Definition von künstlicher Intelligenz, Anwendung im Marketing und in der prudsys RDE Recommendation Engine & Zukunftsvision

Definitions, AI in marketing, AI in retail & what prudsys AG has to do with it

When it comes to labeling the next disruptive technology, no term is used as frequently as that of artificial intelligence (AI for short). Regardless of the medium in which I move these days, the topic of AI is one that confronts me everywhere I go.

For example:

  • AI was one of THE topics at dmexco 2017.
  • Microsoft is currently advertising a webinar entitled “Artificial intelligence in retail” on Facebook.
  • t3n leads with: “Künstliche Intelligenz als Geschäftsmodell: Wie Unternehmen sie einsetzen können” [Artificial intelligence as a business model: how companies can use AI] (an online version of the article “Das Fundament der künstlichen Intelligenz” [The fundamentals of artificial intelligence] in Issue 48/2017).
  • A few days ago I received an online invitation from Joachim Graf at to the online event “How AI, IoT and ML are reinventing marketing and commerce”.
  • In Issue 41 from 13 October 2017, the “Lebensmittel Zeitung” ran the headline: “Künstliche Intelligenz begeistert Händler” [Retailers enthusiastic about artificial intelligence].

However, before we get into what all the hype is about, we should first clarify what we are even dealing with in terms of technology.

What exactly is artificial intelligence?

Scientific literature uses the English term artificial intelligence, or AI.

  • AI is real science, mathematics, algorithms and formulae.

There is still no standardized definition of AI. It is a huge field and the definition varies depending on the point of view and background of the viewer.

Definitions of artificial intelligence

Different sciences have contributed their own ideas, questions and techniques to the topic of artificial intelligence: philosophy, mathematics, economic sciences, neuroscience, psychology, computer science, control theory and cybernetics as well as linguistics. Russel & Norvigs’ definition of AI as “the search for the best agent program in a certain architecture” (my translation) seems unsuitable for the purposes of this article.

Dr. Vishal Sikka, CEO of Infosys, points out in his report that there is still no scientifically recognized definition of AI. Instead, he refers to MIT professor Marvin Minsky’s definition of AI: “the science of making machines do things that would require intelligence if done by men.” He finds good examples of this definition in the fields of visual perception, language recognition, machine learning, decision making and in the processing of natural language.

In his interview appearing in Lead Digital, Head of Webmetrics Alexander Korth breaks it down into one short sentence: “by definition, artificial intelligence (AI) simply means that a machine is doing something for which a person would need intelligence.” In the same breath he restricts the universality of intelligent algorithms: “AI solutions can only be applied to a certain task or certain type of task. The assumption that an intelligent algorithm can be assigned to a strange task is incorrect.”

Wikipedia defines artificial intelligence as “The attempt to simulate human intelligence, i.e. to build or program a computer so that it can work on problems independently.”

John McCarthy formulated it in a very similar way as far back as 1955: “The goal of artificial intelligence is to develop machines that behave as if they were intelligent” (quoted by René Büst in t3n).

In practice, “a definitive” AI is nowhere to be found – Oliver Schonschek published a very interesting article about this (“Moral, Kreativität, Lüge: Gibt es überhaupt Grenzen der künstlichen Intelligenz?” [Morals, creativity, lies: Does artificial intelligence have any boundaries at all?] from 27.9.2017) in As a point of fact, artificial intelligence appears in a variety of functions, applications and services such as:

  • Language: artificially intelligent services hear, speak and respond.
  • Perception/vision: artificially intelligent applications “see” pictures, i.e. they recognize and evaluate picture information.
  • Learning/remembering: artificially intelligent services collect data, evaluate it and learn from it.
  • Response: artificially intelligent programs respond to “incentives” in the form of data or events and “do” something.

Artificial intelligence in marketing

In marketing in particular, AI services help analyze ever growing volumes of data, recognize patterns in user behavior and derive predictions for a wide variety of applications. When it comes to accurately generating advertisements, segmenting target groups for mass mailings and many other tasks that are extremely time and labor intensive for humans, AI services have distinct advantages. They help marketers perform routine tasks. At the same time, the creation of appropriate campaigns is becoming more and more significant.

Where and how AI is used in marketing today and what the future may hold is illustrated in the article “Realität und Zukunft: So wird KI im Marketing eingesetzt” [Reality and the future: this is how AI is used in marketing” by Georg Loewen, Head of Marketing DACH at Selligent, which appeared in Internetworld Business on 13 July 2017. In this article he includes – similar to the abovementioned services – recommendations in online shopping (using recommendation engines like the prudsys RDE), chat bots in customer service, digital assistants in daily life (such as wearables, Google Home, smart fitness apps, the Netflix AI) and personalized search results.

