The problem
A forecast nobody calls a forecast
At 4am, somebody in a bakery decides how many croissants, loaves and buns to make. That decision is a forecast. It's just rarely treated as one.
Usually it's based on yesterday, a feel for the day of the week, and a memory of what happened last time it rained. Often it's surprisingly good. But it has blind spots.
Why the current process breaks
- Memory is selective. We remember the day we sold out, not the five days we binned a tray.
- Waste goes unrecorded. Leftovers are donated, discounted or thrown away, but rarely counted per item — so the feedback loop never closes.
- Too many variables. Day of week, school holidays, weather, a local festival, a big pre-order. No one holds all of that in their head at 4am.
- It depends on one person. When the head baker is off, the bake list gets worse.
Why it matters
Bakery margins are thin and the product is perishable. Over-production is a direct cost. Under-production is a lost sale — and a regular who finds the shelf empty twice may start going elsewhere.
The AI thought
Give instinct a memory
This is a textbook forecasting problem, and forecasting is one of the most mature, least glamorous uses of machine learning.
A model trained on item-level sales history learns the patterns a baker already senses — Saturdays are bigger, rainy mornings sell more of certain items — and adds the ones that are hard to hold in your head. Crucially, it can express uncertainty: "probably 40 to 55" is more honest and more useful than "47".
What should NOT be automated
The final call. The baker knows things the data doesn't: a new café opened down the road, a regular is ordering for a party, the oven is playing up. The model drafts; the baker decides.
Every bakery already runs a forecast. It's just stored in one person's head.
AI doesn't replace that instinct. It gives it a memory and a sparring partner.
The solution
A bake list drafted overnight
1. Capture the right data
Most modern tills can export sales by item and hour. Pair that with a ten-second end-of-day waste log — a tablet with item names and a number pad is enough.
2. Forecast tomorrow
Each evening, a model estimates demand for each item, using sales history, day of week, holidays and the weather forecast.
3. Draft the plan
The output is a suggested bake list with a confidence indicator. High-confidence items can be accepted at a glance; low-confidence ones are where the baker's judgement matters most.
4. Close the loop
Actual sales and waste feed back into tomorrow's model. Over weeks, it learns this bakery, not bakeries in general.
How it works
From problem to outcome
Problem
Production quantities are set by habit and memory, leading to daily waste on some items and sell-outs on others.
Input
Sales by item and by hour from the till, day of week, local events and holidays, the weather forecast, and pre-orders.
AI / Agent
A forecasting model estimates demand for each item for tomorrow, with a range rather than a single number.
Decision
It drafts a bake list, flagging items where it's least confident so the baker can apply judgement.
Action
The baker reviews, adjusts and approves the list before the morning shift. End-of-day leftovers are logged in seconds.
Outcome
Less waste, fewer early sell-outs, and a baker whose instinct is backed by evidence.
Level 1
Manual
Quantities set by memory and habit; waste rarely recorded.
Level 2
AI-assisted
A suggested bake list each evening; the baker adjusts and approves.
Level 3
AI-automated
Ingredient orders and prep schedules generated from the approved plan.
Level 4
AI-agentic
A planning agent coordinates production, supplier orders and end-of-day markdowns across outlets.
Implement this
0/8 done
Identify the workflow
Write down how tomorrow's quantities are decided today — who decides, when, and based on what.
Identify repetitive tasks
The nightly bake-list decision and the end-of-day waste count are the core repetitive loop.
Define required inputs
Item-level sales from your till (export weekly to start), plus a simple daily waste log. Weather and events can be added later.
Select the appropriate AI capability
Time-series forecasting. A spreadsheet model or a lightweight tool is enough at first — you don't need a data team.
Build the workflow
Each evening, generate a suggested bake list for the next day. Keep it to your top ten items initially.
Add human approval where required
The baker always has the final say. Treat the forecast as a second opinion, not an instruction.
Measure the result
Track daily waste and sell-out times per item for a month before and a month after.
Iterate
Add more items, factor in weather and events, and connect pre-orders as the model earns trust.
Business impact
Where the value lands
- Cost. Less over-production means lower ingredient and energy spend per item sold.
- Revenue. Fewer early sell-outs on bestsellers.
- Decision quality. A consistent plan even when the head baker isn't in.
- Sustainability. Less food waste — which matters to customers as well as margins.
The takeaway
The right question isn't 'can AI bake?' It's 'can AI help us decide what to bake?' — and that one has a practical answer.
Share this thinking
The newsletter
Think better about AI.
One practical AI idea. One real-world problem. Delivered periodically.