I just designed The French Laundry’s Wine Pairing Menu with Notebook AI
I'm not even a Somm...just your friendly neighborhood winemaker
This post is inspired by a discussion series on Substack: Wine Conversations: Can AI replace Somms (1)? from the team at Tannic Panic and referred by Dave Baxter
This question isn’t theoretical. Consider the following: A restaurant with a well-curated wine cellar loses their sommelier of 5 years, the same sommelier who designed the wine list. Could the Front of House (FOH) create wine pairings for the multi-course Michelin starred tasting menu with the assistance of AI?
AI could handle this task now, if given proper inputs: a detailed wine list, menu ingredients, the final menu, pricing constraints, and well-crafted queries. I decided to test my hypothesis will a couple real-world examples.
Using Notebook AI, I gathered data from two real-world Michelin starred restaurants, The French Laundry *** (TFL) in Yountville and Acquerello ** in San Francisco. Both sets of data are available to download from their websites.
Here’s the data set:
Current Tasting Menu, with supplements
Current Wine list, complete with By-the-Glass (BTG) selections
TFL has 1,274 By the Bottle selections, Acquerello as 1,937
The process took me about 10 minutes total, which included developing a very simple prompt for Notebook to work over the data I supplied. One note1, I put a pricing constraint to select wines either from the BTG list or less than $150/btl at TFL and less than $125/btl at Acquerello.
Getting back to the premise, Can AI replace Somms?
Insofar as guiding someone’s choices, on the spot, in a restaurant with a tasting menu where the guests want to deviate from the pre-selected wine pairing? The answer is NO. Magic can happen between the guests at the table and the Somm. Perhaps there are dietary concerns, or guests don’t want white wine, or “big reds give me heartburn”, who knows, but those on-the-fly decisions are best left to a human (at this point). Just as digital/iPad wine lists have not replaced sommeliers, AI will not replace human interaction on the floor of a Michelin restaurant. Are guests going to start doing prompt engineering at the table? Perish the thought (but there probably are people doing it at the table today).
Should a Somm at a Michelin restaurant with over 2000 wine choices on their menu use AI to help develop pairing suggestions (I didn’t say completely do it)? Yes, they should. The human Somm can then apply their expertise and experience to refine the AI draft list, curate the list to the specific tasting menu at hand or the BOH substitutions that come in last minute, sea bass didn’t show, we’re going with halibut kind of thing.
In the following section, I’ll detail the specific prompts I used with NotebookAI and give you the results of each pairing menu, designed by Notebook and Google’s AI Gemini. This is a thought experiment, nothing more. These AI tools are coming in hot and we in the food and beverage community are adapting, using, creating, and dealing with changes.
If you’re a working sommelier, I would like to hear your feedback and comments, on this rather simple process of AI prompting and the results the AI produced. I’d also love to know how long it takes to develop a pairing menu like the ones I’m highlighting. These took me 10 minutes. Do you taste the entire menu the chef has created? What’s your current inventory of wines? So many questions…
I have included the edited versions of the pairing (1 page) and unredacted (4 page) versions. There are two differences, the unredacted versions give the AI narrative rationale for each pairing suggestion (instant FOH notes basically) and the format for each is slightly different, despite using the same prompts for both, other than the dollar limitation.
Prompt for Acquerello: I would like to design a wine tasting menu to pair with the Spring menu, using the attached wine list. I'd like to limit the potential list of wines to the by the glass list and wines on the bottle list to $125 ($150) or less.
Super interesting! I do think the main selling point is to produce rapid drafts, to reduce work load and speed up the process. Yet I can’t help but think that this initial screening, which is bound to be predicated on previous lists and more generic notions of what goes together can lead to the actual somm overlooking better combinations or option.
Great stuff here, David - I shared your article with the rest of the official Wine Conversation crew for this month so we can treat this as an unoffical Entry 3.5 in the convo, and when we collect it at the end of the month we'll be sure to keep it included then, too. I try to limit the total number of official participants so it isn't too unweildy to organize and execute, but I really like the idea that these conversations can inspire unofficial side entries, we have another one of these likely to post near the end, too! And that's the joy of Substack and the reason I started this exercise in the first place - we're all here, and we're all independent, we have to combine forces now and again! Crossovers aren't just the purview of Marvel... ;P
Oh, also, going forward, in case you didn't know, if you want to link to another writer or article, the best way to do that is to use the "@" symbol just like in social media then the name, tag them that way which sends us a notification that you've done so. For an article, just copy/paste the URL direct into the text of the post, and it auto-embeds it and again notifies us. I went 3 days not realizing you'd posted this!