Skip to content

Recommending wine with food

The cashier-side suggestion AI and the customer-facing menu AI both make wine recommendations. This page explains the logic they follow, so your staff can recognise a good pairing call (and override one when the AI is wrong for your house).

This is also a training reference: a server reading this page understands why the AI suggests what it suggests, and can defend or replace the pick at the table.

What it does

A good pairing isn’t magic — it’s four practical levers. When the AI matches wine to food, it weighs these in order:

  1. Cut or echo the dominant fat. Fatty dish → wine with acidity or tannin to cut. Lean dish → wine with body that doesn’t overpower.
  2. Match the dominant flavor, not the protein. Duck with cherry glaze pairs with the cherry, not the duck. Tuna with soy pairs with the soy, not the tuna.
  3. Respect the regional pair. Sangiovese with tomato. Riesling with pork. Champagne with anything fried. These work for reasons centuries old — don’t be cute.
  4. Match the moment. Summer lunch → lighter, brighter, lower-alcohol. Winter braise → heavier, riper, higher-tannin.

How to use it

The default direction, by dish

DishLeverWine direction
Tomato-based pastaEcho aciditySangiovese, Barbera, Frappato
Cream-based pastaCut fatVerdicchio, unoaked Chardonnay, Franciacorta
Margherita pizzaEcho + cutSangiovese, Aglianico, Negroamaro
Grilled red meatTannin matches charNebbiolo, Cabernet, Syrah
Roast poultry with sauceMatch the sauceCherry sauce → Pinot Nero. Lemon → Vermentino. Mushroom → Barbera.
Fish, simply preparedLight body, high acidVermentino, Gavi, Soave
Fish, with sauceMatch the sauce intensityButter → Chardonnay. Herbs → Sauvignon Blanc.
Mature cheeseEcho or contrastEcho: aged red. Contrast: late-harvest white.
DessertThe wine must be sweeter than the dishMoscato, Passito, fortified

When the AI gets it wrong for your house

The AI follows generic principles. Your house may not. Override by writing a Voice entry in the venue knowledge section that captures your style:

  • “We never recommend Bordeaux — natural wine only.”
  • “Our house pairing for the carbonara is the Falanghina, not the Frascati. Use that as default.”
  • “We pour Lambrusco with everything fried, including the calamari.”

Once written, the AI uses your override every time. Don’t fight the AI’s defaults with every guest — write the rule once.

When to override at the table

Guest tells you they only drink white. Pair within the constraint. The four levers still apply — the colour just changes.

Guest tells you they want something “interesting.” That’s a signal to step away from the safe default. Reach for the less-obvious wine on the list, with a one-sentence reason: “This is a Frappato — same family as Sangiovese but lighter, much rarer. Goes with the tomato in a different way.”

Guest tells you they want the best wine. Stop reading the food. Pour the wine they want; the food will be fine.

What happens behind the scenes

When the AI generates a pairing:

  • It reads the cart (what’s currently ordered) and the dish in question.
  • It looks up the structured facts about the dish in the menu knowledge graph — what’s the protein, the dominant flavor, the region, the technique.
  • It scans the wines currently available at your venue and ranks them against the four levers.
  • It then reads any voice or signature entries in your venue knowledge that mention “wine,” “pairing,” or the specific dish, and overrides if you’ve stated a house preference.
  • It generates one recommendation with a one-sentence reason.

It does not recommend wines that aren’t on tonight’s menu, and it does not quote critic scores or invent provenance.

Examples

  • 🍕 Standard — Customer orders pizza Margherita and asks for a glass. AI: “The house Aglianico — its tannin lifts the buffalo milk, and the volcanic minerality matches the basil.” Reason follows the levers; specific wine is from your active list.
  • 🍅 House override — You’ve written a voice entry: “Falanghina with carbonara.” Customer orders carbonara. AI: “Falanghina, our house pairing — the apple note refreshes the egg yolk.” It used your wording, not its own.
  • ⚠️ Edge case — no good pair on the list — Customer orders something the venue doesn’t pair well (e.g. a curry-leaning special, on an Italian list). The AI surfaces the closest option and says it’s an unusual pair: “The Gewürztraminer is the closest match we have — it can handle the spice.” Honest beats clever.