Can Ai Be Your Sommelier of the Future? By Joanna Dabrowska

When I was leaving AWS at the beginning of December 2022, Open AI was just starting on their global conquest. I am not sure do we all remember Will Smith eating spaghetti video, but this was a next step after our first Chat GPT (or other LLMs) musings: AI image and video generation. I cannot remember my first prompt in Chat GPT ever, but I remember I have used it initially very heavily to write “angry” corporate emails in the most polite British manner. I also used it, and I continue to use it now as well, to work on my copy, edit all my typos, and check if I have made any grammar mistakes. Although I am native, I would always rather double check my grammar (old school upbringing here, Granny would be mortified otherwise).

Three years in, and the AI generated videos of almost every scene, humans or animals, are slowly flooding social media. Some are genius, my favourites are the little movie productions. It is a fantastic outlet for amateur movie production, and in many cases a good edit can fool a distracted viewer. DeepFakes that were quite easily recognised as fakes a few years ago are now not so easily detected, and believe me both parts of deep fakes — the video generation and the actual “fakeness” detector — are fuelled by AI. We are still far from perfection, as the recent festive billboard at the riverside in London has shown. And if I am honest, in the art playground, I do hope it will stay where it is, as I firmly believe that the technology advancements should focus on the areas of research like health, wellbeing, automatic house chores and how to make our pets live longer, rather than replace our creativity.

Amongst all disruption and uncertainty that AI has brought in every single workspace this year, and against all my best bets, pitches and wishes, wine seems to remain steady on the old course. And whereas I have found many applications where data and AI could truly benefit the wine industry, I did ponder for quite a while whether it is possible, and more importantly, if it is sensible to think that AI will replace our beloved hospitality, and more precisely, sommeliers.

What does a sommelier do?

Although most of us have a rough idea of what a somm does (yap about wine, of course), I think it's essential to dive a little deeper into the actual role and its requirements. One of the recently published job adverts in London states that the sommelier will be responsible for:

• Displaying excellent wine knowledge and being able to recommend and match appropriate wines with chosen food;

• Organisation and maintenance of the wine list and cellar;

• Training and supporting other staff members;

• Ensuring all beverages are correctly posted through the internal SKU system.

To be a successful sommelier, you will need solid wine and food knowledge, a strong understanding of fine dining service, and previous experience in a similar role. In many cases, strong rapport-building skills complemented by clear and confident communication are also essential, particularly in luxury-oriented establishments. This is also, typically, where we encounter them. Beyond their encyclopaedic knowledge, one of the most crucial aspects of the role is the actual wine service. A sommelier is responsible for serving wine at the correct temperature, in the proper glass, and with precise etiquette. Most importantly, they open the bottle. That alone is a skill, especially when dealing with an aged Bordeaux or Burgundy. Nobody wants to drink wine with the cork pushed into the bottle. Trust me, if anything belongs in my version of the Last Christmas lyrics, it is that. Opening such bottles often requires the use of a Durand, or another bit of specialised equipment involving heat and glass. Sabrage qualifies as well, for those who enjoy a touch of drama (me!).

What is AI?

To understand whether AI can replace a sommelier, it is worth explaining what AI actually is and how it works. To understand it properly, we need to go back in time to the 1940s and 50s, to Alan Turing and his machine, and the famous test he designed to determine whether the participant in a conversation is human or artificial. The test itself is simple: if a machine is able to convince a human that it too is human, then the machine has passed. That is when we begin speaking about AGI, which stands for Artificial General Intelligence, a hypothetical form of AI with human-level cognitive abilities, capable of understanding, learning, and applying knowledge across a wide range of tasks.

Ever since then, we have been in pursuit of AGI and along the way we have developed many different types of AI. These include machine learning, computer vision, natural language processing, robotic process automation, and more recently, generative AI. Since the GenAI hype wave began, nearly everything in this space has been labelled AI, which is confusing because AI is nothing new. In fact, we have had it since the 1950s, used extensively in research, development, academic studies, and experimental projects. It did, however, require a particular blend of statistics, mathematics, computing, and programming skills (hello, data science!).

