Last summer, I invited some friends over for coffee on the terrace. The app said it would be sunny. We ended up having coffee indoors.
Whatever annoyed me also kept me thinking. Not because I’d missed out on the sunshine – but because I wanted to understand what the app had actually told me.
What does "a 30% chance of rain" actually mean?
We read a percentage and think: either it’s right, or it’s wrong. The app said "sun", so it was wrong.
But that's not what the app said.
A weather model analyses a vast amount of historical weather data and identifies variables that distinguish one day from another: air pressure, temperature, humidity, wind direction and many more. And, of course, the model knows on which of these days it rained.
When it comes to forecasting a new day, the model essentially asks: how similar is this day to the days when it rained?
The similarity determines the probability.
So “a 30% chance of rain” means that, out of 100 days with these weather conditions, it has rained on 30 of them. On the afternoon I went for coffee, I happened to be there on one of those 30 days.
The app wasn't wrong. It made the right guess and I was just unlucky.
Why that makes a difference
It is a natural reaction, and an understandable one, to interpret a forecast as a simple yes or no statement. Numbers suggest precision, and precision suggests certainty.
In practice, this happens quickly: the model says, “This customer has a high probability of churn,” and the first reaction is, “Is that true or not?” It’s an understandable question. But it misses the point of a forecast, because the answer is neither yes nor no, but always: what is the probability, and under what conditions?
A forecast is always a probability. It is never a certainty.
Same principle, different application
Churn prevention is a classic example: the model uses patterns, contract duration, payment history, complaint history and product changes to determine whether a customer resembles those with a history of cancelling their contract. No judgement. Just a probability.
Another example is intent recognition in customer service: a customer writes or calls, and the model assesses in real time whether they wish to cancel their subscription, make a complaint or simply ask a question. It’s just like with the weather: “Does this enquiry resemble the enquiries we recognise as indicating an intention to cancel?” This enables an immediate, appropriate response even before the customer has fully articulated their concern.
Asking better questions
A model that predicts correctly in 80% of cases is excellent for some decisions, but not for others. This does not depend on the model itself, but on the question being asked and the consequences of an error.
Understanding probability might sound like a statistics lecture, but it is the most practical foundation for making sound decisions based on AI.
What my rainy afternoon spent drinking coffee made clear to me: a good forecast doesn’t mean the result will always come to pass. It means that the model has made the best possible use of the available data.
If you keep that in mind, you’ll automatically ask better questions and make better decisions.