While generative AI takes the headlines, an older branch of machine learning keeps taking the profits. Forecasting, ranking, recommendation, anomaly detection, classification — the workhorse models trained on your own data — are quietly responsible for more revenue than every chatbot combined. They are less demo-friendly and more bankable, which is exactly the trade a business should want.
The pattern behind all of them is the same: wherever an organisation makes the same judgement repeatedly from data — how much stock, which lead, what price, is this fraud — a model can make that judgement more consistently, at scale, and without fatigue.
The workhorses in the field
Demand forecasting turns inventory from guesswork into optimisation — retailers run leaner stock with fewer stockouts. Recommendation systems drive double-digit shares of revenue at every serious commerce and content company. Anomaly detection watches transactions, machines and networks for the pattern that precedes the loss. Churn and lead scoring point human attention where it changes outcomes. Predictive maintenance replaces parts before they fail instead of after.
None of these makes a splashy demo. All of them move a line on a P&L, which is why they survive budget reviews that trim flashier projects.
Why ML projects actually fail
The failures are rarely mathematical. Models fail from data problems — inconsistent, biased or simply absent history; from deployment problems — a model that lives in a notebook instead of a workflow; and from missing feedback loops — no one tracking whether predictions stayed accurate as the world drifted. The remedy is treating ML as engineering: pipelines, monitoring, retraining, ownership.
Our own rule saves clients the most money: start with the baseline. A simple heuristic, honestly measured, is the yardstick any model must beat by enough margin to earn its complexity. Sometimes the heuristic wins — and hearing that early is worth more than a model in production that should not be.
Where to start
The best first ML project chooses itself: a repeated, costly decision; historical data that records both the decision and the outcome; and a workflow where a prediction can actually change what someone does. Score your candidates on those three axes and the roadmap writes itself.
Machine learning that pays rent is not a moonshot programme. It is a series of unglamorous models, each attached to a number a CFO already cares about. That is the version we build.