Recent platforms powered by the names such as OpenAI (ChatGPT, Sora), Anthropic (Claude), Google (Gemini), and Microsoft (Copilot) have made large language models (LLMs) and generative tools for images and video accessible to the everyday user. Though it is fair to say that not everyone using them is fully aware of their limitations.

Where do we encounter AI?

Today, AI is quite literally everywhere. If we are speaking about generative AI, it has made its way into all forms of content creation. Emails, blog posts, social media captions, video scripts, images, music, even poetry. If we are talking about machine learning, it powers your email’s spam filter, your energy usage forecast, and the algorithm deciding which adverts you see online. Robotic process automation helps train both digital and physical systems, while computer vision drives tools for image recognition, medical diagnostics, and autonomous navigation.

AI is replacing recruiters. Agentic AI is now being used to automate internal processes within companies. Algorithms are being trained to take over an increasing number of repetitive and routine human tasks. As a data scientist, I understand just how far we still are from AGI. And I also see clearly how inflated the hype around AI has become. At this point, you can hardly avoid it, unless you make the rather charming decision to retreat into the countryside and live off your land somewhere in Montana. I must admit, I think about that from time to time as well.

Can AI replace a sommelier?

Is it possible for AI to take over sommelier roles right now? Let us return to where we are in our current AI development journey and compare it with the responsibilities of a sommelier.

Excellent wine knowledge, personalised recommendations, and food and wine pairing skills are essential to the role. Technically, we can train an algorithm using available datasets on wine facts, styles, and characteristics. We could also add a feature that allows us to match wine data with food data and then generate pairing suggestions based on chemical compounds or other relevant criteria. The output could certainly be used as a recommendation.

AI is already supporting a variety of organisational and maintenance tasks. In the context of wine cellar management, both commercial and private, it is an excellent application. Many existing systems are now being enhanced with analytics and AI powered components that present information in a more visual and user friendly way. Adding simple checks to flag errors, similar to the way a word processor alerts you to a spelling mistake, is not a significant technical challenge. Ensuring that items are entered correctly into an inventory system is well within the capability of AI.

We could also use AI, and here I mean the entire spectrum of artificial intelligence rather than only generative models, to assist in training and supporting other members of the hospitality team. Information can be tailored to different levels of expertise, and digital learning tools can be adapted to various learning styles.

The key ingredient

The challenge in training algorithms, their performance, and the likelihood of a successful prediction is closely tied to the quality of the input. In other words, the quality of the data. The famous phrase “garbage in, garbage out” still explains many of the failures in AI projects today, and often in analytics as well.

Algorithms rely on data collected from databases, which still need humans to input and curate that information. These databases can be incredibly helpful for humans too. They allow us to store knowledge in one place, and whether you are studying or simply updating your portfolio, such systems can be powerful learning tools. But in both cases, data must be accurate and uploaded in a clear, and consistent way that works for both humans and machines.

How does a sommelier collect their data? Naturally, they rely on many of the same tools and systems I have just mentioned. However, the sommelier has one key advantage over machine learning. And that is a simple human conversation. Yes, data can be collected through dialogue and interaction, and yes, conversational AI is very real. But for now, a human still needs to verify whether the data is useful and whether the system is working as intended. That process takes time, sometimes hours or days.

Meanwhile, a skilled somm can uncover a guest’s preferences within just a few minutes of conversation. They can offer the perfect wine for the mood, the moment (hello, Hannah Crosbie), or the meal, often with a side anecdote about the vineyard, the winemaker, or the winery dog..

All of this is achieved through small talk, rapport, and a sharp understanding of a guest’s needs, many of which are never actually spoken out loud. AI can do many of these things. Sometimes it can even beat a sommelier in speed or efficiency. But it has one disadvantage that cannot be ignored. AI does not possess emotional intelligence. It cannot pick up on sarcasm. And it certainly cannot walk across the room to meet your eye as you sit waiting at the table. At least not yet!

