AI-Powered Insights: The Liberation From Spurious Correlations

March 7, 2024

Chapter 4

THE VALUE - Applications of Causal AI in Marketing

Dr. Frank Buckler

Founder, Success Drivers

Theory is patient. Not every better method has to be used straight away. It is therefore worth taking a look at where Causal AI can be used and what practical added value it provides.

I will touch on a few fields of application as examples. This may give you the impression that Causal AI can be used universally. You might think to yourself “yes, of course, it just makes more sense”. This realization alone will inspire you. 

Just as the benefits of mantras such as “Lying doesn’t pay” or “Being human wins” are not obvious in the short term, life experience shows that they work. Just as a human character with integrity “pays off”, more logically consistent approaches are also more successful in the long term.

But what can you do with Causal AI? I differentiate between three activities: 

  1. Explain: Explain what causes success

  2. Decide: Choosing the best marketing decision

  3. Generate: Create marketing content that is more effective 

You can also aim for several activities with one model. However, differentiating the goal leads to greater clarity of thought.

Explain Better with Causal AI

Since its creation, artificial intelligence has been used to automate decisions. Which product should be suggested to the customer now? Which customer should we send a mailing to stop them from canceling? Will this customer be able to repay their loan?

For these questions, it is not necessarily problematic that AI is a black box. After all, what counts in the first instance is the right decision, not an understanding of how this comes about.

But if we want to optimize the marketing mix, if we want to decide which initiatives we can use to improve the customer experience, we need to explain what drives success. If we want to understand how we can improve our products so that they sell like hot cakes, if we want to fundamentally uncover the hidden reasons and contexts that lead customers to buy or abandon, then we should ask “why”. 

Today, this “why” question is primarily answered by qualitative market research and the qualitative exchange between experts. Statistical modeling also takes place from time to time, but due to its limitations, it is only used in borderline areas. Every method has its value and its place. But Causal AI can now cover a field that neither qualitative research nor statistical modeling could serve well.

Marketing Mix Modeling

The situation at a gambling provider was tense. Turnover had been falling for years and the managing director rightly wondered whether the millions spent on advertising had been well invested. 

In the lottery business, the advertising channels include traditional media such as print, posters and radio, as well as shopping radio in supermarkets and advertising at points of sale and on websites. The media planners determined the distribution of advertising according to general principles. 

One well-known belief, for example, is that radio advertising is the fastest-acting form of advertising. This is why it is used on a massive scale for short-term topics, such as jackpots, in order to increase sales. However, these beliefs do not reveal exactly what advertising pressure is needed. 

Our team differentiated the use of media according to the content of the advertising. Was a jackpot advertised, was it an image advertisement or was it a special promotion? Other important influencing factors were the size of the jackpot, the current press coverage, the specific days of the week and the time of the month. 

It is always important to include all influencing variables in a model, regardless of whether they are controllable or not. This is because false findings (confounders) can only be avoided if the model is as complete as possible. If, for example, radio advertising is mainly placed on weekdays, but sales are usually lower during the week than at weekends, a negative effect would be wrongly attributed to radio. The careful collection of possible success variables therefore plays a central role.

The results of the analysis showed surprising results. Non-classical media were many times more effective than classical media. Many media interacted strongly, so that a combined use was recommended. It was also surprising that short-term media such as radio did not have a short-term effect, but had to be used for as long as possible. However, this only superficially contradicts the belief. The analysis shows that radio only creates cognitive awareness. In order to sell a ticket, the customer must first visit
the point of sale. Almost nobody does this just to buy a lottery ticket. The purchase of a lottery ticket is usually an impulse purchase, which is merely encouraged by the awareness built up beforehand. Therefore, the movement habits of the target customers define the speed at which the advertising effects take effect, and not the short-term effect of the radio.

As a result, we were able to show that advertising for these gambling products is so worthwhile that a budget reallocation brings considerable increases in effectiveness. The optimization of the media plan led to further continuous growth in the following years.

The causal analysis approach also provided further strategic insights. The jackpot amount was seen as a central lever to drive the lottery ticket business, as the correlation between jackpot amount and revenue was obvious. However, communication that focused on jackpots (causally) caused the slump in sales after the jackpot peak. Customers learn that it is not worth buying a ticket without a high jackpot. Thus, the communicative focus on the jackpot leads to the destruction of what lottery is all about – a nice weekly habit.

So what is the difference to classic marketing mix modeling – a discipline that has been practiced for over 100 years (Unilever built its first MMM model in 1919). If there is one field in which statistical modeling is mature, it is MMM. A lot of experience and intuition is used to bend the models so that they “somehow” fit.

