AI helps you discover the wine that’s right for you

In 2021, Vivino built AI on top of all their review and profile data to create algorithms that generate a personalized match score for every user and every wine.

In this series of articles, I have been analyzing the relationship between wine, personalization and AI. This is the first article. In the second post, I have been considering nudge marketing for wine, an experience goods. In the last article, I have been analyzing benefits of AI for winemaking.

I had the pleasure of interviewing leading players in the wine and the tech industries. For this chapter, I have discussed with (sorted by alphabetical order):
• Katerina Axelsson Co-Founder & CEO at Tastry
• Vijay Bhagwandas Co-Founder & CEO at Tasting Intelligence
• Pam Dillon Co-Founder & CEO at Preferabli
• Magalie Dubois Assistant professor at Burgundy School of Business
• Julian Perry CEO at Wine-Seacher
• Tristan Rousselle Founder & Deputy CEO at Aryballe
• Charles Slocum Chief Business Officer at Tastry
• Gérard Spatafora Managing Director at E-Studi’OZ.wine
• Andrew Sussman Chief Technical Offer at Preferabli

I also contacted 7 wine critics. None of them replied.

Artificial Intelligence (AI) in the Food & Beverage industry valued at USD 3.07 billion in 2020 has seen tremendous growth. It is expected to reach USD 29.94 billion by 2026 (Mordor Intelligence, 2021). Some of the key players include TOMRA Sorting Solutions , GREEFA, Sesotec, Key Technology, Raytec Vision, Rockwell Automation, ABB, Honeywell International, among others.

“Wine appreciation: connoisseurship or snobbery?” by Barry Smith in “The World of Fine Wine” magazine is one of the classics of wine critics’ analysis. It was in 2014, updated in 2018. The article starts like this:

“We don’t have to venture far these days to realize that all is not well in the world of wine criticism. The popular press has seen a surge of recent articles detailing how easily wine experts are fooled and connoisseurs misled. The blogosphere is virulent in denouncing what bloggers see as the ludicrous talk of wine critics. Even the pages of this magazine have seen their fair share of soul-searching about the point and purpose of wine writing.
Behind these complaints and concerns lies the suspicion that wine talk lacks substance; that the words we use to describe the taste of wines do not engage adequately or accurately with a genuine subject matter. From here it is a short step to the dismissal of expert opinion and, along with it, the presumed sensibilities of the connoisseur.”

The beauty of wine is to try something new.

Julian Perry, CEO at Wine-Searcher

AI in the drinks industry is helping brands by obtaining the data from their past records and strategizing it with the help of AI-implemented algorithms. This can help in predicting sales, understanding the behavior of consumers, managing the supply chain, and a lot more which can lead to the growth of companies. AI has helped a large number of players in the drinks industry to tap into their consumer preferences, get actionable insights and personalize the experience for their consumers.

So interactive labels, recommendation algorithms, quality assurance, or even AI-generated blends have become a reality. This is just the beginning. The future of tech in the beverage trade is a big conversation to have.

1. AI Sommelier is not AI but Digital Sommelier (Smart Search)

When I was still living in Hong Kong, as director of a wine research institute, I had the pleasure to meet David Garrett, then founder of Entaste. This was a platform “connecting the different links of the wine industry value chain (wineries, restaurants, wine shops, consumers)”. The Entaste platform was the latest high-tech trend in restaurant wine list consultation. It was an iPad-based digital wine list intended to “help wine lovers make smart choices.” This was in 2011. Helping customers to choose a wine, in the form of interactive wine lists thanks to the technology of the moment goes back a long way. It was a virtual sommelier, using algorithms to suggest “the best wine” according to the customer. This app no longer exists.

In 2018, Clara Wine gained popularity as the first AI wine subscription service. To get started, subscribers take a quiz on the Clara Wine website 2018, which categorizes each user into one of 12 different groups based on their aroma, flavor, and body preferences. This app does not exist anymore.

We have not found one solution as a clear winner.

