Machine Learning, Explained Without the Maths

You already do this in your head

Anyone who has run a business for a few years makes predictions all day without calling them that. The owner of a builders' merchant looks at an order and knows it will probably be paid late. A restaurant manager glances at the weather and the football fixtures and orders more chips. Nobody taught them a formula. They saw enough cases that the pattern stuck.

Machine learning is the same process done by software, on far more cases than a person can hold in memory, and applied consistently at three in the morning. That is genuinely all it is. The maths exists to make the pattern-finding precise, but you do not need it to understand what the thing is doing or whether it will help you.

Ordinary software versus software that learns

Most software you use follows rules somebody typed in. If the invoice is over 5,000 pounds, send it for approval. If the customer is in Scotland, add this delivery charge. Those rules are exact, predictable and easy to check. They are also only as good as the person who thought of them.

Machine learning flips the order. Instead of writing the rules, you collect examples where the outcome is already known, and the software works out rules that explain them. Here is the difference in practice:

If you can write the rule down in a sentence, you usually should. We have a longer piece on when rules beat machine learning, and the honest answer is: more often than vendors admit.

A worked example with old invoices

Picture a wholesale supplier with eight years of invoices, around 60,000 of them. For each one they know the customer, the amount, the payment terms, the time of year, how many previous invoices that customer paid late, and, crucially, whether this invoice was paid on time.

Feed those examples to a learning algorithm and it starts noticing things. Customers who were late twice in the last six months are late again quite often. Large invoices raised in the last week of December drift. New customers on 60-day terms are riskier than new customers on 30-day terms. Some of these the credit controller already knew. Some she did not, and some she half-knew but could never quantify.

The output is a model: a compact summary of those patterns that can take a brand new invoice and say, roughly, this one has a high chance of being paid late. The credit team then chases those first. Nothing magic has happened. The software has read eight years of history faster and more evenly than a person could.

The three ingredients every project needs

1. Examples with known answers. Past cases where you recorded what actually happened. No outcomes, no learning, at least not of the predictive kind.
2. Useful information about each example. The details that plausibly influence the outcome, recorded at the time a decision would have been made.
3. A decision that changes because of the prediction. Chasing a customer earlier, ordering more stock, routing a ticket. A prediction nobody acts on is a report.

The third one is where most failed projects actually failed. We have seen perfectly good models built for questions nobody was going to do anything differently about.

What machine learning cannot do

It cannot predict things that have no pattern in your history. If the reason an order gets cancelled is a phone call you never logged, the model cannot see it. It also cannot cope gracefully with a world that has changed: a model trained before a price rise, a new competitor or a pandemic will keep applying old patterns until someone retrains it.

• It does not understand cause and effect, only association
• It will be wrong a known fraction of the time, forever
• It cannot explain itself the way a person can, although some models are far easier to inspect than others
• It will copy biases that exist in the historical decisions it learned from
• It does not know when it is out of its depth unless someone builds that check in

A model is a very well-read intern with no common sense. Brilliant at spotting what usually happens, hopeless at noticing that today is unusual.

How this relates to AI and chatbots

The large language models behind modern chatbots are machine learning too, trained on an enormous amount of text to predict the next word. The difference for a business is mostly practical. A language model arrives pre-trained and you steer it with instructions and your documents. A predictive model for your invoices has to be trained on your own data, because nobody else has your customers.

Both are useful and they solve different problems. If you want the wider picture, our guide on how machine learning relates to AI untangles the terms.

What it costs and how to start sensibly

A first useful model for a mid-sized business is usually a few weeks of work, not months, provided the data exists and is reasonably tidy. Most of that time is spent understanding and cleaning the data, not on the learning part. The expensive surprises come later: plugging predictions into the systems people actually use, and keeping the model accurate as the business changes.

When SpiderHunts takes on a machine learning project, the first thing we do is try to beat a simple baseline, such as 'customers who were late last time will be late this time'. If the model cannot beat that comfortably, we say so and you keep your money. It sounds obvious. It saves a lot of people from buying something they did not need.

Frequently asked questions