Are Big Food Companies Really Embracing AI?

While some companies like NotCo have positioned themselves as the OpenAI of the food world, the truth is that the AI transformation of the food industry is still in the first inning. That is partly because the food system itself, a mix of legacy CPG giants, agricultural suppliers, ingredient developers, and regulators, moves at a glacial pace.

In my recent conversation with Jasmin Hume, founder and CEO of Shiru, she confirmed that the industry is still in the early stages, in large part because food companies have massive amounts of data and strong confidence in their own research and development.

“Food companies have world-class R&D teams, and those scientists want to see proof before adopting a new tool. It’s a lot of tire-kicking in the first meetings,” said Hume.

This slow pace does not mean AI is not making inroads. It is simply happening beneath the surface. From discovery platforms like Shiru’s to optimization tools in manufacturing and retail analytics, AI is slowly reshaping how food gets made.

But the true question is not if the food system will use AI, but who will own the models that make it useful. The answer is usually tied to who owns the data. Legacy food and ingredient companies have decades, even centuries, of proprietary chemical, biological, and sensory data. This makes them both powerful and hesitant to engage with AI models that might use that data to build their own foundation models.

Big food brands “are not going to very quickly turn over that data,” said Hume. Many are debating whether to build their own in-house systems, using models fine-tuned on proprietary data that never leaves their servers. Others are beginning to explore partnerships with companies like NotCo and Shiru that specialize in the discovery layer.

That need for validation may be the biggest differentiator between food AI and other industries. As Hume put it, “You have to bring it into the lab and make sure that it actually works. Otherwise, the predictions are worthless.”

When Hume and I discussed whether large players like Microsoft or Google would eventually dominate vertical-specific foundation models, she acknowledged that possibility. However, she stressed that today’s large foundation models are not yet equipped to deal with the physical and regulatory realities of food. “There’s a ton of very specific know-how that goes into making those models usable for applications like protein discovery or formulation.”

For Shiru competitor NotCo, this highly specific data and domain knowledge are what the company is banking on to solidify its position as a key player in building a foundation model for food, a term now featured prominently on its website.

“I think what people need to understand is that AI is truly about the data sets and the value of the data sets that you have and the logic of the algorithms,” said Muchnick in an interview I had with him in July at Smart Kitchen Summit. “It’s really hard to get to where we were, and specifically also because we weren’t just an AI company. We are a CPG brand, and we learned a lot from being a CPG brand.”

In conversations with AI experts outside of food, several have said we are starting to see the big foundation models open up to allow companies to train them with vertical or highly specific domain knowledge. One pointed to Anthropic’s Model Context Protocol (MCP), which lets a foundation model connect to external data sets to process answers.

Another example is Thinking Machines’ newly announced fine-tuning API called Tinker, which could make it significantly easier for a food brand to train a model with domain-specific knowledge by removing the heavy infrastructure and engineering overhead typically required for custom AI development.

For Shiru, NotCo, and others developing food and ingredient-focused AI, there is still significant opportunity because the field is still so early.

“We’re just starting to see companies thinking about their own internal instances,” said Hume. “A lot of this is in progress, boardrooms are having these discussions right now.”

One of the biggest holdups for food brands is that data ownership and business-model alignment remain unsolved. Who owns the training data and the resulting outputs is a key question, and without clear answers, many companies will hold their data close, limiting the ability of shared platforms to reach critical mass.

For that reason, Hume believes partnerships and licensing models, not open data exchanges, will drive progress in the near term. Shiru’s model focuses on IP discovery and licensing, which allows the company to build intellectual property value without requiring massive manufacturing investments. “Our IP portfolio has doubled year over year since 2022,” said Hume. “Now the focus is on monetizing that through licensing and sales.”

The topic of food-specific foundation models and the adoption of AI by food brands is a fast-moving one, so you’ll want to make sure to listen to this episode to get caught up. You can listen to our entire conversation below or find it on Apple Podcasts, Spotify, or wherever you get your podcasts.

Related