, reading about how these traditional FMCG companies operate was a huge lightbulb moment for me. Here is a breakdown of what these enterprise R&D labs are doing—and my takeaways as a software developer looking at the technical side of it.

L’Oréal: 4x Faster Product Formulation with Predictive Science

L’Oréal has actually been quiet about this, but they’ve been applying predictive AI in their research labs for about four years now. Fabrice Megarbane (President of L’Oréal’s Consumer Products Division) recently mentioned that their technology predicts how specific molecules will interact with skin and hair before any physical lab mixing happens.

Instead of spending months mixing raw ingredients in test tubes, their scientists run digital simulations to test thousands of variables upfront.

  • The Real-World Result: They recently took molecules previously reserved for skincare products and repurposed them to create a collagen-based shampoo designed to add fullness to hair.
  • The Impact: L’Oréal reports that predictive science has made product formulation 4x faster.

My Developer Take: In web dev, when we test an API, we often mock responses or write integration tests before deploying to staging. What L’Oréal is doing is essentially running “unit tests” on chemical properties. By filtering out poor molecular candidates programmatically, their human scientists only spend lab time on the highest-probability formulas.

Mondelēz: Generative Recipes & Supply Chain Flexibility

If you love Cadbury, Toblerone, Oreos, or Chips Ahoy, you’re already eating food shaped by AI algorithms. Filippo Catalano, the Chief Information and Digital Officer at Mondelēz, shared how their internal AI tools generate novel recipe ideas—even unusual ingredient combinations—before chef and food scientists review them.

A few key highlights from their setup:

  • Fewer Physical Batches: The AI tool reduces the volume of physical baking samples required during R&D.
  • Tested Winners: Mondelēz used this system to help develop Gluten-Free Golden Oreos and their recent Chips Ahoy recipe update. They noted that 60% of biscuit recipes generated by the tool scored better across nutrition metrics, sustainability goals, and cost targets.
  • Solving Supply Chain Hiccups: Ingredients fluctuate constantly in price, quality, or availability due to weather or logistics. If a single supplier runs dry, Mondelēz uses AI to quickly find alternative recipe adjustments without altering taste or texture.

My Developer Take: This sounds remarkably like a constraint satisfaction problem combined with generative pipelines. In code, when we build an agent, we give it rules (e.g., “Max budget: $X”, “Must be gluten-free”, “Must match target crunch factor”). The agent evaluates candidate combinations against those bounds. Seeing it applied to baking chocolate chip cookies is brilliant.

Nestlé & IBM: Chemical Language Models & Clean Labeling

Nestlé is currently tackling a massive operational challenge: removing all artificial food colorings from their products globally by the end of 2026 (building on what they’ve already accomplished in the US).

Finding natural color substitutes isn’t as simple as swapping dye A for dye B—it affects shelf life, heat stability during production, and taste. CTO Stefan Palzer noted that AI is helping screen natural alternatives at scale.

Beyond food formulas, Nestlé partnered with IBM Research to build a generative AI tool for sustainable packaging:

  • The Tech Stack: The project uses chemical language modeling combined with IBM Research’s regression transformer.
  • What It Does: It links raw molecular structures of potential packaging materials with physical-chemical properties to discover high-barrier materials (protecting food from moisture, oxygen, and temperature) while factoring in recyclability and cost.

Other Big Moves in FMCG & Health

It’s clear this isn’t an isolated experiment:

  • Barry Callebaut + NotCo: The chocolate maker joined forces with NotCo to run AI-driven plant-based ingredient matching for chocolate formulas.
  • Haleon + Microsoft: Consumer health leader Haleon locked in a 5-year partnership with Microsoft to deploy AI across clinical content development, R&D forecasting, and supply chain management.

My Key Takeaways as a Developer Learning Agentic AI

When you first start building AI applications, it’s easy to get stuck thinking in terms of standard CRUD apps with a chat box stuck on the side. But studying these enterprise implementations opened up three major insights for me:

1. Human-in-the-Loop (HITL) is Non-Negotiable

Notice how none of these companies replaced their food scientists, chemists, or packaging engineers? Filippo Catalano from Mondelēz specifically emphasized that human experts review every single AI recipe idea before it enters physical testing. The AI acts as a candidate generator and candidate ranker, handling the heavy lifting of combinatorial search.

2. Domain Models > Generic Prompts

You can’t ask a general LLM to accurately predict the oxygen barrier property of a biodegradable polymer. Nestlé and IBM had to use domain-specific chemical language models and regression transformers. When building AI agents, matching the right specialized tool or sub-model to the right sub-task is critical.

3. The Power of Simulating Before Execution

Whether it’s L’Oréal testing skin-molecule interactions or Mondelēz swapping out a single ingredient due to supply chain shortages, the core value proposition is compressing feedback loops. Reducing months of trial-and-error down to minutes of computation saves huge capital costs.

Wrapping Up

Moving from traditional MERN stack architecture into agentic AI systems has been an exciting journey so far. Seeing how real-world industries—from beauty to food manufacturing—are structuring their AI pipelines gives me a much clearer picture of how to architect state machines, tool calls, and automated evaluation loops in code.

What are your thoughts? Are you seeing AI tools integrated into your industry’s core workflow, or is it mostly hype in your space? Let’s discuss in the comments below!

You May Also Like