What machine learning is, and what it is not

A conventional computer program produces its output by following instructions a person wrote. If a returns policy says that an item may be returned within 28 days of delivery unless it is a sale item, in which case 14 days, then a program that implements that policy contains those numbers because somebody typed them. You can read the program and know exactly what it will do. If the policy changes, somebody changes the program.

A machine learning model produces its output differently. Nobody typed the rule. Instead, the model was shown a large number of examples in which the answer was already known, and an optimisation procedure adjusted a set of internal numbers until the model's outputs on those examples came as close as possible to the known answers. What the model contains afterwards is not a policy but a set of adjusted numbers that happen to reproduce the pattern in the examples. You cannot read it the way you can read a policy, and if the world changes, nobody changes the model: you show it new examples and let the numbers move again.

This single difference is the source of almost every property that follows, good and bad. Because the model is derived from examples rather than written, it can capture patterns nobody could articulate, which is why it can beat a spreadsheet at forecasting 40,000 product lines. For the same reason it can only capture patterns that were present in the examples, it will happily reproduce whatever was wrong or unrepresentative about them, and it cannot tell you why it produced a particular answer unless you build additional machinery to ask.

It is worth fixing four terms that get used interchangeably and should not be.

Artificial intelligence is the broad label for systems that perform tasks that would ordinarily require human judgement. It is a field, not a technique, and it includes approaches that involve no learning at all.

Machine learning is the subset of those approaches in which behaviour is derived from data rather than specified in advance. Calderwood's demand forecast and fraud model are machine learning.

Deep learning is a subset of machine learning that uses neural networks with many layers. It is the technique behind image recognition, speech, and the large language models underneath Calderwood's support assistant and its product description generator. It is not a synonym for machine learning and it is not automatically better than the alternatives.

Rule-based automation executes instructions faster and more consistently than a person, but it learns nothing. Robotic process automation that copies order data between Calderwood's three customer systems is automation, not AI, and calling it AI in a board paper is a small dishonesty that costs credibility later when the board asks why the "AI" cannot handle a case it was never told about.

The practical importance of the distinction is that it determines who is accountable for the behaviour and how you would fix it. If the support assistant states a 30-day returns window when the policy says 28, and the assistant is a rules engine, then somebody typed 30 and you correct it in one place in five minutes. If it is a generative model, nobody typed anything. The number came out of a statistical process over text, and correcting it requires you to work out whether the correct policy was ever available to the model, whether it was retrieved, whether it was retrieved and then overridden, or whether the model produced a plausible number from general knowledge of retail because it had nothing better. Those four causes have four different fixes and only one of them involves the model.

The failure mode when this distinction is misunderstood is systematic. Organisations apply machine learning to problems where the rule is known, deterministic and cheap to write, which is how you end up with a model that predicts something you could simply look up. And they apply rules to problems where the pattern is genuinely too complex to articulate, which is how Calderwood ended up forecasting 40,000 lines with last year plus a percentage for as long as it did.