Shahar Rosentraub, Chief Product Officer at AKA Foods, evaluates the limitations of zero-shot AI in industrial food formulation, assessing why general-purpose language models lack the domain-specific grounding required for reliable R&D application — and identifies structured, data-driven deployment as the more scientifically credible path forward.

A step back in time

Ask ChatGPT or Claude to design a plant-based cheese with the texture of mozzarella, a clean label, regulatory compliance in three regions and scalable industrial processing.

It will give you an answer in seconds. But it won’t work.

And in food R&D, “won’t work” is not theoretical. It means wasted pilot runs, delayed launches, regulatory rework and erosion of team confidence in AI initiatives.

In theory, it sounds simple. If AI can generate poems, images and even software code, why not a new plant-based cheese or beverage formula? But industrial food engineering is not a language problem. It is a biological, chemical and sensory system operating under strict regulatory and commercial constraints.

This is where the idea of “zero-shot” prediction runs into reality.

Zero-shot refers to a model’s ability to solve a task it has never seen before. In food terms, that would mean generating a brand-new formulation, predicting texture, stability and consumer acceptance, without prior experiments or product-specific data. The appeal is obvious. If that were possible, we would effectively have an in-silico food lab.

But food systems do not behave like text or images.

Industrial food processing sits at the intersection of chemistry, biology and mechanical engineering. Small changes in temperature, shear, ingredient quality or processing time can produce dramatically different outcomes. Agricultural inputs vary by season and geo-
graphy. Phase transitions, enzymatic reactions and structural trans-
formations are highly non-linear. Much of the know-how that governs success is tacit, developed through years of hands-on experience.

A 1% shift in moisture, a change in supplier crop quality, or a minor adjustment in shear can mean the difference between a stable emulsion and a product that fails after two weeks on shelf.

Most importantly, food is judged by human perception. Taste, aroma, texture and mouthfeel cannot be fully inferred from composition alone. Even with perfect compositional data, we do not yet have perfect simulators of human sensory variability.

Large language models create the illusion that zero-shot generation is achievable everywhere. But these models are trained on general knowledge, not on structured, proprietary R&D datasets, detailed processing parameters, sensory panels or regulatory frameworks. Without grounding in domain-specific data, generative outputs risk being eloquent but disconnected from industrial reality. Internet-scale knowledge is not the same as industrial-scale formulation data.

Zero-Shot vs Grounded AI in Industrial Food R&D

This does not mean AI is not ready for food. It means that zero-shot, from-scratch formulation is the wrong benchmark. The near-term opportunity lies in structured, few-shot and hybrid approaches.

At AKA, we see this daily. When AI systems are grounded in a company’s own historical R&D data, ingredient performance, process parameters and sensory outcomes, they can dramatically accelerate iteration. The breakthrough does not come from asking a model to “invent” a product. It comes from structuring fragmented knowledge so that models can reason across it coherently.

In practice, that means unifying formulation data, processing records and sensory feedback into a consistent, machine-readable framework. Once grounded in that context, AI can surface non-obvious relationships, reduce redundant experimentation and support more confident decision-making. In some cases, that means identifying redundant experimental paths. In others, it means flagging formulation risks before they reach pilot scale.

In this model, AI does not replace food scientists. It amplifies them.

The real constraint in food and beverage is not computational power. It is fragmented knowledge. Critical data sits in silos across R&D labs, sensory departments, sales teams and ingredient suppliers. Unlocking value requires unifying that knowledge into structured, accessible systems that models can reason over responsibly.

Zero-shot AI in industrial food engineering remains an aspiration. But grounded, domain-specific AI is already delivering measurable gains.

The companies that move first will not be those waiting for a magical formulation engine. They will be those investing in structured knowledge, pragmatic deployment and systems that enhance human expertise rather than attempt to bypass it.

In a margin-sensitive industry where speed and precision define competitiveness, disciplined AI deployment will be a structural advantage, not a novelty.

About the author

Shahar Rosentraub is Chief Product Officer at AKA Foods, where he leads the development of AKA Studio, an AI platform designed to structure and unify proprietary R&D, formulation and sensory data for food and beverage companies.

With a background in product strategy and applied AI systems, he focuses on bridging advanced machine learning capabilities with the practical realities of industrial food development.