The limits of artificial intelligence when applied to marketing
Many companies in highly competitive industries use big data and artificial intelligence to analyze their customers’ behaviour. But it is important to understand the current limits of this tool.
With the development of big data analysis models and recent successes in the field of deep learning in artificial intelligence (AI), predictive marketing seems to have become the answer for many companies in highly competitive situations.
Essentially, it involves analyzing the behavior of prospects and customers from the massive amount of data available in order to deduce typical consumer profiles that then make it possible to propose the most suitable products or services.
The benefits of AI
An example often given to illustrate the applications of predictive marketing is subscription services providers having access to the sites their customers visit, and they use this information in order to anticipate possible cancellations of their subscriptions. Once detected, customers are contacted before their decision in order to make them an attractive offer.
There is, in fact, an immense amount of personal data, particularly on social networks, which provides valuable information to respond in near real-time to the needs and desires of consumers.
In its ultimate version, a predictive marketing AI algorithm’s job will be to anticipate needs not yet formulated by customers themselves akin to the Precogs in the film Minority Report which identify=ied murderers before they acted. This will likely never be possible in the real world, but something resembling it might very well be possible in 50 years.
Without wishing to question the obvious advantages of intelligent data analysis, however, it is important to point out certain limitations and risks associated with AI.
Correlation and causation
The first limitation stems from the fact, well known to statisticians, that correlations are not indicative of a causal relation. We can illustrate this point with a humorous example: the number of documentaries produced per year obviously has no impact on the number of pigs slaughtered during the same period and vice versa. However, the two curves show an almost perfect similarity with a correlation coefficient of 0.974 (the maximum is 1).
Such correlations are frequent and it is necessary to implement context-sensitive analyses to detect them and reject them. A possible approach to solve this type of problem is to use algorithms from artificial intelligence research.
For example, DeepQA, the software architecture on which IBM’s Watson supercomputer is based, brings out several hypotheses from the analysis of massive data, which it then classified according to indices of relevance and confidence, before choosing the “best” solution to the problem posed. While the example we have given of correlation not equalling causation is obvious, others are much more difficult to assess depending on the areas and problems being considered. Therefore, it will generally require the joint input of AI and human specialists to obtain meaningful results.
A second limitation is that of the algorithms generally used in recommendation systems. To simplify, they are mostly based on the principle of associating a client to a typical profile that emerges from the data analysis.
The best-known example can be summarized by the formula “customers who bought this also bought that”. Like television channels, whose programs depend on a small number of typical profiles, we obtain solutions that are ideal for these hypothetical profiles but they are far from ideal on the individual level.
This limitation causes significant friction and might even damage your sales if the prediction algorithm is not implemented correctly. This is why it is recommended that you hire AI consultants to help you implement AI algorithms in your business.