Artificial intelligence in the prudsys AG range of solutions

At prudsys AG, we have been involved with the topic of AI since 1998. Our personalization solution, the prudsys Realtime Decisioning Engine (prudsys RDE for short), is based on artificial intelligence and achieves the best possible results in the fields of recommendations, marketing automation and dynamic pricing for retailers and customers.

Jens Scholz, founder and current CEO of prudsys AG, said it best in a recent interview (blog entry in German). “prudsys AG has been involved with artificial intelligence and the automation of decisions since its founding. […] We found the challenge of measuring data in real time, learning directly from that and being able to play back the results in real time to be particularly exhilarating.”

Dr. Michael Theß, co-founder of prudsys AG, has published his knowledge in several specialist books and articles and has given several talks in the fields of data mining, realtime analytics, artificial intelligence and the application of these things in a personalization solution for retail. One of his favorite books is the over 1000-page tome entitled “Artificial Intelligence: A Modern Approach” by Stuart Russel and Peter Norvig – THE international standard textbook for artificial intelligence. The 3rd edition, published in 2010, is next to me now (online PDF here).

prudsys employs specialists on its data science team who are constantly conducting research on innovative methods and continually optimizing the AI underlying the prudsys RDE. It is mainly intelligent functions in the areas of learning/remembering as well as response that are used.

Adaptive machine learning and reinforcement learning

From a historical perspective, the development of our AI technology started with data mining. Along with the availability of large volumes of data came the desire to analyze ever larger datasets (keyword: big data), recognize patterns and use this knowledge profitably. Data mining continued to develop in the direction of machine learning, which far exceeds simple pattern recognition in datasets. Following the learning phase, “the machine” generalizes its knowledge and is then able to apply it to unknown data. In the beginning “the machine” learned only from static datasets but there was soon a need to learn adaptively (or incrementally). When it comes to adaptive machine learning, “the machine” learns from every new dataset that it processes. There are considerable advantages to this when it comes to both speed (keyword: real time) and quality. The latter improves even further if “the machine” also receives a reward for every “good” action. We refer to that as reinforcement learning or learning through rewards and punishment. “The machine” learns how to act in potentially occurring situations in order to maximize the benefit of the overall system.

That was a very simplified summary. Wikipedia provides a decent, detailed overview of each individual keyword. What I find exhilarating is how one aspect leads to the next. The correct term for “the machine” is algorithm or agent, which brings us back to Russel & Norvig’s definition.

Once again, in detail: how intelligent is the prudsys RDE?

All of the prudsys RDE applications are based on data that provides information about the customer and his needs. This includes, for example, behavioral data, historical data, transaction data as well as CRM and ERP data. The prudsys RDE recognizes patterns and regularities in the data and can thus even evaluate unknown data due to advances in the learning phase (= machine learning). In other words, the prudsys RDE automatically learns personalization rules in real time – e.g. on the basis of clicks, shopping carts, purchases, external and internal search queries, clicked categories and banners, or self-selected events. The software collects all of this data independently, and most importantly, continuously.

However, the prudsys RDE algorithms do not only recognize and respond to customer behavior (e.g. with individual recommendations), they also measure the acceptance of the generated content at the same time (= reinforcement learning). If a generated recommendation is accepted by the customer, the system rewards itself. If the recommendation is not accepted, there is no reward. This way, the prudsys RDE learns gradually and can better cater to each individual user in a more targeted way with each interaction.

Translated to the use of the recommendation engine in retail, this means that if you use an algorithm that works according to the principle of reinforcement learning, you can maximize the benefit of your “whole system” over the long term – in other words, the crucial variables such as sales, turnover and profit.

But there’s more: the prudsys RDE can deviate from learned patterns by way of so-called exploit and explore mechanisms, e.g. to recommend products or content that has never been recommended or have not been recommended for a very long time and are thus not part of the learned set of rules. The prudsys RDE ensures that its scope of learning and action remains flexible by trying out alternatives or new variants. For the end customer this means that he is always offered new content or products and does not continuously receive the same recommendations in a kind of filter bubble.

This is a great advantage, especially for quickly changing product ranges such as in the fashion industry: products from new collections or brands flow automatically and immediately into the recommendations.

What does that have to do with chess?

The easiest way to explain how the prudsys RDE works is using the analogy of a chess computer: the software recognizes the customer’s “moves,” predicts the chain of all possible subsequent “moves” and acts accordingly (e.g. product recommendations) so that it wins “the game” (e.g. the target parameter of sales optimization).

