A veggie-food AI designs recipes no human tasted first — how does that know-how become something the whole company can use?

This explores how an AI system's know-how becomes knowledge a company actually owns and uses. NotCo's Giuseppe is the example: it searches plant ingredients for combinations that reproduce the taste and texture of milk, mayonnaise or meat. One caveat first. None of the retrieved notes describes the Giuseppe case in detail, so this answer explains the idea the case illustrates rather than retelling it. The closest note is Can machine tacit knowledge spread across organizations unlike human tacit knowledge?. It takes the classic theory of the 'knowledge-creating company' and adds a new kind of tacit knowledge: the kind a model holds but cannot explain.

That note's main claim is the surprise. Human tacit knowledge, such as a chef's feel for flavor, stays with the person and spreads slowly through apprenticeship. Machine tacit knowledge is different: once a model has learned it, it can be copied anywhere at almost no cost. Giuseppe 'knows' that some odd pairing of plants tastes like dairy without being able to say why. That is tacit knowledge that scales. The same note adds a condition, though. Portable doesn't mean organizational. The model's suggestions only turn into company knowledge through feedback loops in which people taste, reject, refine and give names to what the machine proposes. The model supplies the hunch, and the people around it make it shared and meaningful. Whether a model can have tacit knowledge at all is still debated philosophically. Do language models possess tacit knowledge in Davies' sense? argues that it plausibly can, though the evidence so far is thin.

An industrial case study shows the same movement running the other way. In Can codified expertise let non-experts match specialist output?, a company wrote its experts' rules into the structure around an AI agent, and non-experts then produced work that experts rated highly. The quality gain came from writing expertise down into that structure, not from using a bigger model. Put next to Giuseppe, you get two halves of one loop. Machines produce insight that people turn into shared understanding, and people's expertise gets written into machine systems that anyone can then use. Knowledge moves back and forth between people and systems, and each pass makes more of it the company's own.

The critical notes explain why the human step can't be skipped. Does AI-generated knowledge have the same structure as hearsay? argues that AI output works like hearsay: it can't be traced to a source or checked directly. A suggestion from Giuseppe is a claim nobody can verify until a person tests it in the kitchen. farrell-and-colleagues-argue-large-ai-models-are-cultural-and-social-technologies sees large models as lossy summarizers of what humans already know, closer to markets and bureaucracies than to individual minds. On that view, Giuseppe becomes part of the organization the way a bureaucracy does: it gives knowledge a structure but doesn't understand it. Food testing is a useful case because tasting is the step that turns hearsay into evidence.

The less obvious lesson concerns what happens once machine knowledge enters the workflow. Does AI really compress all layers of knowledge work equally? argues that AI shrinks the 'doing' part of knowledge work, while deciding what to do and delivering the result stay the same or grow. Giuseppe fits that pattern. Searching through ingredient combinations gets much faster, but choosing which leads to pursue and getting a product to market still depend on people. Does easier tool-building actually solve enterprise adoption problems? adds that even strong capabilities stall without decisions made across departments. So the Giuseppe case is less about a clever model and more about building the routines that turn a model's hunches into decisions the company stands behind.