MenuStart a projectStart

AI & data

Café prep forecasting from Square sales data

Prep forecasting uses your own Square sales history to suggest how much of each item to make tomorrow. It takes account of the day of the week, the season, the weather and local events, and the kitchen adjusts the suggestions before anything is made.

An example of how the studio would build it. No live services are connected.

From today’s till data to tomorrow’s suggested prep list.

  1. Step 01

    Day closed

    After closing, each item’s sales for the day are read from Square, including the time it last sold.

  2. Step 02

    History stored

    Sales are saved by item, day and hour in a PostgreSQL database, alongside the weather, school holidays and local events.

  3. Step 03

    Quantities forecast

    A forecasting model estimates tomorrow’s demand for each item from recent weeks, the same period last year and the weather forecast.

  4. Step 04

    Prep list shared

    Suggested quantities appear in a Google Sheets prep list, with a short note on why any number has changed.

  5. Step 05

    Kitchen adjusts

    The person running the kitchen adjusts the numbers, and what was actually made is recorded so forecasts can be checked against it.

The result

Fewer early sell-outs and less left over at closing.

The forecast only suggests quantities: it never orders stock or changes the menu. An item that sold out early is treated as demand that went unmet, not as the most it could have sold.

Before we connect it.

Where this is useful

Useful for cafés, bakeries and delis that make food fresh each day and either sell out early or throw too much away. A sensible first version forecasts a handful of best-selling items and compares the suggestions with what the kitchen would have made anyway, before anyone relies on them.

What we need to connect

  1. 01Read access to Square orders and the item catalogue, with items named consistently over time.
  2. 02Several months of trading history, ideally including the same season last year, plus a record of what was made each day if you keep one.
  3. 03Your opening hours, closed days, local events and the items you want forecast first.

What needs careful handling

When an item sells out at noon, the sales figure understates real demand, so the time of the last sale is used to adjust for it. New items, menu changes and one-off events have little history to learn from, so their forecasts are marked as low confidence and left to the kitchen’s judgement. Unusual days, such as a road closure or a heatwave, can distort the history, so they can be marked and left out of future forecasts.

Questions people ask.

Is this AI, or just maths?

Mostly well-tested forecasting methods, which are easier to explain and check than a language model. An AI model can add a plain-English note on why a number moved, but the quantities come from your own data.

How accurate will it be?

We can’t promise a level of accuracy, because it depends on how steady your trade is and how much history you have. Running it alongside your current approach for a few weeks shows whether it helps before anyone relies on it.

Stock level dropped → Sent on approval
Automation 50

Shopify low-stock reorder automation

Monday arrives → Posted to Slack
Automation 89

Weekly KPI digest in Slack

All automation examples