Steganox Engineering · 27 May 2026
A model is a maintenance commitment
Shipping a machine learning system is not like shipping a feature. A feature behaves the same on the day you deploy it and eighteen months later. A model degrades quietly as the world drifts away from the data it was trained on, and it does so without throwing an error. You need monitoring, retraining, a labelled evaluation set that stays current, and someone who understands why the numbers moved.
That is a real ongoing cost. It is worth paying when the problem requires it. It is a poor trade when the same outcome could have come from a well-written rule, a lookup table or a better-designed form.
Four questions before you commission a model
Can you write the rule down? If a domain expert can describe the decision in a page of conditions, start there. Rules are inspectable, testable and free to run. You can always add a model later for the cases the rules handle badly, and by then you have a baseline to beat.
Do you have labelled examples? Not data. Labels. Ten million rows of unlabelled logs will not train a classifier. If labelling is going to take three months of expert time, that is part of the project cost and it belongs in the plan.
What is the cost of being wrong? A recommendation engine that suggests a mediocre product costs you a click. A model that declines a loan application costs someone their house and you a regulatory conversation. The higher that cost, the more the system needs explanation, appeal routes and human review, which is often more engineering than the model itself.
Would a person notice if it stopped working? If the answer is no, you do not have monitoring, and an unmonitored model in production is a liability accruing quietly.
Where it does earn its place
Machine learning is worth it when the pattern is real but nobody can articulate it, when the input is unstructured (images, free text, audio), or when the decision has to be made thousands of times a day at a speed and consistency people cannot sustain.
Those are common. We build these systems regularly. But we have also talked several clients out of one, and in every case the resulting project was cheaper, shipped sooner and did the job.