We are regularly approached for a machine learning project and conclude that the client does not need one. This is not modesty. A model carries costs — training data, validation, monitoring for drift, explainability, and a permanent dependency on someone who understands it — and those costs are only worth paying when the problem genuinely requires inference.
Three questions that usually settle it
- Can the rule be written down? If an experienced member of staff can state the decision as a set of conditions, encode the conditions. It will be faster, cheaper and defensible in an audit.
- Is the outcome you want to predict actually recorded? Models learn from labelled history. If nobody has been recording which cases turned out to be fraudulent, there is nothing to learn from yet.
- What happens when it is wrong? If a wrong answer has consequences for a citizen or a patient, you need explainability, and that constrains which techniques are acceptable regardless of accuracy.
The dashboard that answered the question
One retail client came to us for demand forecasting. The underlying question was which lines were running out before replenishment. That is a stock-cover calculation against sales velocity — arithmetic, not inference. A dashboard answered it in a fortnight, was understood by the buying team without explanation, and required no retraining.
Where models do earn their keep
Genuine pattern problems where the rule cannot be articulated: document classification across inconsistent formats, anomaly detection in high-volume transaction streams where normal behaviour differs per account, computer vision for defect detection at line speed, and forecasting where the seasonality is real and the history is long enough to learn it.
If you do build one, build the monitoring first
A model degrades as the world moves away from its training data, and it does so silently. Decide before deployment what you will measure, what threshold triggers retraining, and who is responsible for looking. A model in production without drift monitoring is a decision system that nobody is checking.
The most useful thing a data practice can tell a client is that the expensive option is unnecessary. It costs us a larger project and buys a client who returns.
Sources and further reading
External links, provided for verification. CyberEdge is not responsible for third-party content.