An MMM with Causal AI naturally offers these advantages in particular:

  • Avoids incorrect attribution due to confounders that are not (or not correctly) integrated. The seasonal information, which is always included as a variable in a causal AI model, is an example.

  • Recognizes non-linear effects (such as saturation effects) without specifying them in advance at great expense and recognizes interaction effects (such as the mutual reinforcement of radio advertising and POS advertising). The classic consideration of such hypotheses would involve extreme effort.

  • Natural modeling of short, medium and long-term effects. By far the largest part of the advertising effect is long-term, because it works again and again over the long term through the formation of mental structures. Disregarding this leads to misdirected marketing.

  • Avoids incorrect attribution due to strongly correlated variables. In particular, online channels (earned and owned media) correlate very strongly with paid media. Here we see strong distortions in conventional mix modeling.

Customer Experience

What makes customers happy? What annoys them? How can you retain them? These are questions that an entire industry is now dealing with. 

Customer satisfaction surveys have been around for a long time. However, the NPS concept has introduced a very simple methodology. Its advantage is that it is easy to ask a question at each touchpoint and, if necessary, to receive an open-ended explanation. Today, CX software providers such as Qualtrics, Medallia and InMoment support the design, implementation and evaluation. However, success is sparse and CX experts argue about why this is the case.

In my experience, a key reason for this is that customer feedback is analyzed too superficially, resulting in crucial misjudgments.

The streaming speaker brand SONOS decided to have the touchpoint surveys analyzed with Causal AI. The NPS had been falling for some time. The explanation given was that with market penetration, other, less enthusiastic customer types were changing the customer mix. But they were not sure.

Using Natural Language Processing , we analyzed the open text fields and categorized each topic that customers mentioned into one of 80 categories. In this way, tens of thousands of responses can be categorized with the highest precision. Today, the accuracy is even higher than if a human were to categorize it. This is because humans get tired and tend to deviate from their own definitions depending on their mood.

This type of AI application in the CX area is standard today, even if very coarse, unspecific text AI models are still often used.

A look at the frequencies only paints a seemingly clear picture. Almost every second customer based their rating on the good sound of the loudspeakers. It is not surprising that the consensus within the company was that sound quality was the key factor in customer satisfaction and loyalty.

Many other companies that I have come to know fall into the same trap. Restaurant chains believe “tastes good” is crucial, washing machine manufacturers believe “washes well” and insurers believe “has good service” are the decisive levers.

Using the categorization data, we created a causal AI model and the result showed that sound quality was nothing more than a hygiene factor. What made customers really loyal was the experience of smooth functioning. This could sometimes be disrupted by various technical reasons. However, this experience of smooth operation was not mentioned so often. No wonder, as there was potential for improvement.

The following graphic provides an insight into the Key Driver Matrix, which shows the frequency of naming on the vertical axis and the significance on the horizontal axis (positive on the right, negative on the left).

On the right you can see the impact simulator of the tool. Improving the software architecture for smooth functioning promised the potential to increase the NPS by 4 points. Based on reference values determined with modeling, these 4 points can now be converted into additional sales.

The analysis reveals many other valuable insights that would have remained hidden in the conventional way. For example:

  • “Great Sound” is an emotional, non-technical statement: an AI score that measures the tonality/emotionality of speech is integrated into the model. According to the model, this “Great Sound” variable only has an indirect effect via this emotionality. In terms of content, this means that by “Great Sound”, customers mean the good feeling that the music brings and not the technical sound characteristics that the traditional hi-fi industry has always focused on.

  • The product itself is very emotional. 30% of the variance is not explained by what is said, but how (with what emotionality) it was said. In comparison with insurance, this value is only around 10%.

  • Some topics are hygiene factors like “Easy of Use” and some are excitement factors like the “Voice Assistant” feature.

The topic of customer experience is a useful introduction to the topic of Causal AI. As a rule, companies are sitting on a lot of data that just needs to be analyzed in a more meaningful way. In my books “The CX Insights Manifesto” and “CX Insights Playbook”, I describe in detail how this can be achieved.

One of the advantages of using Causal AI in the customer experience area is that

  • Various data sources such as binary categorization of open-ended responses can be integrated into a model with metric emotion scores and any Likert scales (problematic in statistical modeling)

  • Indirect effects are taken into account. Topics in open-ended responses often have different levels of abstraction. For example, friendliness and good service are not causally independent. Rather, friendliness leads to good service. Taking these interdependencies into account makes it possible to better measure the causal effects and thus derive the right decisions.

