“How much Riesling do I have left?” — what an AI assistant may do at a wine estate
Read-only, always with its source: why an assistant should not touch entries — and still saves time every day.
The most useful question to put to a business system is rarely complicated. It is usually: How much do I have left? Or: What is coming up tomorrow? Or: When did I last treat this parcel?
Those are questions whose answers are in the system. You only have to know which screen.
The real problem is not the answer but the route to it
Every business system can say how much Riesling is in the cellar. The question is how many clicks that costs and whether you still remember the way when you have not taken it for two months.
That is exactly where an assistant is useful: it does not replace the system, it replaces searching in the system. You ask in ordinary sentences and get the figure that would have been there anyway.
The most important limit: read only
An assistant that is allowed to make entries is not a relief but a risk. Not because the technology is unreliable, but because the responsibility cannot be delegated: a cellar entry is a record. Who made it, and on what basis, has to be unambiguous.
That is why read access is the right limit. The assistant answers questions, prepares things and points things out — but it does not make entries, it does not submit declarations, and it sends nothing. Anyone who wants it to do more should first work out who explains at an inspection why an entry looks the way it looks.
Always with its source
The second condition matters just as much: every answer has to say where it comes from. “2,100 litres in the cellar” is an assertion. “2,100 litres in the cellar, source: cellar book, as of today 14:20” is a checkable statement.
The difference is not cosmetic. Language models can produce plausible-sounding figures that are wrong. An assistant that answers from business data and supplies the origin can be checked — you can go and look. One that merely names a figure asks for trust that is not technically justified.
An assistant without a source is a search engine you cannot contradict.
Where it actually saves time
Not on the big analyses — you do those once a year anyway and take your time over them. It saves time on the small questions that come up twenty times a day:
- How much of this wine is left in stock, how much in the cellar?
- When was the last treatment on this parcel, and with which product?
- Which deadlines fall due in the next four weeks?
- What did this customer last order?
- Which item sold worst this year?
Each one of those is answered in two minutes on your own. Twenty times a day is forty minutes.
How to recognise a usable assistant
- It answers from your data, not from general knowledge. “How much Riesling do I have?” is not a question a language model can answer out of itself.
- It names the source and the time it is current as of.
- It says when it does not know. That is the most reliable mark of quality there is.
- It changes nothing.
What it is not
It is not an adviser. It knows what is in your data — not whether the wine will be good. It knows the pre-harvest interval of the product you entered, but it does not decide whether today is the right day to spray. It takes the looking-up off your hands, not the judgement.
That is less than the advertising copy says — and precisely why it is usable. A tool whose limits you know is one you can use day to day. One that promises everything is one you check after every answer, and that saves nothing.