Recommendations & more

Apart from recommendations, the prudsys RDE offers many other intelligent services for use in omnichannel retail. Or vice versa: different applications of artificial intelligence are used to implement other services within the prudsys RDE.

In addition to reinforcement learning, machine learning features other algorithmic approaches that the prudsys RDE uses:

  • (Adaptive) supervised learning: “The machine” learns from the existing data what impact the various attributes have on target features. The goal is to predict the target feature on the basis of attributes.
  • (Adaptive) unsupervised learning: this is where “the machine” tries to recognize statistically relevant patterns in the data without knowing a target feature in advance.

Supervised learning algorithms are used, for example, to predict shopping cart cancellations and returns in the field of marketing automation. The goal is to identify at an early stage those customers who do not buy or who often return goods in order to use appropriate measures to positively influence the buyer while shopping. This occurs, for example, by automatically generating so-called purchase incentives. However, the calculation of optimized prices is also based on unsupervised learning algorithms. The target feature to be optimized here, depending on corporate objectives, is the maximization of profit, turnover or sales. After the learning phase, the algorithm calculates the price that maximizes the target parameter.

The prudsys RDE uses unsupervised learning algorithms (especially so-called clustering) to enrich the system with information via interrelationships between products. Human users create those relationships by placing certain products in the shopping cart together and/or purchasing them. This information flows directly back into “the machine” for subsequent action following pattern recognition.

What does artificial intelligence mean for future retail?

“Artificial intelligence (AI) will revolutionize retail”, concludes marketing performance company OMD in its study entitled “Retail Revolution” vis-a-vis acquisa. In conjunction with Goldsmiths University London, experts surveyed more than 15,000 consumers in 13 European countries regarding their perception and behavior in relation to artificial intelligence (AI) in retail. According to expert opinion, the focus was predominantly on the optimization of the communication process between companies (brands) and consumers with the help of intelligent services.

Alexander Korth still feels that there will be great differences between retailers in five years’ time, depending on industry and size. However he does see a common trend: retailers are building up more knowledge about their customers by merging it from previously non-networked systems and enriching it with external data. These customer profiles result in retailers communicating the right message to customers using the right channel.

Will artificial intelligence make marketing redundant?

No, but it will change. Expectations are extremely high: a study conducted by KRC Research found that approximately 55 percent of marketing decision makers surveyed “are convinced that artificial intelligence will change the marketing industry more than social media has managed to change it”, summarizes Georg Loewen. Many solutions are already working with AI, we just don’t see it.

Adobe DACH has also broached the challenges of modern, data-driven, digital marketing and the necessity of intelligent tools “that allow for not only the collection and understanding of information but also for drawing the right conclusions from that information – in order to generate the right result for the customer in real time”.

In an interview with the magazine acquisa on 2 August 2017, futurist Gerd Leonhardt concisely summarized: “Machines are better than us at everything that is routine and that has to do with numbers and mathematical intelligence.” Even if routine jobs fall by the wayside, there is an opportunity for marketing because that is exactly what “gives us the freedom to concentrate on what really matters: developing relevant stories, building relationships, understanding our customers and developing campaigns together. Those are all tasks that only humans can accomplish.”


Artificial intelligence has long been an established reality. For most people, however, it is invisible and works in the background. We are dealing with it on a daily basis when we shop online or read an online newsletter. The topic is currently experiencing a hype that is stirring up ethical concerns and fear in society. Even internet pioneers like Bill Gates and Elon Musk are very critical about current developments in artificial intelligence. Irmela Schwab (writer for W&V and Lead Digital) clearly illustrated the limits of artificially intelligent assistants in the 10/2017 edition of Lead Digital. She points to the introduction of the General Data Protection Regulation in the EU which came into effect in May 2018 and to its effects on the use of data in marketing and in terms of training data for AI. And she pleads for a strategically thought-out introduction of artificially intelligent systems: “For the human-machine duo to work well as a team instead of fighting against one another, exact future scenarios must first be considered and appropriate rules put in place”, says Schwab, summarizing her most important challenge when it comes to dealing with AI. In practice, many industries benefit from artificially intelligent helpers which in turn make life easier for people. Applications in the field of smart homes and digital assistants are slowly becoming commonplace in our daily life.

Good to know: Artificially intelligent systems always take care of just one thing, this is what they are trained for and they do it extremely well. “Computers are good for working in closed environments – chess and Go, click patterns and machine control. As soon as you leave behind the known, as soon as there is more than one skill needed, the natural stupidity of artificial intelligence comes to the fore: chess computers are quite unskilled when it comes to playing soccer, welding robots do not play the piano well” says Joachim Graf, futurologist, journalist and editor at in his sharp-tongued commentary.