Product & Innovation

Which product features are important to customers? What should be taken into account when developing new products? Which product features should be placed at the center of product communication? These are the classic questions that arise in the product innovation process. Interestingly, they also arise again and again in the life cycle of the finished product, because the consumer changes and the competition changes with it.

For this type of question, we at Success Drivers have founded the platform, which provides modern standardized solutions for price and product optimization as well as brand and touchpoint optimization. All of these solutions use Causal AI. They also apply an innovative market research method developed in the field of neuroscience. The “Implicit Association Test” can measure the unconscious opinions of consumers in a very simple way using a reaction time-based query. It is precisely this information that is decisive for purchasing decisions.

The price and product optimization of the APPLE VISION PRO was one of the first use cases we carried out a year before the launch. Setting up the study only required a product image and a description that briefly describes all relevant features and uses the language of the brand. The tool suggests a price range and product features based on LLM. The latter were then revised once again by Expert Judgement.

Then 250 computer users in the USA were interviewed and half of them were so-called “early adopters”. These are people who generally buy innovative products first without waiting for the product experiences of others. The willingness to pay was first measured implicitly and the extent to which the product can actually fulfill the twelve features was also implicitly queried. Features included “long battery life”, “ultra high display resolution” and “easy intuitive gesture user interface”. 

The result was a profit-maximizing price of USD 1,999 across all consumers and USD 3,499 for early adopters. The price-profit function is calculated by multiplying the calculated price-sales function by the price and then subtracting the retail margin and the estimated unit costs.

Causal AI was used to determine what leverage the features have in order to increase the willingness to buy (which the price test measures). Features that have a high leverage, but are not perceived as a given or are even doubted, should be improved communicatively or technically.

The following illustration shows the dashboard of the tool. It was the revolutionary gesture-based user guidance and the 3D applications that had a huge impact on the willingness to buy. At the same time, however, there was clear skepticism as to whether the user guidance would really work so intuitively.

Apple communicated precisely these aspects at the launch and showed in videos how exactly the user guidance works as well as trustworthy people in their usage situation.

Since then, the methodology has been successfully applied in many other markets. For example, a major shoe brand understood that, despite all the interesting features that product development gave its shoe variants, something completely different is the main reason to consider a shoe: The design has to fit your style. The customer simply has to like the shoe. This sounds banal. But sometimes it is precisely this banal evidence that experts need in order to see the wood for the trees.

In conventional market research projects, conjoint measurement or MaxDiff in particular are used for this type of question. The application of Causal AI combined with neuroscience measurement methods offers the following advantages:

  • 50 or more product features can be evaluated instead of a maximum of seven with conjoint measurement

  • The product can be experienced qualitatively in its entirety with a picture and comprehensive product description instead of as a bullet list of a maximum of seven properties, as is the case with Conjoint.

  • The importance of the attributes is derived from an implicitly measured and calibrated absolute willingness to buy instead of only in relation to a competitive set that will not exist in this form in reality and that can change depending on the context.

An alternative is the MaxDiff method. It forces the respondent to make a trade-off between attributes. It is an intelligent form of direct inquiry to find out which attributes are important. Because it asks directly about importance, this method is rationally biased and blind to the unconscious association processes that determine purchasing decisions. The new method (Supra Product Optimizer) addresses both.

The most important thing, however, is of course the validity of the method. Is it true what the method determines? This is exactly what Causal AI is designed for – causal attribution, the reason for the purchase.

Communication / Advertizing

What makes a good advertisement? What should advertisers pay attention to so that the advertising works? What guidelines and provisional rules help us to avoid the risk of landing an advertising flop?

Questions that have long been left to the land of qualitative knowledge. However, advertising test procedures have always shown that there is an enormous disparity. Most commercials are only very moderately effective and only a few are enormously effective. However, the unknown formula that is sure to generate advertising with a high ROI is still controversial today.

In 2016, insurer Metlife asked us to look deeper into their advertising test data to better understand what makes advertising successful. In 2017, we then developed this method further for this purpose and tested it extensively in six sectors. 

I was subsequently invited to speak at several conferences. While my presentations to market researchers were received with reserved interest, the feedback from “creative” audiences was rather poor.

I still remember giving the final presentation at the Shoppers Brain Conference in Amsterdam. The presentations were evaluated afterwards by the participants via an app and as I was the last one, I was hoping to receive particularly good feedback. After all, in my presentation I showed a clear method of how advertising could be optimized holistically and systematically for the first time.

The result gave my ego a “kick in the butt”. It was the worst feedback I had ever received. The slightly angry questions from the audience after the presentation should have given me pause for thought.

