Wineries
AI in the cellar: from Pommard to Italy, it works for one winery
Two different tools, built for different problems, but converging on the same lesson: AI works only when it learns from a single winery.
For decades the wine world has welcomed every technological innovation following the same pattern. First comes distrust, then the cautious trial, and finally silent adoption once nobody notices anymore. In recent months the same pattern is repeating itself with artificial intelligence, but at a different scale. In France, Joannès Violot Guillemard, a winegrower in Pommard, built on his own the application that manages the winemaking of his estate. The estate has 8 hectares of vines plus a négociant business, for a total of 36 cuvées in full fermentation. In Italy, Francesco Gaspari has developed AgentoWine, a conversational assistant designed for wineries that receive visitors. Two different tools, built for different problems, but converging on the same lesson: AI works only when it learns from a single winery.
The law applies to a single winery, even when the Excel spreadsheet disappears
The Pommard application was tested during the 2026 harvest and adopted immediately. The winegrower dictates the temperature, density and acidity of each cuvée into his phone, and the data is recorded and sorted automatically. The system calculates the quantities of products to be added and sends alerts such as: cuvée number 30, add such and such grams of tannins. Everything appears on a touch screen installed in the cellar. The practical gain is clear: no more scattered sheets of paper, and staff are no longer reprimanded for wrong or forgotten annotations.
Similar tools already existed on the market. But the existing professional package cost 100,000 euros, a figure out of reach for an estate of that size. The homemade alternative cost 1,000 euros in subscriptions to AI services and an application-building program. It also required development time: three hours a day during the month of July, working alone. The winegrower is not an IT technician. But he has already built a second tool in 30 minutes to coordinate orders and invoices. His mother, who handles the administration, would have needed three days for that task.
There is, however, a limit that technology has not solved. By law, in France, winemaking records must remain on paper, handwritten. The digital system therefore coexists with the administrative requirement, and the double manual annotation remains the real bottleneck. Anyone evaluating similar tools must keep two levels distinct: that of operational efficiency, where the gain is immediate, and that of regulatory compliance, where nothing has changed. The digitalization of the vineyard does not eliminate bureaucracy, it doubles it, until the regulations catch up.
Artificial intelligence works only in a closed loop
The most important point of contact between the two cases concerns content control. The Pommard app works in a closed loop. The winegrower entered his own winemaking criteria. The tool does not go looking for protocols or manuals elsewhere. AgentoWine follows the same logic on a different plane: the assistant does not draw on generic answers from the internet. It uses exclusively the content provided and approved by the single winery.
This architecture answers the most serious problem of conversational assistants: the invention of information. Prices, real availability and allergens are topics where an error has direct consequences on customer trust. When the system does not have enough data to answer with confidence, it does not formulate a hypothesis: it forwards the request to the winery and flags in the management panel that knowledge on that topic is missing. The assistant declares its limits instead of hiding them, and the winery fills the gaps by updating the content. The guarantee, however, must be understood precisely: the method does not prevent errors in itself. If the uploaded data is wrong or outdated, the wrong answer arrives with the same confidence. The closed loop guarantees consistency with what the winery has approved, not absolute truth.
In the end, I do it the way I feel it. It’s like the consulting oenologist, whom I never listen to. Luckily, otherwise all wines would taste the same.
This remark from the Pommard winegrower overturns the most widespread oversimplification, that of AI deciding in the operator’s place. Quite simply, it decides in no one’s place: it records, calculates, remembers. The part that requires sensitivity remains human. Gaspari has imposed the same boundary on the hospitality side. Some wineries asked him to install a totem with the assistant in the tasting rooms. He refused: when the visitor is inside the winery, the relationship must be human, because that is where the bond with the business is built.
What it does not guarantee, and what instead surprises in the data
AI in the cellar does not guarantee the replacement of staff, and the promoters of both tools say so openly. AgentoWine, in the words of its founder, was born not to replace human time but to free it. It handles repetitive requests. It brings visitors to the winery already oriented. The declared ethical limit is precise: the assistant accompanies the guest up to the threshold, then steps aside.
The data collected has already produced discoveries that no one had foreseen. Across all the client wineries, almost 40% of conversations happen outside opening hours, at night or on weekends. At those moments there is no one in the winery to answer. Before the assistant, those questions were lost. The system also intercepts silent visitors, who would have browsed without leaving a trace, and reveals unexpected needs. Some wineries have modified their wine tourism packages precisely on the basis of the questions received. The differences between nationalities are marked. The German visitor asks about hours, travel times and parking with precision. The English-speaking one asks open questions and asks for help choosing the experience suited to a broader itinerary.
Gaspari tells the origin of the project through a concrete episode. In December 2024, during a tasting in a Piedmontese winery, a couple arrives. The couple was looking for a Barolo from the current vintage, to give as a gift for a baby’s birth according to tradition. They had received no reply to a booking email, and had shown up in person. But that vintage had been harvested only a few months earlier: the bottle they were looking for did not yet exist. Nobody had made a mistake on purpose, the email had ended up in a full inbox. The problem was the structure, not the winery.
What changes from here
The two experiences point in a similar direction. The Pommard winegrower has already been approached by colleagues, and gladly shares his model. But he warns that every business must teach the AI its own products and protocols. Gaspari is working on extending the assistant beyond the website, to email, messaging and social media, where requests currently arrive fragmented, always with supervision from the winery. The question of legal recognition remains open: as long as paper records remain mandatory, the real advantage remains operational, not administrative. And the basic criterion remains valid for anyone evaluating these tools. One must always ask what data feeds the system, who approved it, and what happens when the system does not know.
Sources: