The problem
Cakes, scones and traybakes binned at closing cost twice: ingredients and the baker's time. If a café throws away 8 items a day at about $1.46£1.10€1.31A$2.11C$2.07 each to make, over 6 days that's about $70£53€63A$102C$100 a week, about $302£227€270A$436C$427 a month. Most owners know it happens; fewer know which items on which days.
What the AI does
Export 8 weeks of item sales by day from your till and write down your usual daily bake for each item. Upload both to ChatGPT and ask it to work out, for each item and weekday, the number baked, the most sold and the average left over, with a chart. Its data analysis runs the sums in Python, so you can ask to see the method. You then set a bake number for each weekday. Full data analysis is on Go (about $9£7€8A$13C$13 a month) and above.
A worked example
A café in Shrewsbury bakes the same 9 lines most days and bins what's left at 4pm. The owner exports 8 weeks of item sales by day from the till and uploads it with the usual bake list.
Before
About 8 items binned a day, about $302£227€270A$436C$427 a month
After
About 5 items binned a day after setting bake numbers by weekday, about $189£142€169A$273C$267 a month
What it's worth
About $113£85€101A$163C$160 a month less waste, against about $9£7€8A$13C$13 for ChatGPT Go. That is 3 fewer items binned a day at about $1.46£1.10€1.31A$2.11C$2.07 each to make, over 6 days a week ($26.33£19.80€23.56A$38.02C$37.22 a week, times 4.3), for about an hour a month exporting and checking.
How to set it up
- Export item sales by day for the last 8 weeks from your till's reports, as a CSV or Excel file. Remove any card or customer details.
- Write your usual daily bake in a simple sheet: item, weekday, number baked.
- In ChatGPT Go (€8, about $9£7€8A$13C$13 a month) or above, upload both and ask: 'For each item and weekday, show the number baked, the most sold and the average left over, and suggest a bake number that would have sold out on no more than one day in four.'
- Ask for a chart of leftovers by weekday and check two or three figures against the till.
- Pin the new bake numbers up in the kitchen and run the same check again after a month.
What goes wrong
- Till data shows what sold, not what people wanted. If an item sold out at 11am, demand was higher than the till shows, so note sell-out times for a fortnight first.
- Weather, school holidays and local events move sales. Treat the numbers as a starting point, not a rule.
- Free data analysis is limited and may not get through 8 weeks of data. On personal plans, turn off 'Improve the model for everyone' in Data Controls if you'd rather your sales figures weren't used for training.