What Machine Learning Can — and Cannot — Do for Your Business

There is a version of the machine learning conversation that goes like this: ML can do anything. Feed it enough data, give it enough compute, and it will solve your hardest problems. This version is wrong — and believing it leads to expensive disappointments.

There is another version that goes like this: ML is overhyped, the results never match the promises, and you are better off with a good analyst and a spreadsheet. This version is also wrong — and believing it means leaving real value on the table.

The truth is more nuanced, and more useful.

What Machine Learning Is Actually Good At

Machine learning excels at a specific class of problems: finding patterns in large datasets that would be too complex or too time-consuming for humans to find manually.

Prediction from historical data. If you have historical data about an outcome — customer churn, equipment failure, fraud, demand — and that outcome is influenced by patterns in the data, ML can build a model that predicts future occurrences with meaningful accuracy. This is the core use case, and it is genuinely powerful.

Classification at scale. Sorting large volumes of inputs into categories — documents, images, transactions, support tickets — is something ML does faster and more consistently than humans. The quality depends heavily on the quality of the training data, but at scale, even imperfect classification can be enormously valuable.

Anomaly detection. Identifying what is unusual in a stream of data — unusual transactions, unusual sensor readings, unusual user behaviour — is a natural fit for ML. Humans are good at spotting anomalies they have seen before; ML can spot anomalies it has never seen.

Personalisation and recommendation. Matching content, products, or experiences to individual users based on their behaviour is a problem ML handles well, at a scale that would be impossible manually.

Where Machine Learning Struggles

Understanding the limits of ML is as important as understanding its strengths.

Problems without historical data. ML learns from examples. If you do not have historical data about the outcome you want to predict, you cannot train a model. This is a fundamental constraint, not a temporary one.

Problems that require causal reasoning. ML finds correlations. It does not understand causation. A model that predicts which customers will churn cannot tell you why they churn or what will make them stay. For that, you need human judgment, experimentation, and domain expertise.

Problems where the world changes rapidly. ML models are trained on historical data. When the world changes — new products, new regulations, new competitive dynamics — models trained on old data can become unreliable quickly. Maintaining ML systems requires ongoing monitoring and retraining.

Problems that require explanation. Many ML models, particularly deep learning models, are difficult to interpret. In regulated industries, or in contexts where decisions need to be explained to customers or regulators, this is a significant constraint.

Problems where errors are catastrophic. ML models make mistakes. In most contexts, this is acceptable — the model is right often enough to be valuable. In contexts where a single error has severe consequences, the risk calculus changes significantly.

The Questions Worth Asking

Before investing in a machine learning solution, ask these questions:

Do we have the data? Not just data in general, but the right data — labelled, clean, and representative of the problem you are trying to solve.

Is the problem stable? Will the patterns in historical data still be relevant in the future? Or is the environment changing in ways that will make historical data less useful?

What happens when the model is wrong? Every model makes mistakes. What is the cost of a false positive? A false negative? How will errors be caught and corrected?

Can we act on the output? A prediction is only valuable if it changes a decision. Who will use the model's output, and how will it change what they do?

Is ML the right tool? Sometimes a simpler approach — a rules-based system, a statistical model, a well-designed process — is more appropriate than ML. The goal is to solve the problem, not to use the most sophisticated technology available.

Getting the Most from Machine Learning

The organisations that get the most value from ML are not necessarily the ones with the most sophisticated models. They are the ones that:

• Start with well-defined problems where the value of a solution is clear
• Invest in data quality before investing in model complexity
• Build feedback loops that allow models to improve over time
• Treat ML as a tool that augments human judgment, not replaces it
• Measure success by business outcomes, not model metrics

Machine learning is a powerful tool. Like any tool, its value depends on using it for the right problems, in the right way, with the right foundations in place.

If you are trying to figure out where ML fits in your organisation, we are happy to help you think it through.