A Real Use Case: How We Used AI to Create a New Ingredient
From idea to concept in hours, not weeks—powered by AI and domain-specific intelligence.
When people hear about AI in food innovation, the first thought is often: “Isn’t that just a fancy name for Googling recipes faster?”
Not quite. True AI-driven innovation is about more than speed; it’s about depth and precision. Instead of manually searching through fragmented databases, nutrition papers, and regulatory guidelines, AI can combine these elements into one coherent process—powered by structured data and a model that understands the context of food science.
In this case study, I want to share how we used AI to conceptually design a new snack ingredient—purely as an experiment—and how we achieved in hours what traditionally takes weeks.
Step 1: Building the Right Foundation
AI is only as good as the data and logic behind it. For this experiment, we didn’t just throw a generic large language model (LLM) at the problem. We leveraged:
A curated ingredient database: This contained detailed data on nutritional profiles, functional properties, sensory characteristics, and regulatory status of hundreds of possible ingredients.
A domain-specific AI model: Unlike a general-purpose LLM, this model was fine-tuned for the food and nutrition space, meaning it understood not just words but concepts like glycemic load, emulsification properties, and dietary compliance.
The combination of structured data and specialized intelligence is critical—because generic AI often hallucinates. In food, hallucinations are dangerous, especially if they lead to unrealistic or unsafe ingredient suggestions.
Step 2: Defining the Objective
We gave the AI a clear brief:
Goal: Create a concept for a high-protein, shelf-stable snack that is plant-based and meets EU regulatory standards.
Constraints:
At least 15 grams of protein per serving
Less than 5 grams of sugar
Compatible with a clean-label positioning
Traditionally, a food scientist would research possible ingredient combinations, cross-reference nutritional tables, check for allergens, validate functional properties, and sketch formulations. That’s a multi-week process involving iterative tests.
We wanted to see if the AI could do this in hours, at least at the conceptual stage.
Step 3: The AI Workflow in Action
Here’s what happened in practice:
Data Retrieval with RAG: The model used Retrieval-Augmented Generation (RAG) to pull structured data from the ingredient database, ensuring all suggestions were rooted in real, validated facts—not guesses.
Formulation Suggestions: The AI proposed several combinations of proteins, binders, and natural flavorings that aligned with our brief. For example, it suggested pea protein isolate combined with chicory root fiber for functionality and sweetness balance.
Iterative Refinement: We then asked the AI to improve texture and sensory appeal without adding artificial ingredients. It responded with alternatives like adding aquafaba as a natural binder and using cacao nibs for crunch.
Within three prompts and less than an hour, we had a hypothetical formulation that met all the criteria. Was it ready to launch? Absolutely not. But the speed at which the ideation happened was impressive—and a strong indicator of what’s possible at the front end of innovation.
Step 4: Why This Matters
Food innovation is expensive and time-consuming. The early stages—where you go from “idea” to “first concept”—often determine whether a product moves forward or dies in committee. If AI can reduce weeks of research to hours of conceptual work, the implications are huge:
Lower costs at the ideation stage
Faster iteration cycles
Better use of R&D resources for testing only the most promising concepts
In our experiment, what traditionally takes weeks (and often involves multiple experts) was condensed into an afternoon of strategic prompting and review.
Important Caveat: Theory vs. Reality
This was a theoretical exercise—meaning we didn’t physically create the ingredient. Why? Because physical formulation, sensory testing, shelf-life validation, and regulatory checks are still essential and cannot be skipped. AI accelerates thinking and planning, not the laws of chemistry or biology.
What it does give us is a head start—allowing teams to move forward with better ideas, faster.
What’s Next?
The next frontier is integrating AI not just at the concept stage but through predictive modeling for shelf life, cost optimization, and even market forecasting. But none of that is possible without structured data and domain expertise. AI isn’t a magic wand—it’s a power tool, and like all tools, it works best in skilled hands.