Where it genuinely saves hours
The first and clearest: repeated questions. The same ten enquiries arrive every day, many of them outside hours and in different languages. A well-built assistant answers those and leaves you the conversations worth having.
The second: drafts. Translating the menu into four languages, writing descriptions for thirty products, turning scribbled notes into a presentable quote. What it saves is not the typing; it is the blank page. The review stays yours and is not negotiable.
The third: sorting what arrives in bulk. Classifying email, extracting details from a long form, summarising twenty reviews into what they have in common. Dull, frequent tasks with little downside if one goes wrong.
Where it does not
Not for setting prices or discounts or anything with financial consequences without a person confirming it. Not for publishing text unreviewed, for the same reason you would not publish a machine translation: a small error in a price or a condition costs you a customer or a complaint.
And not for replacing the person who serves. In a small business, personal attention is usually precisely the advantage over the large one. Automating the part that was already mechanical makes sense; automating the conversation that differentiates you is self-defeating.
Grounding, which is the whole difference
A language model, by default, produces a plausible answer. When it does not know something, it will often fill the gap with something that sounds right. In general conversation that is a curiosity; on a restaurant’s website it is a customer standing at a locked door on a Sunday.
The technical answer is called grounding: the assistant can only answer from a set of documents you control — your menu, your prices, your hours, your terms — and is explicitly instructed to say it does not know when the answer is not there.
That changes two practical things. One, keeping the assistant current means keeping the documents current, not retraining anything. Two, when it gets something wrong you can see which document it came from, which turns a mistake into something fixable rather than a mystery.
What the law asks for
Two things, neither complicated. First: if a person is interacting with an AI system, they have to know. The EU AI Act sets transparency obligations precisely for systems designed to talk to people. One line at the start of the chat deals with it.
Second: conversations contain personal data, and the GDPR applies as it does to any other processing. You need a lawful basis, clear information about what is stored and for how long, and to know where it is processed. If your supplier cannot answer that precisely, that is the answer.