If you had claimed back then that there would soon be an AI that could write text brilliantly and produce photos and videos without being able to distinguish them from reality, I would have been laughed out loud in front of this audience.

It is the self-image of the creative that they cannot be replaced by the mechanistic. Excitingly, I even agree with this in essence. But you have to understand what is really “creative” in the creative process and what is simply the application of beliefs, smoke and mirrors and dusty pseudo-knowledge. 

The creative spark and holistically inspired inspiration will continue to define us as human beings for a long time to come. I believe that those who know how to use AI can be more creative and more effective as humans – not less.

But back to the process that uses Causal AI to distil what makes advertising successful and which specific guidelines significantly increase ROI. The methodology, which we call Creative.AI, consists of three components:

  1. Advertising test: In a survey, we measure the emotional reaction to an advertisement, the willingness to buy the product, the brand strengths and some other typical indicators of an advertising test.

  2. Profiling: Using a coding rule that we adopted from the company 601 Communications, one of our employees looks at the advertising and categorizes its content according to around 200 characteristics. These characteristics include the story structure (e.g. problem solution), the content-related stereotypical message (e.g. this brand is your friend) and advertising techniques (e.g. voiceover, expert or brand song)

  3. Causal AI: The characteristics of an advertisement cause an emotional reaction, which in turn causes brand acceptance and ultimately the desire to buy.

It turns out that Causal AI can explain the willingness to buy around three times better than using statistical modeling. Both the way we test advertising and profiling could be further improved today. However, the crucial thing about the approach is that it does not simply stop at perception, as was previously the case, but takes up the actual characteristics of the advertising.

The most exciting thing is the findings that the methodology revealed in many industry studies. Universal, recurring insights and industry-specific findings emerged.

It is universal that advertising works by making people happy. Advertising that triggers negative emotions such as anger, disgust or contempt is generally a waste of money. Yet it is easy to trigger such emotions unintentionally. For example, when a meat eater sees someone biting into a vegetarian patty (back then, these were not as meat-like as they are today), many of the meat eaters watching feel disgust. This reaction prevents any positive advertising effect.

Many advertising techniques are also universal in their effect. Here are a few examples: The voiceover technique confuses the viewer. If the consumer is supposed to learn something specific in the commercial, it is best to show a speaker who says the message in clear words into the camera. The most effective technique to make the viewer happy is the “looser trick”. This is the same trick that Dick & Doof or Tom & Jerry have been using for 100 years. There is one person in the commercial who you can laugh at because something happens to him or he is being clumsy.

However, the emotional messages of the commercial are industry-specific. For example, spirits are usually advertised in a way that emphasizes their quality or the friends in the (drinking) community. What works, however, is the promise of enjoyment. For investment products, the message “We are like a friend at your side” is the most effective. 

You can already see from this very brief description that it is by no means a blueprint for an advertisement. What emerges are guidelines that a creative person can fill in with their work. It is more of a lighthouse that ensures you are sailing safely in the right direction.

I’ve regretted ever since that the creative world doesn’t make more use of this approach. When I met my esteemed acquaintance Jon Puleston, who is Director Innovation at KANTAR, last year, I realized why this is the case. Kantar itself runs an advertising testing methodology called LINK and has a large department of data scientists. They have now developed a method they call LINK.AI. They use deep learning AI systems to automatically categorize advertising into several hundred properties. An AI model is then trained on the basis of a huge database of past advertising tests. With this AI model, Kantar can now predict the results of advertising tests quite well, making the advertising test superfluous. 

This product sells like hotcakes. It sells so well that Kantar has introduced a few advertising tests as mandatory for all those who want to use the forecast.

When I heard about this, I realized where the need actually lies. Decision-makers want a forecast. They want to know whether an advertisement is good or not. Kantar delivers that. What it doesn’t do is describe what would increase success instead. This kind of thing is not (yet) so easy to sell. But who knows how things will develop? We’ll come back to this in the context of Gen AI.

Marketing Strategy

After several successful Causal AI projects at Deutsche Telekom, I received an email from the US subsidiary T-Mobile USA in the summer of 2013. David was the Insights Director at the time and explained to me that the new brand and product strategy was working wonders and driving growth. What was giving them a headache, however, was that T-Mobile didn’t know what exactly was attracting customers. Was it the then innovative flat rate? Was it the fact that T-Mobile had completely removed the contract lock-in? Or was it the fact that an iPhone was still available for USD 0?

We got the chance to take a closer look at the extensive brand tracking data to better understand what drives customer behavior. On the day I presented the results, I was a little unsure how satisfied David would be with the depth of insight. I didn’t realize that the results would have a significant impact on the future of the company and the group as a whole. Before Christmas, I received this thank you email from David:

What had happened? Our analysis had shown that none of the presumed success drivers were a direct cause of the growth. Instead, it turned out that the company’s positioning as the “Robin Hood” of the industry was the central lever for success. The innovations “no contract commitment”, “flat rate”, “good low-cost devices” were the perfect justification that made the positioning credible. They had an indirect effect and enhanced the positioning. Each of the components could be copied. The Robin Hood status was not. So the brand decided to develop a continuous stream of unusual features and launch new ones every quarter. One example was the “Free Global Roaming” feature. The strategy became known as “Uncarrier Moves”.

The strategy worked. T-Mobile grew year on year and took over its competitor Sprint seven years later. Today, the brand has grown from a small, loss-making provider with inferior mobile networks to become the market-leading, highly profitable mobile communications company in the USA. 

In 2022, I met Tim Höttges, CEO of Deutsche Telekom, to attend his keynote speech (picture below). What he showed there was the icing on the cake.  The parent company has essentially risen from the seventh largest telecommunications group in the world to number one in the world thanks to the development of T-Mobile USA.

It’s amazing what the findings of a Causal AI analysis can achieve. Today I think to myself: We should have invested our fee in T-Mobile shares. But ok, what’s better than seeing your own work put into practice. 😉

It is a prime example of how marketing strategies should be founded. I had already shown similar examples in the previous chapter, such as Kindernothilfe. The examples show that it is worthwhile for every industry to shed light on its own understanding of the market in a different way. 

For the beer market, for example, we found that it is not the quality or the special taste, but the promise of refreshment that makes up the basic benefit of a Pilsner beer. We also found that the purchase of body cleansing products (shower gel, etc.) is most sustainably influenced by the fragrance experience. These are both plausible findings, but no-one had previously argued for them in this way.

There are probably thousands of methods and ways to understand what drives the buyer of a product category. Many of them complement each other, some are more useful than others. What most of them have in common is that they deliver plausible results. Therein lies the danger. Because people tend to use plausibility as an indicator of truth. But plausibility merely expresses whether something is congruent with existing prior knowledge. That doesn’t really help if you want to learn something new.

There are many methods that claim to reveal the “why” behind customer behavior. Very few of them have the scientific ideal of causality in mind. Therefore, there is no perceived lack of “why” explanations. Our brain generates these automatically one way or another. It’s like sport. The fans always know why things aren’t going well. That’s just how our brain works.

However, I hope these examples provide a few indications that it is worth approaching the “why” in a discoverable but measurable way. The advantages of Causal AI for the marketing strategy can be summarized as follows:

  • Holistic: The existing expert knowledge is used to set up the model optimally instead of merely challenging the results as before.

  • Fills a gap: There is a methodological gap between qualitative research and quantitative modeling (which takes a confirmatory approach). Causal AI fills this gap by allowing quantitative but unknown areas to be explored in a discovery-based but knowledge-based way.

  • Networked: Causal AI inherently models the networking of all system variables instead of assuming a reductionist input-output relationship. In this way, causal effects are measured holistically instead of only showing partial effects as usual.

  • Complexity-affirming: Causal AI can (if set up correctly) reveal unknown non-linearities and unknown interactions of any kind and thus enable a deeper, more qualitative understanding.

  • Validatable: In contrast to qualitative research approaches, validity is measurable, reproducible and therefore comparable.

The greatest leverage lies in explaining the reasons for success and failure. Causal AI delivers this “why” like no other methodology. Unfortunately, many decision-makers do not realize that plausibility does not have much to do with quality. As a result, there is no urgent demand for answers to the “why” in many companies. Because “answers” are a dime a dozen. Whether they are valid is another matter. Decision-makers who can separate the wheat from the chaff here have a strategic advantage, as the examples have hopefully shown.

Decide Better with Causal AI

When it comes to operational decisions, the tables turn. Here, the success of the decision can often be measured promptly. When AI became fashionable again in the 2010s after another “AI winter”, it was almost always about predictive analytics. It’s about using AI to make better and more flexible operational decisions. 

What is meant by operational decisions? In marketing, we can distinguish between two types of decision: On the one hand, decisions on the design of communication, product, packaging and distribution. These decisions affect all customers. On the other hand, there are customer-specific decisions that can be found in direct marketing, sales, customer service and customer management.

Better Decisions on Communication, Products & Packaging

Which advertising idea should we choose? Which new product concept should we launch? Which packaging design will be best received? Which price maximizes profit?

