Four Situations Where you Should Not use Machine Learning
Machine learning (ML) is a great technology. But does it really have to be used all the time? This post is about four situations where you shouldn’t use machine learning.
There is (rightfully) a lot of hype around artificial intelligence and machine learning (ML). As we said, great technologies and use cases are emerging from it. You can also find many examples that use machine learning in our blog articles. But in which situations is machine learning not necessary?
Remember, there are always exceptions to the examples to come. However, your AI project will probably be five times more difficult if these examples apply.
1. No Data “Cold Start Problem”
Without data, there is no machine learning. And by data, of course, we mean data related to your application problem. You want to predict the weather with machine learning, then you need historical weather data. You want to predict your next sales from repeat customers, then you need historical sales data.
It’s also called a “cold start problem” if you don’t have enough or any or data for a use case of ML. Machine learning algorithms and models cannot generate recommendations from it.
A possible solution would be if you use data with similar characteristics so that the algorithms apply them to the new situation. But even these must first be available.
Imagine you are launching a new product and want to know which of your customers might be interested in this product. A new product always means no historical sales data on it. Here the problem can be solved by using data from similar products for your ML system.
An example where even ML systems can’t help you would be if you want to predict which offers will be best accepted by your customers without any data at all.
2. Use Case has Zero Tolerant for Mistakes and There is NO Supervisor Available.
Almost all machine learning systems work with probabilities. With what probability is the animal in the picture a cat? With what probability is the chosen route the shortest? What is the probability that it will rain tomorrow and what is the probability that you like the suggested Netflix movie?
So even the best ML systems with over 90 percent accuracy are wrong from time to time. That is the nature of machine learning.
So in application, you have to ask yourself, what kind of damage does a “not accurate recommendation” do? If Netflix suggests 9 movies that you think are great and one that doesn’t appeal to you – no big harm. Just like Amazon product recommendations: basically any recommended product you DON’T click on is wrong. The damage from products not being accepted? Not great. The benefit, however, even greater, because 25% of Amazon sales come from these recommendations.
Let’s move on to use cases where there is no margin of error. In medicine, machine learning systems are now being applied to assist in diagnosis. And “assist” is the most important word here. Medicine is a field in which errors can be fatal. And yet ML systems are being used. This is only possible because the physician ALWAYS remains in control, checking and balancing the ML system’s recommendations against other factors.
So if you have a use case that allows zero tolerance for error and you don’t have a professional available to check (supervisor), then you better not use machine learning.
3. Your employees are not open to new things
The most successful machine learning applications in B2B are hybrid. That is, humans use AI software as support. A recommender in an online store runs without humans in the middle. A sales team using sales forecasting software to prioritize sales activities is an example of a hybrid machine learning application.
However, working with algorithms also requires a change in internal processes. The technical term “Algorithmic Management” deals precisely with this issue: How should an employee react to the recommendations of algorithms? What control mechanisms should be observed to decide whether to accept or reject the recommendation.
A machine learning system is only as good as its users. A lack of willingness on the part of users to open up to new technologies makes the introduction of machine learning enormously difficult – if it doesn’t even fail because of this.
4. Your Rule Based Solution Works Fine
When using machine learning, the primary focus should always be on the added value to a specific goal. In other words, to what extent can the new technology fulfill the goals better than existing systems. We do not recommend using machine learning just because it is “in”.
Rule-based applications are systems with rules set by a human. A classic and simple example is an “If, then…” function of Excel: “If the customer has not bought for more than 3 months, then mark him red”.
Many companies analyze their data with Excel spreadsheets and as long as that works well – i.e. serves the goal – and is not too costly, then there is no need for machine learning.
From our experience, this is true for small and young companies that don’t have much data yet either. At some point, rule-based systems reach their limits and then machine learning is the perfect solution.
Four Situations Where You Should NOT use Machine Learning – Conclusion.
AI applications based on machine learning offer great opportunities for companies: Processes can be made more efficient, planning reliability can be increased, and costs can be significantly reduced.
But there are also exceptions in which machine learning should not be applied. In principle, these exceptions arise due to the nature of machine learning: data is needed, a tolerance for error must be granted, and working with algorithms requires a change in daily routines.
If you want to know how AI can be applied in sales, we have an exciting post for you here.