FeelEat puts smart fridges inside companies. For every one of those fridges, someone has to decide what goes in: which dishes, and how many of each. Get it wrong in one direction and the dish people wanted is gone. Get it wrong in the other and food goes unsold.
We built the AI demand forecasting model behind FeelEat's Planning module. It works out which dishes, and how many of each, should go into each company's fridge. It is 98% accurate, measured against actual sales. Stockouts are down 58% and excess stock is down 41%.
I worry about the model's accuracy more than anything else in this project, and that accuracy depends on the data going into it.
What goes into an AI demand forecast
For each company it serves, the forecast looks at:
- that company's buying pattern and buying history
- the age and gender mix of the people who work there
- the kind of work they do
- the weather
- the product type: main, dessert or drink
- the diet type: vegan, vegetarian, fish, chicken (the list is longer than that)
- the price
Look at that list again. Every item on it is data that someone has to collect, keep correct and deliver on time.
In any project like this, picture a staff mix that changes, or a dish filed under the wrong diet type. The model won't complain; it will carry on producing plans that look exactly as confident as yesterday's, and a forecast that is wrong without anyone noticing is worse than no forecast, because people trust it.
That is why I call the model the easy part. The inputs are where the daily work is.
Forecasting model accuracy: measured against what?
Measure forecasting model accuracy against what actually happened. Our model for FeelEat's smart fridges is 98% accurate compared with actual sales, and that is the figure we quote. Then track the business results next to it: here, 58% fewer stockouts and 41% less excess stock.
A score on old data tells you how well a model fits the past. A comparison with actual sales tells you how it does in the real world, and that number can go down. Everyone who reads it, us included, sees the same thing.
If a vendor shows you an accuracy number, ask one question: accurate against what?
The output is a plan
What the model produces is a plan: these dishes, this many of each, for each company's fridge. Whoever acts on a forecast needs a quantity they can use.
From there, the hard design questions are about people. What does the plan look like on screen? And what happens when someone knows something the data doesn't, like a team that is away for the week?
Stockouts and waste: the numbers that matter
98% is the headline. The two numbers I watch more closely are these:
- 58% fewer stockouts. Fewer people opening a fridge and finding their dish is gone.
- 41% less excess stock. Less food stocked that nobody buys.
A forecasting project that improves accuracy and never measures these is a science project.
Before you start one
What I'd tell anyone planning a forecasting project:
- List your inputs before you choose a model. For us that list ran from buying history to price. For each input, write down who owns it and how you will notice when it is wrong.
- How will you measure accuracy? Decide before anything is built, and measure against what actually happened.
- Design the output for the person who acts on it. For us that is how many of each dish go into each company's fridge. Work backwards from there.
- Track the business numbers from day one: stockouts, waste, or whatever your version of them is.
- Keep a person in the loop. Someone who knows the business should be able to look at the plan and change it when they know better.
One more: if your sales history is thin or messy, don't start with a forecasting model. Start by recording sales properly.
That is the work we do in AI and machine learning projects, and you can see the model in the forecasting model case study. On the people side, agents earn autonomy the way new hires do. If you have a pilot that looks good on paper but nobody uses, tell us where it's stuck.