These types of decisions are evaluated in particular through test surveys. AI can be used to predict the outcome of the survey. The use of causal AI can make such a model more stable, so that the predictions are actually more valid and more likely to come true.

We have already touched on a few examples of this.

Such as Kantar’s “Link” model, which predicts the results of an advertising test. With the help of Causal AI, such a model could achieve better performance in future predictions. Causal AI models suffer less from model drift because they not only interpret the characteristics of the advertising causally better, but also integrate or control confounders.

Or the new product forecasting model that we built for Mintel. It enables an initial screening of new products before the actual launch. With a success rate of 81%, the model is very useful.

Deep learning systems (providers such as aimpower or neurons) have already been established for testing advertisements and videos in particular, which use image information to predict the results of real eye tracking studies. It is already possible to predict with 90% accuracy where consumers will look at images and how long they will do so. Of course, there are some borderline areas where this does not yet work so well. For example, AI knows that people like to look at faces, but has not learned that dog owners look dogs in the eye in the same way. Another example is the “black square” image. People look at the edges, while the AI assumes the gaze is in the center. These examples show that this AI lacks the contextual information that a causal AI would ideally take into account. Causal AI is expected to make further developments in this area in the future.

Better Decisions on Personalized Actions

I had already touched on this example above under the topic of confounders: A local telephone service provider had the issue of cancellation rates high on its agenda. Measures were developed and implemented – with moderate success. One measure was the so-called “cuddle” calls. Customers were called at home to ask whether they were still satisfied. The analysis terrified the management. Customers with whom a cuddle call was made were twice as likely to cancel as all others. The conclusion was that the call was waking “sleeping dogs”.

The problem was big enough that we were called in to set up a cancellation forecasting model. 

Example: Churn prevention

The procedure is quickly described. In the first step, the customer data is prepared in such a way that the characteristics and behavioral data of a customer that were current one year ago are used as predictor variables. The target variable is whether the customer canceled in the following period. The modeling should learn to predict this cancellation. The AI can use thousands of customer data records to create a forecasting model.

The churn scoring of all customers themselves then takes place with the trained AI model. To do this, the same property variables are only updated to the current status. The model was trained with properties from a year ago. The AI model uses this to calculate a pseudo cancellation probability – a value between 0 and 1. For various reasons, this value is not a “real” probability. 

Because the value does not represent a “real” probability, it must first be determined at which threshold value an action should be triggered that attempts to prevent churn.

As outlined in the first chapter, every churn prediction model makes two types of error: false alarm or missed opportunity. In the case of a false alarm, churn is predicted even though it will not happen. This means that the costs for the campaign were spent in vain. In the case of “missed opportunity”, the model fails to predict churn. The customer is lost.

If we set the threshold at just 0.1, we will consider most customers to be cancelers. This will minimize the error of the missed opportunity, but the false alarm error will be enormous.

If we set the threshold value to 0.9, we only have a few false alarms. However, the number of missed opportunities increases.

The optimum threshold value can only be determined if I know what a campaign costs, the probability of it preventing a cancellation and the customer value that is lost if a customer cancels. So by simulating the return on investment for all possible threshold values, we can optimize it.

Campaign management

Back to the example of the telco company. The question was how effective the cuddle calls actually were.

The properties of the model can be revealed using the simulation techniques described in the chapter above. In this example, we were able to show that the cuddle call was indeed useful. The calls were increasingly answered by people who were at home during the day. It turned out that these people had a higher cancellation rate per se. Therefore, these people canceled more often than others. However, among those who would have accepted a call, fewer of those who accepted the call canceled.

How does the evaluation of measures (campaign management) normally take place today? Just like in our telco company. The example makes it clear that this type of control works by chance at best.

Clean campaign control requires causal AI. This not only makes it possible to measure how much a campaign reduces the probability of cancellation. We can carry out this simulation for each individual customer. The effect may be greater for some customers than for others. Especially if the number of customers above the threshold value is greater than the budget would allow. In this case, we may select those customers for whom the planned campaign is particularly effective.

This shows that, ideally, a forecast should not be considered independently of the intended actions. This is because the aim of the cancellation forecast is to carry out actions that avoid cancellation.

Customer Lifetime Value

Customer lifetime value has been a hot topic in marketing literature for thirty years, but hardly any company goes beyond a current sales or margin analysis. The reason for this is a lack of information about how likely the customer is to remain a customer in the future or how strongly they themselves recommend the company each year.