Julian Perry, CEO at Wine-Searcher

Most computer-generated wine recommendations have traditionally been pretty much worthless. Jeff Siegel, wine critic and blogger at Wine Curmudgeon, received an email from America’s leading online retailer: wine.com. The email said that Jeff Siegel would like a bottle of Veuve Clicquot for $64, as well as a Kim Crawford sauvignon blanc for $20, a California pinot noir for $12 and a sweet red blend for $12. Jeff Siegel doesn’t like any of the suggested wines.

The formulas used to make these recommendations depend on things that aren’t necessarily wine-related — such as income, age and gender — and use information about past purchases probably more than they should.

Artificial Intelligence (AI) is used when a machine mimics the cognitive functions that humans associate with other human minds. Compared to human-programmed intelligence, AI is able to create its own algorithms through the process of Machine Learning (ML).
AI is a big data hound — effective construction and deployment of AI and ML systems requires large data sets to recognize patterns.

While machine learning is based on the idea that machines should be able to learn and adapt through experience, AI refers to a broader idea where machines can execute tasks “smartly.” Artificial Intelligence applies machine learning, deep learning and other techniques to solve actual problems.

Ask how specifically does your system learn? and see how often their software is simply a function of straightforward, automated programming.

Pam Dillon, Co-founder and CEO at Preferabli

AI sommeliers are digital sommeliers a.k.a smart search engines. They do not use AI per se as such machine learning or Large Language Models (LLMs) / deep learning techniques. They are just algorithms. Others call it ‘augmented intelligence’. A number of personal sommeliers don’t always have a good reputation. Sippd, Bright Cellars, Matcha, or SommSelec report using science or an expert to avoid confusion in wine buying.

The goal is to instill a confidence to ask more questions, making more wine choices without intimidation.

2. AI recommendations based on crowdsourced opinion (Collaborative filtering)

Mass market segmentation in the wine and spirits industry has taken the shape of personas using characteristics such as age, knowledge, income and gender.

“What’s Your Whisky?”, a digital app recently launched by the beverage producer Diageo, is a prime example of exactly how this can work. It is designed to help consumers discover their favorite whiskey based on their answers to a series of questions about their personal taste and flavor preferences.

IBM has created an artificial intelligence system for food giant McCormick & Company that compiles a huge database of ingredients, flavour formulas, consumer satisfaction tests and sales figures in various markets to build algorithms that lead to concrete creations of seasonings. A titanic collection work that would be impossible for a human to perform and memorize.

Tasting Intelligence is an aggregator of consumer comments and ratings. It analyzes tasting notes and sentiments from online tasting reviews and Social Media posts with the assistance of Artificial Intelligence. Buyers can get a better understanding of tastings with the latest tasting reviews throughout the world (tastings, sentiments, emoji).

In the end, it is meant to help with discovering flavors and to give you a modern feeling with tasting experiences.

Vijay Bhagwandas, CEO at Tasting Intelligence

In 2021, Vivino built AI on top of all their review and profile data to create algorithms that generate a personalized match score for every user and every wine. Until now, Vivino users primarily relied on the Vivino Rating, a five-star wine rating system to show the average score of a wine. “Match for You”, however, uses machine learning and the insights Vivino has on each user to give each one of them a unique match score for every single wine in the world — all 13 million of them.
Once you’ve rated at least five wines, every wine in the database will have a “match score” that’s generated for consumers, letting drinkers know if there’s a high, medium, or low chance that they’ll like it. These match scores are generated specifically for you. The feature is available for all users once they rate five wines. “Match for You” brackets are:

• Great Match (70%–100%): These are wines that are sure to please, and will likely get rated four stars or higher.
• Average Match (40%–70%): These wines may not be absolute favorites, but could be worth trying.
• Low Match (0%–40%): Based on Vivino’s data, it is unlikely these wines will be enjoyed.

CEO Heini Zachariassen believes that the wine industry will be able to build on the big data Vivino and other wine tech companies have gathered, and use machine learning and artificial intelligence to continue to push the boundaries of the industry in the future.

“I see data-driven AI as a first-level vision in an AI strategy for wine discovery issues. Other AI techniques, especially those based on senses, seem to me to be future technologies for the subject that concerns us: wine recommendations.” – David BECK.