What the future holds

I think by now you have enough information to assess it for yourself. Hypothetically, yes, we could develop an AI sommelier. Once we reach the level of AGI, it is entirely possible that we will be interacting with such systems without even realising it.

But trust me, it is still a long way to Tipperary. And that is not necessarily AI's fault. As humans, we only understand a very limited portion of how our own brains actually function. These are incredibly complex mechanisms, and it is that very complexity that AGI aims to replicate.

Now add to that the full range of human communication skills. Ours are the most sophisticated in the animal kingdom, and they rarely rely on verbal conversation alone. Body language, gestures, even a raised eyebrow or a slight pause can carry meaning. AI can be trained to read those cues, to some extent. But can it replicate them in a way that feels natural and intuitive? No. Not yet. And when it tries, it tends to be, well, cringe to say the least.

Then there is something we share with animals, something deeply personal and completely unique to each of us. It is called instinct. Or if you prefer, intuition. We do not understand it fully, and we certainly cannot recreate it. Not in code, not in silicon.

AI, at its core, is designed either to mimic how humans act through machine learning or how humans think through neural networks. Algorithms are essentially instructions, written and executed by people, to achieve a specific goal or make a prediction. LLMs, in particular, are built to work with text. They summarise, extract, tidy up, translate, and perform other editorial tasks. And yet, they still fail at many of these.

If I had to imagine a future for an AI sommelier, it would not be as a replacement. It would be a tool. A well-designed knowledge system that professionals could access to support their work with wine recommendations.

Is it worth it?

From a financial perspective, it would be one of the most expensive ventures imaginable, and sadly, considering current wage levels, it would likely deliver little return on investment, if any at all.

Then there is the question of accuracy. We have not yet fully accounted for just how much personal information AI would need in order to make the best wine recommendation for you. It would need to know who you are, your age, your spending power, your mood, your intention, whether you are a wine geek, whether it is your birthday or a friend’s, whether you are sharing, whether you are dining alone, whether your date failed to arrive. Not everyone wants to offer that kind of detail, nor spend their evening in a restaurant providing it.

A successful sommelier, on the other hand, can read you within seconds. You do not need to reveal anything overtly. An experienced somm can glean what they need through a simple comment about the weather. Through conversation, a great sommelier can gently challenge your assumptions. They can persuade you to try something new, or encourage you to look at a food and wine pairing from a completely different angle. And because it is a real human interaction, the conversation often wanders. A small story might emerge, a recommendation, a detour into a vineyard tale or a travel memory. That influence can shape not only what you drink tonight, but perhaps even where you book your next holiday. Never underestimate a somm’s enthusiasm for recounting their wine region adventures.

And if we zoom out a little further, we face a much larger question. If we replace the sommelier, why stop there? Why not the entire hospitality experience? But then what would remain? Dining out, going out, enjoying a restaurant — these are not just about food and drink. They are about being looked after. Being seen. Having your wine knowledge complimented. Engaging in a brief but meaningful exchange with a stranger. Sharing a moment over a bottle you have never tried before.

Would we really want to return to the days of private dinners in grand houses, with limited options, the same crowd, and no spontaneity? That is the very opposite of what dining out offers. And no matter how much we like our somm, the interaction has a natural limit. We enjoy it, then we return to our lives. It is not so with AI. Can you imagine an AI somm constantly following you, sending wine recommendations at five in the morning because it detected you were awake? Or popping up with a new bottle suggestion during a funeral? A new Black Mirror episode practically writes itself.

We go to restaurants for experience, connection, and conversation. We leave satisfied, perhaps with a few thoughts to take home, and ready for the next item on the agenda. We might return one day and be greeted by the same sommelier, as you would be by an old friend. We might build a rapport, a simple human connection that is, in many ways, the essence of life itself. Can AI truly replace that?