LSM Agents
CAVA · Deep dive

Christos, the sommelier who never hides what he is.

An AI sommelier built on Anthropic Claude, across six languages — designed so transparency is the principle, not a footnote.

At /sommelier on cava-griechischerwein.de there is Christos. He is an AI sommelier powered by Anthropic Claude, and he does the one thing a good shop assistant is there for: listen, ask, and recommend a bottle that fits your evening.

Recommending like a good shop assistant

A shelf of 316 products is a barrier for many customers, not a help. Christos turns that around. Instead of having to filter by region or grape, the customer describes what they want — a bold red for lamb, something light for the terrace, a gift for someone who loves Naoussa. Christos answers the way a person behind the counter would: with a specific bottle and a reason for it.

That is the difference between a search bar and a conversation. A search asks the customer to already speak the catalogue’s language. A conversation meets them where they stand — at taste, occasion and food. That is exactly how wine is bought in a real shop, and exactly how it should feel online.

Six languages, the same posture

Christos works across the full six-language interface of CAVA: German, Greek, English, French, Italian and Spanish. Greek wine is a diaspora product — a customer in Paris or Madrid should get the same care as one in Charlottenburg. The language changes; the adviser’s posture does not.

Transparency is the design

Christos never hides that he is an AI. From the first second, the customer knows they are talking to a machine — not a human in disguise. This is not caution born of fear of the rules. It is the more honest way to give advice: you should know where the recommendation comes from.

The path of the data is just as open. Processing happens in the US, and the privacy policy says so plainly — no hidden data flows, no surprises in the fine print. Your inputs are not retained as training data. What you tell Christos is used to answer you, not to feed a model.

Why disclosure now is an advantage

The EU AI Act, through Article 50, brings binding transparency obligations from August 2026: anyone interacting with an AI must be able to tell. Many shops will have to retrofit that later — as a notice that looks bolted on, because it was.

At CAVA, disclosure was part of the design from the start. Christos is built disclosure-first: the customer always knows they are talking to an AI, and where their data is processed. What will be a box-ticking exercise elsewhere is already lived practice here. A requirement taken seriously early becomes proof of trust rather than a compliance burden.

Trust is the real feature

A sommelier who never sleeps is nice. A sommelier you believe sells wine. That is why transparency here is not an accessory but the function itself: a customer only asks Christos for a bottle for a special evening once they know what they are dealing with. That is precisely what makes an open AI better than a hidden one.

All prices exclude 19% VAT.

At a glance

  • Live at /sommelier on cava-griechischerwein.de
  • Powered by Anthropic Claude
  • Six languages: DE, EL, EN, FR, IT, ES
  • US processing disclosed plainly in the privacy policy
  • No training-data retention
  • Disclosure-first, ahead of AI Act Art. 50 (in force August 2026)

— Dimitrios

Back to CAVA
The next build

Want an AI that says openly what it is?

Start at 499 €