3. AI recommendations based on individual preferences (Content-Based Systems)

A content-based system analyzes both user and product data in detail. Only content-based, AI-driven recommenders are intrinsically focused on individual preferences, and so are best suited for the wine and spirits industry.

In October 2020, the world’s largest flavor and fragrance creation company, Firmenich, announced to great fanfare that it had just achieved a natural grilled beef flavor designed entirely by artificial intelligence.

Researchers at Carlsberg, one of the world’s largest brewers, are using sensors and analytical tools to map and predict flavors more quickly. The goal: to create a catalog of flavors from samples that will help brewers become more efficient when developing blends. According to Carlsberg, its AI is already able to tell the difference between a lager and a beer.

Our mission is to build software that behaves like human being.

Andrew Sussman, Co-founder and CTO at Preferabli

Preferabli captures the individual preferences of wine, beer and spirits shoppers, enabling merchants to provide personalized shopping experiences. Preferabli analyzes the digital signal, including consumer ratings and sales, and makes recommendations based on the inventory.
With a dozen Masters of Wine and Master Sommeliers, Preferabli analyzes each product taking into account more than 700 characteristics.
“Our machine learning allows shoppers to say what they like, even love, and merchants to understand and present what each buyer really wants. We think of it as artificial intelligence with a human touch,” said Pam Dillon, co-founder and CEO of Preferabli.
Preferabli offers a suite of software solutions that integrate with merchants’ existing systems to increase sales, optimize inventory and purchasing decisions, and build customer loyalty.
With a dozen Masters of Wine and Master Sommeliers, Preferabli analyzes each product taking into account more than 700 characteristics.

Since research into robotics began, humans have been trying to create machines in their own image. From machines that can move like humans do to ones that can feel and think as people do, humans have long been in search of ways to, for all intents and purposes, recreate themselves in machine form.

While this may have seemed an unattainable goal in the past, systems are starting to surface that can differentiate between as many tastes, smells, and textures as any human. One institution dedicated to developing human-like machines is the Massachusetts Institute of Technology (MIT).

Consumer behaviors and preferences for the food and beverage industry have been evolving dramatically over the past decade. From the movement for more natural ingredients to more discerning taste palettes, food and beverage manufacturers are striving to improve product quality controls across the manufacturing process.

The 5 flavors that shape taste are sweet, salty, sour, bitter and umami. The rest is from the nose.

Tristan Rousselle Founder & Deputy CEO at Aryballe

Aryballe combines biosensors, silicon photonics and machine learning to deliver digital olfaction solutions. Digital olfaction is generally defined as the digital capture and production of aromas. Similar to our sense of smell, it mimics the process by which our brains identify and differentiate between odors.
Odor has such a strong impact on consumer behavior and perception of quality; however, it is also extremely subjective by its very nature — informed by our own personal experiences and preferences. Smell has a major impact on consumer satisfaction, and manufacturers want to be able to guarantee their customers experience the intended outcome of their product.

As manufacturers continue to test and roll out new products in the industry, digital olfaction can help analyze new formulas and predict which products will resonate best with consumers. Sensory analysis still heavily relies on human panels — which can be expensive and prone to subjectivity. Digital olfaction combines biosensors, advanced optics and machine learning to mimic the human sense of smell which can then be used for objective classification and characterization of flavors.

“The olfactory AI would allow to guide wine lovers to refine their nose. This is a good method for professionals to do without focus groups (note: I am a former market research director) but also sommeliers during blending. With the taste AI, the winemaker wants to build a wine with a certain typicity, the sensory AI helps him. In the same way, it helps him to maintain a certain level of consistency depending on the harvest – I am thinking of our friends in Champagne.” – David BECK.

4. Sensory based AI — Person’s palate is as unique as a fingerprint

Wines are already being matched with consumer preferences as a way to reduce wastage and market specific wines to consumers. With a mix of machine learning, sensory science and analytical chemistry, AI can pinpoint the palate preferences of individual consumers so retailers and distributors can make wine recommendations and see increased sales.

