Machine Learning: A Practical Guide Before You Start

What machine learning actually does

Machine learning lets a computer find rules from examples instead of you writing every rule by hand. If you can state the rule clearly - "10% off purchases above one million", you do not need machine learning; ordinary code is faster, cheaper, and inspectable.

It earns its place when the rules are too complex or too numerous to write manually: forecasting demand from dozens of factors, recognising buying patterns, scoring credit risk. The point is not sophistication but whether the pattern genuinely exists in your data.

When machine learning is not the answer

With only dozens or hundreds of rows, machine learning will almost certainly lose to an experienced human judgement. If decisions must stand up legally or to audit, a model you cannot explain becomes a liability. If the real problem is a broken process or data that was never recorded properly, machine learning will not fix it and often masks the root cause. And if you just need reports or summaries, that is ordinary data analysis.

The workflow

The sequence looks simple, but the time goes at the front. Preparing data usually consumes far more effort than training the model.

1. 1State the decision you want to improve, not the technology you want to use. "Cut expired stock by 20%" is measurable; "adopt AI" is not.
2. 2Inspect the data first. How much, since when, who recorded it, how often it is wrong. If those answers are unclear, stop here.
3. 3Build the simplest possible baseline, predicting the average, for instance. Any model that cannot beat it is not worth continuing.
4. 4Train simple models before complex ones. Simple often suffices, and is far easier to explain to a manager or client.
5. 5Test on data the model has never seen, splitting by time if the data is sequential. Otherwise your numbers will look good while misleading you.
6. 6Plan monitoring from the start. Models decay as conditions change, and without monitoring that decay stays invisible until it costs money.

Why machine learning projects fail

In our experience failure is rarely about the algorithm not being advanced enough:

• Data collection starts after the project does. If record-keeping is only being fixed now, a model can only be trained next year.
• Accuracy is chased without weighing the cost of errors. A 95% accurate model can be dangerous if the 5% it misses are the costliest cases.
• Nobody owns the model after launch. An unmonitored model drifts quietly until someone notices the damage.
• Outputs never reach the workflow. Predictions that stop at a dashboard change no decisions.