When Is Machine Learning Actually Worth It for a Business?
Machine learning is not always the answer, and the wrong answer here is the most expensive mistake available. A simple test before you commission a model.
The test that matters before anything else
Machine learning earns its place when you have years of records of a decision being made, the outcome being known afterwards, and the decision happening often enough that a few percentage points of improvement is real money — stock reordering, churn, demand by branch by day, which invoices go unpaid. It earns nothing when the pattern changes faster than data arrives, or when only a handful of examples exist.
Problems worth modelling
- Forecasting — demand, sales, stock depletion, staffing load
- Classification — routing tickets, categorising transactions, triaging leads
- Scoring — expected value, expected duration, risk and priority
- Recommendation — what to reorder, what a returning customer likely wants
- Anomaly detection — suspicious transactions, unusual shrinkage
How the work is actually run
In phases, with a decision point between them, because model development is iterative. The current method is scored first — a model that cannot beat the rule of thumb it would replace is not worth deploying, and that is checked before deployment is funded, not after.
What is included, and what is not
Included: the data pipeline, the model, the evaluation, the serving API, the integration and documentation. Not included: guaranteed accuracy, third-party compute costs, and ongoing retraining unless it is contracted separately. Models age — pretending otherwise is how a model ends up quietly making worse decisions than the spreadsheet it replaced.