What help can an AI offer to decide to drink a wine?

Julian Perry, CEO at Wine-Searcher

In 2014, Scientists in Denmark led by Ph.D. student Joana Guerreiro developed an “artificial tongue” with almost human taste capabilities, reported Vine Pair. The tongue uses a nanosensor to analyze the molecular makeup of an individual’s mouth. It then processes this data to gauge how tannins will hit their unique tastebuds, and how their palette is likely to respond to the astringency of a given wine.

The technical difficulties lie in the management of humidity and ethanol, unlike the human nose.

Tristan Rousselle Founder & Deputy CEO at Aryballe

Gastrograph AI predicts how people will react to new food products, allowing developers and marketers to use consumer tastes to predict which will perform better or worse on a specific market. This system uses the data of thousands of consumers who have rated thousands of products through a mobile app, specifying different parameters and categories. Gastrograph AI models human sensory perception of flavor, aroma and texture to predict consumer preference of food and beverage products.
Gastrograph AI is used to review beer, coffee, chocolate, and other food and beverage products. Through the AI’s ​​self-learning system, it can determine which flavor and preference patterns work best in each place. In each review, a user can input data for 24 flavor-attributes. Each flavor-attribute is rated on its intensity on a discrete scale from zero to five (zero: not present, to five: most intense; initializing at 0) which constitutes the flavor profile of a sample. Once all of the flavor-attributes’ intensities have been entered, the user inputs a value for perceived quality, which is a discrete value from one to seven (lowest to highest “quality”) that reflect the user’s perception of the quality of the product.

Our clients use Tastry AI as a permanently empaneled focus group of millions of consumers and to provide specific and actionable steps the winemaker can take to best achieve their vision.

Katerina Axelsson, CEO at Tastry

The Tastry technology reportedly individualises recommendations by decoding the flavour (taste, smell and feel) and aroma matrices of wines, by chemically analysing thousands of products in the lab and running them against the unique consumer palate of each individual. AI learns about what people like by breaking down their preferences via question responses that identify each palate. Tastry has also built AI that pairs wines with foods, provides the best matches for groups of people, and addresses other common scenarios.
The algorithm interprets the consumer novel chemistry data and decodes the flavour matrix. The chemistry method will pull over one million data points out of a single bottle of wine. Understanding how that is going to taste, smell and feel to a human palate is the first step. It essentially reduces all the chemistry of the wine — just like the human palate does — to a flavour represented mathematically. That’s when another algorithm looks at that flavour, and the flavour of every other wine or subset of wines.
Then the algorithm chooses a set of deceptively simple quiz questions that can be intuitively answered by almost any person but are able to divide different consumers while simultaneously functioning as analogs to many compounds, or groups of compounds, found in wine. Consumers answer a simple quiz once and an entire retail assortment is ranked and matched to the individual through complex AI systems.
The Tastry quiz, is just the initial understanding of the consumer palate. As the consumer rates wines, Tastry AI is constantly comparing the chemistry of wines as they relate to the ratings on an individual level. Once the insight on an individual level exceeds 80% (the average accuracy of the default quiz), the insight garnered through the initial quiz is permanently supplanted with direct chemistry/palate/rating data which tracks preference changes over time. Consumers at such point see efficacy rates continually increase to well above 90%.

“Of course, depending on the transport, the storage, the light, the temperature of exposure, the wine can have an aromatic palette completely different from the one initially analyzed by a sensory AI. I’m not even talking about the corky taste. Nevertheless, a wine proposed by an expert can turn out to be disappointing for the same reasons. It is about the beauty of a beverage that evolves, that is ‘alive’.
An AI that asks questions is not optimal. We can’t imagine an AI that asks questions like a common market research questionnaire (as a reminder, I am a former director of a research institute). This remains a barrier to mass use.
It will be necessary to combine this taste-learning AI with facial and vocal AI emotion recognition. These technologies help avoid conscious and unconscious bias in consumer responses.
I also find interesting the related effects of a learning AI, to help consumption on health to fight against obesity, the management of allergies, prevention of alcohol abuse…” – David BECK.