A practicable customer lifetime value can be calculated if a number of cancellation forecast models are set up in parallel. One that predicts cancellation in one year. Another for the second year. Another for the third year. And so on.

Alternatively, a forecast model can also predict the sales or margin contribution that the customer will make for each of the future years.

The future years must then be discounted using the opportunity interest rate. The sum of all these values gives the customer value.

This customer value is then in turn useful in many applications. In particular, it is needed to find the optimum threshold value for the cancellation model in order to correctly factor in the opportunity costs of the “missed opportunity”.

The advantages of Causal AI for controlling operational decisions can be summarized as follows:

  • Stable over time: Cancellation models in particular are trained with data that is one year old. Causal AI models are more stable over time and suffer less from model drift. Therefore, the current predictions are more likely to be accurate.
  • Suitable for campaign control: The question of the influence of an action on the target value is a causal question that Causal AI can answer better. This means that the predictions of the consequences of an action are also more valid.

Generate better with Causal AI

Generative AI is so much at the center of interest that people often use the term “artificial intelligence” to actually mean generative AI. The fields of application are very diverse. The potential is only just being explored. 

However, this sounding out reveals a peculiarity of us humans. We generate marketing text and we generate advertising images and assess the results with the help of our subjective impressions. Is the result plausible? Does it make a “good impression”? While this way of judging is justified, it has many blind spots.

Anyone who has played around with ChatGPT has experienced that the results change significantly depending on how you “prompt” the system. Can we use Causal AI to find out how we need to prompt in order to generate marketing material that works better?

Generate Better Ads

Dr. Steffen Schmidt is a marketing researcher like no other. He is constantly testing and combining the latest tools. When he sat among the participants at our first Causal AI seminar in 2009, this was immediately clear to me. Since then, he has also been one of the pioneers in the application of the technology in marketing research. In 2023, he skillfully combined Causal AI with various Gen AI tools and massively improved the impact of Samsung Social Media Ads. This is how he proceeded:

Step 1 – Causal AI: He conducted a survey of smartphone users and measured which brand archetypes customers associate with the respective brands. He fed the data into a Causal AI model that was designed to determine which brand archetypes increase the willingness to buy a smartphone. The strongest factor was the “Explorer”, although (or precisely because) most brands are not perceived as Explorers.

Step 2 -Gen AI: He then asked ChatGPT to write a prompt for Midjourney that is designed to produce a social media ad for SAMSUNG that expresses the Explorer archetype. Without choosing further, he took Midjourney’s suggestion. He generated the slogan using the Neuroflash app – an application designed to generate advertising copy. The resulting slogan was “The freedom to go further”.

Step 3 – Test: With the InContext market research solution, respondents experience Instagram, Facebook, TikTok or Amazon as if they were visiting them in real life. The solution uses a clone of the website and can therefore replace the advertising at will. After one minute, a survey that takes reaction times into account measures the willingness to buy the Samsung brand. The comparison with conventional advertising showed an 18% higher market share. 

The result is amazing. The creation process did not require any advertising expertise. A very standardizable process of Causal AI and Gen AI was used in combination. Without further optimization, the result was an increase in performance that would have required a great deal of work, variants and test loops under conventional circumstances and would never have been achieved due to budget restrictions.

The example is so powerful because market researchers are not sufficiently successful in briefing creative people in such a way that they implement the results as intended. 

I have experienced this myself several times. Creative agencies often have a completely different understanding of what a marketing researcher recommends. What is an Explorer Archetype? How do you visualize it? Qualitative summaries often offer so much room for interpretation that the effect is lost.

Gen AI offers a standardizable and validatable interpretation mechanism. This is exactly what the manual process of a creative agency cannot provide.

While Causal AI is the more valid method of generating insights, Gen AI is the more valid method of creatively translating these into marketing material. Will humans become superfluous? Not at all! Humans are taking the wheel. With their expertise and wisdom, they orchestrate both Causal AI and the activation of Gen AI.

Marketing professionals quickly realized that Gen AI may be able to produce beautiful ads, but that they lack recognition because the design brand assets are not used. There are solutions for this and other weaknesses. For example, the addition in Midjourney –sref “URL” can give the AI an image and color language to use.

Where Causal AI consists of AI, filter algorithms and specialist knowledge, a functioning creative AI system consists of GenAI, Causal AI and specialist knowledge.

Creative AI = Gen AI + Causal AI + Expert knowledge

Generate Better Copy

The effectiveness of emails, letters and websites is largely determined by effective texts. On the one hand, this is about the right topics and arguments and, on the other, about appropriate tonality and metaphors.

