When You Should Not Use Machine Learning

Machine learning is powerful, and that makes it easy to reach for it first. But a model is not always the right tool. Some problems are better solved with explicit rules, heuristics, or simple statistics — approaches that are cheaper, more transparent, and easier to maintain. Recognizing those cases before building a model saves effort and avoids systems that are harder to explain than the rule they replaced.

When rules are better

If the decision logic is fully known and stable, encode it as rules. A fraud rule like "block transactions over $10,000 from new devices in a new country" is explicit, auditable, and easy to update. A model trained to approximate that rule would be less transparent and would add maintenance cost without adding value. Rules also win when the cost of a wrong prediction is high and the logic is safety-critical: you want to be able to explain exactly why a decision was made.

When simple statistics are better

If the question is "what is the average 14-day stream count for pop tracks?" you do not need a model — you need a group-by average. A model would add complexity to answer a question that a single SQL query or pandas groupby answers directly. The scikit-learn guide notes that a baseline that is hard to beat is a signal that a model may not be needed (scikit-learn: supervised learning).

When data is too scarce or too noisy

A model needs enough examples to find a pattern that generalizes. With 50 rows and a noisy target, a model will overfit and its predictions will be unreliable; a simple average or a domain expert's rule will be more trustworthy. If the label is unreliable — for example, a proxy affected by many confounders — the model learns the proxy, not the underlying signal.

A worked example

Suppose a team wants to predict which support tickets will breach their response-time target. They have 80 tickets. A rule based on ticket age and queue length catches 60% of breaches and is easy to explain. A gradient-boosted model on the same 80 tickets might reach 62% on validation but is harder to explain and may not generalize. Here the rule is the better choice: the model's marginal gain does not justify the complexity, and the team can reason about the rule when it fails.

Common mistakes

One mistake is using a model where a rule already exists and works. Another is using a model to answer a question that simple statistics answer directly. A third is building a model on scarce, noisy data and trusting its predictions without accounting for their uncertainty.

An exercise

Take a problem you assumed needed a model. Write down the simplest rule or statistic that could address it. Measure that baseline. Only if a model clearly beats it — by enough to justify the complexity — proceed with the model. For the broader learning path, see the Applied AI program previews or the data analysis program.