Why AI cannot answer a question about your restaurant

«Why did margin drop in August?» is the kind of question a restaurant usually answers a week later and approximately. Not because nobody can do the arithmetic. Because the answer is assembled from four sources, and none of them knows about the other three.

Where the data lives

• Sales - in the till and the books: what was sold, at what price, on which day.

• Purchasing - in the invoices: what arrived, at what price, from whom.

• Cost - in recipe cards: how much of each ingredient goes into a dish.

• Shifts - in the rota: who worked and for how long.

• Documents - in the mailbox: supplier emails, e-invoices, photographs.

Each source is right in its own way. The trouble starts where the question crosses a boundary: to answer anything about margin you have to add up sales, purchasing and recipe cards - and be sure all three are talking about the same item.

Three reasons the answer does not assemble

1. One item is named three ways. «Cream 33% 1 l», «Slivki 33 1l», «Cream Valmiera» - to a system these are three items, and the price history breaks into three fragments.

2. Units do not match. The supplier sells a case, the stock is kept in kilos, the recipe in grams. Without the pack multiplier the figures are not comparable.

3. The price depends on which price you asked for. The weighted average in the books and the last purchase price are different numbers, and the gap can be several times over.

What has to exist before an assistant can answer

Between the restaurant's systems and the model there has to be a layer that brings everything to one shape: one item, one record, one unit, one price history. On that layer «why did margin drop» stops being an investigation and becomes a query.

• items are merged and matched to the accounting catalogue;

• units and packs are converted to a base unit - kilo, litre, piece;

• the price history is continuous and tied to a supplier;

• sales, write-offs and stock are read for the same period as purchasing.

Only then does the assistant stop answering in generalities. How it works for us is on the assistant page; the list of connected systems is under integrations.

A five-minute test

Take one item you buy constantly and try to answer three questions: what was the price per kilo in June, what is it now, and how many portions of that item you sold in the same month. If it takes more than five minutes and two exports, the data has not been brought together.