Applying AI in Sustainability:
Mixing Data, Protein, and Generative AI in One Powerful Shake
Why Measure Plant-Based Protein
Across the food sector, a transition is taking place. There is growing evidence and recognition among research institutions, governments and civil society that our current food system is not sustainable. Part of the concern are the impacts on human health and the environment associated with the production and excessive consumption of animal protein products such as red and processed meat. The EAT-Lancet Commission is a global initiative by leading sustainability and human health experts that has proposed the concept of a Planetary Health Diet, at the center of which is the idea of balanced consumption of plant- and animal-based products. To start on the transition, companies first need to understand their range and sales mix.
Current Measurement Approaches
In the Netherlands, most food retailers and supermarkets measure plant-based protein using the Eiweet (“Protein”) methodology, a framework developed jointly by a Dutch NGO ProVeg Netherlands and the Green Protein Alliance. It is based on a flowchart used to classify products into four protein categories: animal-based protein product, plant-based core protein product, plant-based non-core protein product, or a combined protein product.
This current approach presents practical challenges. The classification logic depends heavily on a lengthy appendix of general product categories clustered into four types of protein groups based on their ingredients. For example, beans are a plant-based core protein product because they are a primary source of protein and a plant, whilst an apple pie is a combined protein product because it has apples (plant protein) and butter (animal protein).
In theory, this list should make product assessment easier; in practice, it requires manually reviewing each article to understand which protein group it falls under and why. Applying these rules to analyse Picnic’s assortment means navigating a two-step manual process: first understanding the rationale behind the categorization, then executing it product-by-product. This makes automation difficult and introduces room for human interpretation and error.
Transitioning to AI Integration
To overcome these limitations, we explored how Artificial Intelligence could assist with applying the methodology more efficiently and consistently. Our main motivation was to reduce reliance on manual steps by integrating the information about product groups from the Eiweet methodology’s appendix as examples within the AI prompts, guiding its reasoning. For this proof of concept, we used the Gemini extension for Google Sheets.
Our first design principle focused on breaking the problem into solvable tasks: rather than asking AI to solve everything at once, we divided it into focused sub-tasks. Each sub-task had to be narrow enough to be solved within a single prompt. To aid model performance, we ensured the input text was in English, prompts were concise but example-driven, and edge cases or exclusions were explicitly stated.
This laid the groundwork for what we now refer to as our AI × Protein Analysis methodology, a hybrid approach that preserves domain rules while using AI prompting to scale their application.
Our AI X Protein Analysis Methodology
With these principles established, we developed a step-by-step process that integrates the Eiweet framework rules into AI-assisted decision-making. Rather than redesigning the classification logic, our aim was to translate it into a series of prompts that could be executed systematically in Google Sheets.
1. Data Preparation
We first standardized the ingredient list by translating it into English and removing may contain disclaimers that had caused incorrect classifications during early tests.
2. Detecting Protein Sources
AI was then prompted to assess protein sources in two distinct steps. One prompt checked whether the product contained animal-based protein, while a separate prompt evaluated the presence of plant-based protein. Each prompt was supported with examples of common animal- and plant-based protein sources to guide its reasoning.
3. Distinguishing Core from Non-Core Plant Protein Sources
Next, we evaluated whether any identified plant proteins should be considered core sources. This was the most difficult step, as the Eiweet methodology leaves the distinction between core and non-core plant proteins relatively open for interpretation. Hence, we defined a more specific prompt enriched with examples to help the AI make an informed decision.
4. Final protein label assignment
Outputs from each step were combined through simple conditional logic mirroring the Eiweet flowchart to produce the final protein label.
When the Model Spilled the Beans
As we tested our prompts, we encountered a few unexpected behaviors that offered useful insight into how the AI model interprets ingredient lists. One recurring example was the treatment of honey. Although honey is not considered a plant-based protein source under the Eiweet methodology, the AI occasionally flagged it as one, likely because it associates honey with plant-derived inputs, ignoring the fact that it is primarily a sugar.
Another pattern emerged with dairy ingredients. Whenever the AI encountered terms like milk or yoghurt, it consistently classified the product as animal-based, even when the ingredient referred to coconut milk or another plant-based alternative. Without explicit guidance, the model tended to default to the most common interpretation, which in this case meant assuming a dairy origin.
These observations reinforced the importance of prompt engineering and guidance under well-chosen examples. They also highlighted where human oversight is still needed to ensure edge cases are interpreted correctly.
Value, Limitations, and Future Directions
Integrating AI into our protein analysis workflow has shown clear value. The prompts allow us to apply the methodology more consistently, reduce manual effort, and scale the classification process across a large assortment.
At the same time, the current setup comes with practical limitations. The Gemini extension imposes a 350-cell limit, which restricts how much data can be processed at once. Some edge cases still require manual interpretation, and the model’s behavior can depend heavily on prompt clarity and examples.
Looking ahead, there are several opportunities to strengthen this approach. One improvement is to move beyond spreadsheet prompts and integrate an AI model directly through an API, enabling more scalable usage. This way, we can set up an edge model that “talks” directly to the AI, allowing classifications to run automatically on much larger datasets without relying on manual spreadsheet setups.
Another opportunity is to experiment with alternative prompting strategies and compare their output quality. For example, the same product could be classified using multiple prompt formulations, and the results could then be compared to see which approach produces the most accurate and consistent labels. These steps would help better understand the trade-offs and build a more robust methodology for classifying protein sources.
Curious to learn more about Picnic’s Sustainability initiatives? Check out our website.
Applying AI in Sustainability: Mixing Data, Protein, and Generative AI in One Powerful Shake was originally published in Picnic Engineering on Medium, where people are continuing the conversation by highlighting and responding to this story.
Janneploeger
