Five Signs Your Business Problem Is a Machine Learning Problem
Last updated:
Most problems are not machine learning problems
That is not a gloomy opening. It is a money-saving one. We talk to a lot of businesses who have been told, or have told themselves, that machine learning is the answer. For a good share of them the honest answer is a report, a rule, a cleaner process or a piece of ordinary software.
The ones that genuinely suit machine learning share a recognisable shape. These five signs describe it. If your problem shows all five, it is very likely worth exploring. If it shows two, save your budget for now.
Sign one: the same decision, made many times
Machine learning pays back through repetition. A small gain on each decision, multiplied by thousands of decisions, becomes real money. A decision made twelve times a year rarely justifies a model, however hard each one is.
- Good fit: approving 800 trade credit applications a month, routing 2,000 support emails a week, pricing 300 insurance quotes a day
- Poor fit: choosing where to open the next branch, deciding annual budgets, selecting a new supplier
The poor-fit examples may well benefit from data and analysis. They just do not have the volume to train or justify a predictive model.
Sign two: the answer depends on many tangled factors
If an experienced person can state the rule in a sentence, write the rule. Machine learning is for decisions where judgement weighs many signals at once, and where different experts would describe their reasoning differently.
Ask the person who makes the decision how they do it. If they say 'it depends' and then list eight things that interact, that is promising. If they say 'anything over 10,000 pounds goes to the director', that is a rule, and a rule will beat a model on cost and clarity every time.
Sign three: you have history with outcomes
You need past cases where you recorded both the details and what happened. A list of customers is not enough; you need to know which of them stopped buying. A list of orders needs to show which were returned.
| What you have | What it supports |
|---|---|
| Several years of cases with outcomes recorded consistently | A predictive model, likely soon |
| Plenty of records but outcomes missing or patchy | Start recording outcomes; consider labelling a sample |
| Outcomes recorded, but only a few dozen of the important kind | Wait, pool cases, or use rules plus human review |
| Little history at all | Rules or a pre-trained tool now, learning later |
Check the unglamorous detail too: were the inputs recorded at the time, or filled in afterwards. The difference decides whether a model trained on them will work live.
Sign four: a better prediction changes an action
This sign kills more projects than any other, and it should kill them early. A prediction only has value if someone does something differently because of it. Otherwise you have built an expensive dashboard.
- Name the person or process that will receive the prediction
- Describe what they will do differently for a high score versus a low one
- Estimate what that different action is worth in money or time
- Check they have the capacity to act on the predictions the model will produce
A churn score that goes to a retention team with a budget for offers passes this test. The same score emailed as a monthly PDF to a busy director does not.
A model nobody acts on is a very expensive way to feel informed.
Sign five: occasional mistakes are acceptable
Every model is wrong some of the time. The problem suits machine learning if a wrong prediction is either cheap, caught by a person, or outweighed by the value of the right ones.
Recommending the wrong product is cheap. A human reviewing a flagged transaction catches false alarms. Automatically refusing a mortgage based solely on a score is neither cheap nor easily caught, and it brings legal duties around automated decisions. That does not rule out machine learning, but it does mean a human-in-the-loop design and considerably more care.
Warning signs, and what to do next
Some patterns suggest the problem sits elsewhere:
- The data needed lives in people's heads or inboxes, not systems
- Nobody agrees what a good outcome is
- The underlying process changes every few months
- The main aim is to have an AI project to talk about
- A clean report of existing data has never been produced
If your problem passes all five signs, the next step is small: pull a sample of historical data, build a simple baseline, and test whether a model beats it. That is a few weeks, not a year. SpiderHunts runs exactly this kind of short feasibility study as the first stage of machine learning work, and if the data is not there we tell you what to start recording. For inspiration on where the signs tend to line up, see machine learning use cases that drive revenue. And if what you actually need is a language model reading documents rather than a prediction, our AI integration work is the better starting point.
Frequently asked questions
How do I know if machine learning is right for my business?
What should I try before machine learning?
How long does a machine learning feasibility study take?
Can machine learning help with one-off strategic decisions?
What if my problem passes some signs but not all?
Think you have a machine learning problem?
Run it past the five signs, then tell us what you found. We will give you a straight answer on fit, including when a simpler approach would serve you better.