The Emperor’s New AI – Artificial Intelligence in Marketing: Hype, Failure and Sensible Use Cases

AI has long been ubiquitous – from social media to web shops. Large and small software manufacturers rely on AI functions everywhere, but many tech companies are desperately looking for real use cases. Despite the enthusiasm, there is increasing criticism; this article attempts to identify the “state of the art”.

AI and its limitations: Failures and typical criticism

The internet is full of AI failures: People with excess limbs, nightmarish image landscapes and search engines that recommend pizza with glue, or say that the inventor of the “backflip”, John Backflip, was banished in 1316 because he was accused of witchcraft by his rival William Frontflip. These ridiculous mistakes are often very amusing, but they do raise the question of whether the AI products were tested at all before they went live. One thing is clear: Much of this will be taken care of by further developments. The problem with limbs is now less prevalent, and companies are doing everything they can to minimize the damage to their image caused by failures. To dismiss AI as pointless because of these kinds of problems is simply superficial and wrong.

Sensible use cases for AI in marketing

AI offers a number of useful applications, not only in medicine or the insurance industry, but also in marketing. Natural Language Generation (NLG) automates the creation of text from technical data, such as product descriptions and campaign texts, often in multiple languages and at high speed.

In the area of product data quality, AI helps to categorize products, tag images and extract text from spoken language. It also checks the quality of texts or the abridging of content – although in the latter case, AI often does not deliver the desired added value.

Personalization is another big topic. AI is used for product recommendations, to identify preferences and to recognize trends from large amounts of data. It also improves product searches by taking into account similar formulations.

Expectations and reality

Unfortunately, the constant hype and secrecy creates unrealistic expectations. In reality, none of the AI products on the market is intelligent in the sense of human intelligence. There is no strong AI (artificial general intelligence) (yet), and there is currently nothing that comes close to “real” intelligence. It starts with the fact that there is disagreement about what intelligence actually is. When someone claims that their AI models are on the verge of making the transition to strong AI, that person is usually raising funds for the next round of financing or an IPO.

In practice, we are talking about generative AI, machine learning / deep learning and computer vision. These technologies make it possible to recognize patterns in large amounts of data and make predictions based on them. None of this sounds like independent thought or intelligence – and that is exactly the point.

Is ChatGPT bullshit?

A good example of misconceptions about AI is how large language models (LLMs) like ChatGPT work. These models have neither knowledge nor understanding. They merely generate statistical strings of words. GPT uses the prompt to set a starting point and then strings together the words that are statistically most likely. If you end up in an area where the model has been well trained, it often makes sense. If you end up in an area that has not been well trained, the result is simply bullshit in the sense of Harry Frankfurt (see “ChatGPT is bullshit”).

This is also evident in text abridgements and content checks. LLMs can abridge, but they often distort meaning, leave out key passages, and emphasize unimportant side issues. Another problem is that the AI models are mathematics – not magic. They learn from the data they are fed. If this data is generated by people who follow simple algorithms, the AI will adopt this behavior. For example, if the data reflects a preference for certain applicant groups, the AI will mimic and even reinforce these preferences.

The thing about data

It can't be done without data, but more data doesn't necessarily make it better. Data can contain biases that are reflected in the AI models. Also, the understanding of what is done with this data is often limited. Retailers in Germany, for example, collect huge amounts of data, but use it only to a limited extent for personalizing advertising. The EHI Marketing Monitor for Retail 2023–2026 makes it clear that 29% of retailers personalize poorly or not at all, 24% use simple segmentation, and 47% do not attempt a 1:1 approach. Despite huge amounts of data, personalization is rarely implemented in many cases.

The way data is collected also causes unease. Many customers feel that data collection is inappropriate, which is why ad blockers are widespread. Apple has tightened its privacy policy and automatically deletes even “agreed” cookies every two weeks. Google, on the other hand, is postponing the replacement of third-party cookies because the underlying data collection ecosystem is not yet ready to abandon this data source.

Use data or flip a coin?

A study on the efficiency of data use showed that 3rd-party data brokers are of little use in practice. In a study on reaching the “right” customer, it was found that only 14.3% of IT decision-makers were correctly addressed when 3rd-party data was used. The results were even worse when searching for senior IT decision-makers – only 7.5% were actually correct.

Even the gender was only correct in 42.3% of cases. Even a coin toss is more accurate than that. This shows that using 3rd-party data is not a reliable way to reach the right customers. This emperor is completely naked. 

What can be deduced from data?

What can you learn from data such as the number of seconds a video was viewed or the number of likes a post received? Engagement ultimately only measures the “level of excitement”, but not the cause of the interaction. A high click or like value does not necessarily mean that the content was of high quality or relevant to the user. To actually gain deeper insights, you have to correctly assign cause and effect. AI can recognize correlations, but it cannot determine causality. This is only possible with domain knowledge that AI does not have. It is good at finding patterns, but humans have to do the interpreting.

Understanding AI means better using AI

Artificial intelligence offers an enormous opportunity if you understand its limitations and possibilities. Companies should ask themselves: What use cases are there in their own organization? Where can AI provide insights that would not be accessible otherwise? And is the better performance actually useful and economical, or is the higher cost not justified? This aspect in particular is often overlooked in the wake of hype and FOMO. Those who understand AI can use it effectively and take advantage of the many opportunities it offers.
Don't let anyone talk you out of AI. Test it with a pilot or a proof of concept. Work out your business case, get advice, and get started, preferably with a partner you can work with on an equal footing. We know just the right people…