To make our own emails more effective at Success Drivers, we experimented with the combination of Causal AI and LLMs in 2020. These already existed in the form of GPT2, among others. Our process should increase the open rates of our emails by no less than 500%.

Step 1 – Expert Judgement – we designed 50 variants for subject lines and had them evaluated by experts. We tested the best and worst variants in real mailings. As a result, the variants rated as poor were better than the supposedly good ones. 

Step 2 – Trial & error: We then got creative and tested the wildest variants in small sample sizes in real trials. The resulting opening rates were to serve as the target variable for the following optimization step.

Step 3 – Optimization: In the first step, the subject lines were broken down into basic associations and categorized using natural language processing. This categorization now served as input data for Causal AI. This revealed that a dominant language encouraged addressees to open. We then promoted the LLMs (with the help of the Neuroflash tool, which already existed at the time) so that it suggested subject lines that were dominant and a maximum of five words long. The subject line “straight to the point” turned out to be a direct hit. Previously impossible open rates of 54% became reality.

The same process is used to optimize newsletter headlines, texts, mailings or website headlines. A set of as many examples as possible is needed from which the AI can learn. The number of variants is more important than the sample size that is collected for each variant. For example, we only sent 50 emails per subject line and were thus able to test 40 variants simultaneously with 2000 addressees.

ChatGPT is known for writing great texts and responding to the user’s requests. But how does the user know whether their queries lead to effective variants? A Causal AI-driven cognitive process provides precisely this evidence.

Sounds complicated? Yes, it is elaborate. But would you like to calculate the profit contribution if a campaign only improves by 10%? How much more profit will you receive in absolute terms? It will probably exceed ten times the investment in a short space of time. It quickly becomes clear that these AI-based processes pay for themselves very quickly.

The benefits of Causal AI

Causal AI helps to become more effective in all areas of marketing. It is a kind of upgrade for AI and makes insights, decisions and creation more effective.

Better insights

In marketing mix modeling, Causal AI helps to cope analytically with the increasingly correlated channels. Their influence is determined more truthfully and at the same time the indirect effects, which can be found in long-term effects among other things, are also taken into account.

The drivers of the customer experience can be better understood. The application of Causal AI overcomes the fallacy that the topics frequently mentioned by customers are the most relevant. It enables a simulation of the effect that an improvement in the customer experience will deliver financially.

Causal AI provides innovations and product optimization with new methodological possibilities to reveal which product features and barriers play a central role in promoting purchase acceptance and willingness to pay. You are no longer limited by a small number of conjoint features, nor by the explicit query methodology of MaxDiff.

The impact of communication and advertising can be better understood through Causal AI. This in-depth understanding goes far beyond the aspects that are queried in an advertising test. It is insightful to understand which emotions are “deadly” or what contribution brand building has. It becomes particularly useful when the analysis can provide recommendations for the storyline, the emotional message, the choice of music or the type of humor. These are all tried and tested methods.

A marketing strategy requires effective positioning, target group selection and segmentation. Causal AI can provide insights based on suitable surveys to deliver effective brand positioning both in terms of content and associations/emotions and help to understand which customer types have an affinity for the product category.

Better decisions

Causal AI is also used to create forecasts. These models are more stable and suffer less from model drift. The risk of discriminating against minorities is also reduced.

When it comes to the selection of products, advertisements or packaging design, a Causal AI Model can in some cases avoid testing with the help of market research or a trial test. This allows decisions to be made more quickly and cost-effectively.

When it comes to direct marketing or customer service campaigns, an individual decision can be made for each customer and target customer as to which marketing campaign should be used. Causal AI can optimize this decision. Examples of this are the reduction of cancellation rates, the estimation of customer potential or customer lifetime value.

Better creation

Generative AI supports marketing in the creation of images, videos and texts. The contribution that Causal AI can make is to ensure that among the almost infinite possibilities for prompting AI, a variant is found that is most likely to be highly effective.

More effective advertising motifs can be generated by feeding Causal AI on the basis of suitable market research, which then identifies which content-related and emotional-associative characteristics the image must express in order to be effective.

Effective messaging is needed on websites, in subject lines or mailing headlines as well as in slogans. This can be optimized with Causal AI by trying out many different variants in real experiments. By breaking down the texts into their associative factors, Causal AI can identify the hidden characteristics of effective texts. With this knowledge, Gen AI can then be used to generate new, more effective texts.


Benefits = Better insights + Better decisions + Better creation

Success Drivers GmbH
Johann-Heinrich-Platz 4
50935 Cologne

Success Drivers GmbH holds the copyright for any material, included translations, published at