The Prediction Problems Hiding In Your Support Inbox
Most teams are already running three or four machine learning problems on human judgment alone. Someone reads each incoming ticket and decides which team owns it. Someone scans the new leads and guesses which ones deserve a call today. Someone eyeballs yesterday's numbers and decides whether the dip is real. None of that gets called machine learning, so none of it ever gets a model.
What Counts As A Prediction Problem
The test is simple. If a person looks at a row of information and puts it in one of a few buckets, that is a classification problem. If a person looks at a row and estimates a number, that is a regression problem. Ticket routing, lead scoring, spam filtering and anomaly alerts are all the same shape underneath, which is why one small system can cover all of them.
The useful consequence is that you already have training data. Every ticket a human routed correctly is a labeled example. Every lead that closed is a label. The examples are sitting in a database, generated as a side effect of the work your team already did.
Why These Never Get Models
Historically there were two options and both were bad. You could pay a cloud prediction service per call, which turns a quiet background task into a line item that grows with your traffic. Or you could ask a data scientist to build a pipeline, which means the work has to be important enough to justify a person's time.
Ticket routing is rarely important enough for either, so it stays manual forever. The cost of the decision is invisible because it is spread across everyone who touches the queue.
What Training Actually Involves Now
The workflow is shorter than most people expect. Pick a model from a catalog, add example rows either in a browser admin or through the API, click train, then start calling the endpoint over HTTPS and reading the answer.
The modeling is not the hard part anymore. Choosing the right buckets is. If your categories overlap, no model will separate them cleanly, and that is a definition problem rather than a machine learning problem.
Why Self Hosting Changes The Cost Math
A prediction that costs nothing per call gets used in places you would never justify at a fraction of a cent each. You can score every inbound message instead of a sample. You can re run predictions across your whole history when a model improves. You can leave a monitor running on a low volume queue that only matters twice a month.
Your data also stays on your own server the whole time, which removes the security review that usually kills these projects before they start.
The Takeaway
Stop looking for a machine learning project and start looking for the decisions your team already makes by hand hundreds of times a week. Those are the models worth training, and the training data already exists as a byproduct of the manual work.