The AI that learned to taste - and quietly tells roughly 100 countries which bottle they'll actually love.
Above: the Preferabli mark. The product is invisible; it lives one layer behind the wine list, doing the guessing so you don't have to.
A guest checks into a Marriott in Napa Valley, opens an app, answers a few questions, and is handed a wine list that already knows their palate. Nobody mentions the software. That is the point.
That invisible layer is Preferabli. It is the recommendation engine sitting behind retailers, merchants and hotels - deciding, with uncomfortable accuracy, what you are going to enjoy in the glass before you have tasted it. The company calls itself an AI-driven B2B2C product discovery and recommendation platform. In plain terms: it predicts taste, and rents that prediction to the people selling you a drink.
It runs across roughly 100 countries. It carries 15 patents. It employs around 31 people. Those numbers do not obviously belong together, which is most of what makes Preferabli interesting.
Recommendation engines guess what sells. Preferabli set out to guess what you'll actually like. Those are not the same problem.
- The thesis, in one lineWine is the hardest thing in the world to recommend. There are tens of thousands of producers, vintages that change every year, and a vocabulary - "minerality," "structure," "a hint of barnyard" - that means everything to a sommelier and nothing to everyone else. Walk into a shop with a thousand bottles and the honest reaction is paralysis.
The retail industry's answer was the same answer it gives for everything: recommend what other people bought. Useful for headphones. Useless for a Riesling, because the customer who bought the same bottle as you may have hated it. Sales data tells you what moved off the shelf. It says nothing about whether anyone was happy they drank it.
The customer who bought your exact bottle might have poured it down the sink. Purchase history can't tell the difference between delight and regret.
- Why click data falls apart for sensory productsPreferabli's founders saw the gap clearly: the beverage world had oceans of transaction data and almost no preference data. The two are not interchangeable. Closing that gap is the entire company.
Pam Dillon and Andrew Sussman started the company - then called Wine Ring - around 2012. The trigger was mundane and, in hindsight, prophetic: the smartphone. Software you could carry anywhere, they reasoned, could change how people chose what to drink. Dillon brought 25 years in consumer retail and hospitality tech. Sussman brought the engineering and a stubborn conviction about machine learning that most people, at the time, found premature.
They were early in a way that is easy to underestimate now. Sussman likes to point out that Preferabli's first patent issued in 2014 - three years before the paper that introduced the transformer, the architecture underneath today's entire AI boom. The company spent years building taste-prediction models while the rest of the world was still arguing about whether machine learning was a fad.
Our first patent issued three years before the transformer model was even written about.
- Andrew Sussman, Co-founder & CTOThe bet was not that AI would arrive. The bet was that, when it did, the company with the cleanest preference data would win - and that the data would have to be built, painstakingly, by people who actually knew how things tasted.
Here is the detail that makes Preferabli unusual. Its models were not trained on scraped reviews. They were built by PhDs in physiology and applied mathematics working alongside what the company describes as the largest assembled group of Masters of Wine and Master Sommeliers in the world - people who taste professionally and can describe a wine in terms a machine can learn from.
The result is a platform that profiles an individual palate from a short series of questions, then matches it against a structured sensory understanding of products. Retailers plug it into their email, web and mobile channels. The Wine Society's version is called "My Taste Match." Virgin Wines uses it to predict your favourite bottle. The technology behaves the same everywhere; only the logo on the front changes.
That sameness is the business. Preferabli sells the engine, not the storefront, which is why most people who use it will never see the name. A merchant gets a recommendation layer that improves as customers interact with it; the customer gets a shop that seems unusually good at reading them. The patents - 15 of them, spanning the US, Japan and Australia - exist to keep that engine hard to clone. Anyone can build a quiz. Almost nobody can build the expert-graded sensory dataset sitting behind it.
When the rebrand from Wine Ring landed in 2022, it was not cosmetics. The platform had outgrown wine. Beer and spirits arrived as first-class categories, and the underlying claim got bolder: that a single model of human preference could span anything you taste or smell.
Its training data didn't come from clicks. It came from people who taste for a living and can prove it on an exam.
- On why the data is hard to copyA taste-prediction claim is easy to make and hard to verify. The clearest evidence is the company keeping it: who agreed to put Preferabli between themselves and their customers. The list is not small. Virgin Wines and Marks & Spencer in the UK. Albertsons and Marriott in the US. And, in June 2026, The Wine Society - the member-owned wine community with a history stretching back to 1874 - which is not an institution that adopts technology on a whim.
In January 2025 Preferabli acquired Libation Labs, a Napa wine-tech startup, and folded its Tastefully platform into a broader hospitality push. That is what put an AI concierge in a Napa Valley hotel: not a chatbot bolted on as a gimmick, but a recommendation system tuned to an individual guest's palate, integrated with large language models for fluency without giving up accuracy.
Approximate figures • USD • per public reporting and data providers
A single Series A doing roughly 85% of the lifting. Translation: the patient years were cheap, and the conviction money showed up once the AI thesis stopped sounding eccentric. Figures are approximate.
An institution founded in 1874 just handed its taste matching to an AI. That is either a very good signal or a very good story. Probably both.
- On the Wine Society partnershipStrip away the patents and the partnerships and Preferabli's mission is small and human: help people find things they will genuinely love, and spare them the dread of the wall of bottles. The company frames it as understanding individual sensory preference at scale. The customer experiences it as a shop that finally seems to get them.
Dillon, named by Goldman Sachs among its 100 Most Intriguing Entrepreneurs and a frequent voice on AI and the drinks trade, tends to argue that smarter software is good news for an industry built on subjective taste - precisely because taste is where generic recommendation engines break down. That is a self-interested position. It also happens to be defensible.
There is a quieter ambition underneath it. The drinks industry has long treated personalization as a luxury - the job of a trusted shopkeeper or a sommelier you could afford. Preferabli's wager is that the same attention can be delivered at the scale of a grocery chain or a hotel group, to people who would never describe themselves as connoisseurs. Done well, it makes expertise cheap and discovery easy. Done badly, it is just another recommendation widget. The difference, again, is the data.
Software is getting smarter, and for a business built entirely on what people like, that is a bigger deal than it sounds.
- The Preferabli view of the AI momentTaste is the last thing anyone expected to automate, which is exactly why it is worth watching. If a machine can reliably predict what an individual will enjoy drinking - not what is popular, not what is profitable, but what they will personally like - the same approach reaches well beyond wine. Coffee, fragrance, food, any product where preference is the whole game.
Preferabli has spent more than a decade building the unglamorous part: the data, the patents, the relationships with people who taste for a living. The current AI wave did not create the company. It just made everyone else understand what it had already built.
So return to that Napa hotel. The guest opens the app, answers a few questions, and is handed a list that already knows their palate. A few years ago that was a fantasy, or a parlor trick. Now it is a quiet piece of infrastructure, running in roughly 100 countries, that most of its users will never know the name of. Preferabli is fine with that. The best recommendation, after all, is the one that feels like your own idea.