The Internet of Things (IoT) is the concept of connecting any device (so long as it has an on/off switch) to the Internet and to other connected devices. The IoT is a giant network of connected things and people.

5. Natural Language Processing (NPL) — AI’s wine rating/review writing?

Reading the wine description below, you can almost feel the cool glass sweating in your hand and taste a burst of citrus on your tongue.

Rating: 4.5 out of 5.

While the nose is a bit closed, the palate of this off-dry Riesling is chock full of juicy white grapefruit and tangerine flavors. It’s not a deeply concentrated wine, but it’s balanced neatly by a strike of lemon-lime acidity that lingers on the finish.

Example of a wine review

But the author of this review never had that experience — because the author was a piece of software. An interdisciplinary group of researchers developed an artificial intelligence algorithm capable of writing reviews for wine and beer that are largely indistinguishable from those penned by a human critic.

AI NLP tools (natural language process) that generate language are nothing new. But the technology appears to have advanced to an unprecedented level. Computer engineer Keith Carlson and his co-authors trained their program on a decade’s worth of professional reviews — about 125,000 total — scraped from the magazine Wine Enthusiast. They also used nearly 143,000 beer reviews from the Web site RateBeer.

Natural Language Processing (NLP) is a subfield of AI. It helps machines process and understand the human language so that they can automatically perform repetitive tasks. Examples include machine translation, summarization, ticket classification, and spell check. A well-know “AI writing assistant” for writer is HyperWrite.

Theoretically, the algorithm could have produced reviews about anything. A couple of key features made beer and wine particularly interesting to the researchers, though. For one thing, “it was just a very unique data set,” says Keith Carlson of Dartmouth College, who co-developed the algorithm used in the study.

Wine and beer reviews also make a great template for AI-generated text, he explains, because their descriptions contain a lot of specific variables, such as growing region, grape or wheat variety, fermentation style and year of production. Also, these reviews tend to rely on a limited vocabulary. “People talk about wine in the same way, using the same set of words,” Carlson says. For example, connoisseurs might routinely toss around adjectives such as “oaky,” “floral” or “dry.”

Machines can write reviews that are indistinguishable from those written by experts.

Keith Carlson, Computer Engineer at Dartmouth College

The algorithm processed these human-written analyses to learn the general structure and style of a review. In order to generate its own reviews, the AI was given a specific wine’s or beer’s details, such as winery or brewery name, style, alcohol percentage and price point. Based on these parameters, the AI found existing reviews for that beverage, pulled out the most frequently used adjectives and used them to write its own description.

Although the algorithm seemed to do well at collecting many reviews and condensing them into a single, cohesive description, it has some significant limitations. For instance, it may not be able to accurately predict the flavor profile of a beverage.

“Combining this NPL technique with sensory AI, you have our expert ready to write reviews, put notes and make personalized recommendations. The AI can even help you develop your palate, according to your desires: at 8pm on a Sunday, on vacation at the other end of the world… so no constraints except being connected to your digital twin, to you. You will become a pro, understand the nuances of a vertical tasting of the domain or chateau you love.” – David BECK.

We can ask ourselves some questions about the effectiveness of its AI models. Besides the fact that they depend on the quality of the data — for example the quality and granularity of the coding of each bottle of wine — how do we know that the matching between the consumer and the product is 100% accurate.

Spectrometers don’t perceive taste the way that people do. Spectrometers measure individual compounds. People perceive compounds in combination with each other, through uniquely individual olfactory systems.

Pam Dillon, Co-founder and CEO at Preferabli

The emotional engagement of consumers is not really measurable, in my opinion, on an administered questionnaire, nor in a more qualitative debrief. This ‘primitive’ technique has too many methodological biases, especially when we talk about personalization, individualization of the satisfaction measurement.

One of the only solutions remains, in my opinion, a facial recognition AI of feelings — emotional AI — coupled with sensory AI. I remain on this line even if an article from The Verge published on June 21st announces for example that Microsoft will withdraw its facial recognition tool that claims to identify emotions… for ethical reasons.