AI That Earns Its Place
The interesting question is rarely whether machine learning can be applied, but whether it should be. A surprising number of problems are better served by a rule, a form, or a fixed process.
There is pressure to add machine learning to things. Some of it is genuine opportunity. A good deal of it is a solution looking for a problem.
The question worth asking first
Not can we model this but what would have to be true for this to be worth doing?
Usually three things: the decision recurs often enough that automating it matters; the pattern is genuinely learnable from data you can obtain; and someone will act differently because of the output.
Drop any one and the project produces something technically defensible that nobody uses.
Simple approaches are undervalued
A rules engine that captures 85% of cases, that a domain expert can read and correct, is frequently better than a model reaching 90% that nobody can explain when it errs.
The gap narrows further once you count what the model actually costs: labelled data, evaluation infrastructure, monitoring, retraining, and the on-call burden of a system that degrades silently. Those costs are real and recurring, and they rarely appear in the initial comparison.
Sometimes the honest recommendation is a better form, a validation rule, or removing a step. That advice is less exciting to give and more useful to receive.
Where it does earn its place
Machine learning is genuinely strong where rules become unmanageable: perception tasks such as identifying defects in inspection imagery, problems with many weakly-informative features, patterns that shift and need periodic relearning, and cases where ranking a large set matters more than being precise about any single item.
The common thread is that the decision boundary is real but too intricate to write down.
Deployment is the start of the work
A model that performs well in evaluation has cleared the easy part. Then it meets data drift, an upstream format change, edge cases nobody thought to label, and users who adapt their behaviour to it.
Before launch, be clear on: what is monitored, what threshold triggers investigation, who owns it, how retraining happens, and how to roll back. Without those, you have not shipped a system — you have shipped a liability with good early metrics.
The standard
Every added component must justify its maintenance cost for as long as it runs. Applied honestly, that standard rules out a lot of proposals — and makes the surviving ones substantially more likely to still be delivering value in two years.