In the Code – Better customer understanding thanks to AI

Imagine if you could predict when your customers will order certain services to run their business. You would be much more accurate and faster than the competition. Resource planning would be easier and more targeted and you would have a reliable perspective on the future.

You could better address and optimize specific topics such as :

• Refining existing segmentation
• Planning and executing campaigns for specific customer segments
• Preparing and deploying your own resources.

At an early state the project showed how effective a combination of CRM, Big Data and AI can be.

Companies must continuously personalize and optimize their offers and processes to build and maintain a strong relationship with their customers. CRM (Customer Relationship Management) systems, which contain all of the company's knowledge about its customers support personalization of offers.

With the advent of artificial intelligence, large amounts of data, stored in in-house systems (e.g., CRM, ERP) and enriched with publicly available data (e.g., weather forecasts, etc.), can be used by a company in a customer-oriented strategy. But note that at the same time you must ensure your customer relationship is lasting and that customers continue to entrust you with their data. You need to credibly and transparently ensure that cyber risk is minimal and that you comply with regulations (e.g.: DSGVO/GDPR).

Big data analysis and artificial intelligence are already being used successfully to refine customer segments, predict the needs of these customer groups, determine the optimal time to make a purchase and to plan customer service personnel. These examples show the economic relevance of the correct application of AI and other technologies.

An accurate analysis of customer needs as a starting point is important. Let's say you are a retailer in a perhaps stagnating market. If you can place your offer at the right time, you will be more successful. Precisely predicting when your customers will decide to buy is a big competitive advantage. You can, for example, increase market share.

In the mentioned project, ELCA analyzed historical ERP transactions and publicly available data to determine the next shopping window for specific products for each customer. Since the retailer had customer permission to use the data and a functioning distribution system was also available, we used state-of-the-art technology and algorithms to do this.

The matter was addressed with a classification rather than a regression problem approach. For our customer, it was not the amount (cost, quantity) of the next order that was of primary importance, but the next possible order time, so that the customer was contacted at the right time to support him in his purchase decision. The model therefore had to divide customers into different classes. Such a classification could look like this:

• 0: Customers buying in the next x days
• 1: Customers buying only after x days, and within one year.
• 2: Customers not making a purchase within one year.

The various data sets were then merged to create useful information for the machine learning model. This combination of customer, contact and product model data forms a first set of features defined as account-related features.

Each row in the final machine learning data consists of three types of features: the account-related features related to the current month, and the features of the previous order. Functions related to the current month are mainly built from external data to link price and weather temperature fluctuations over time. Finally, some features of the previous order are used, such as "Was it a large or a small order?" or "When did this process occur compared to the previous order?

The result was an extensive list of 122 functions. These functions are based on the business knowledge accumulated in CRM, but also on analyses of customer journeys. A customer journey is a chart that shows all orders from a particular customer, as well as weather and pricing information over time. Customer journeys were used as a starting point to understand the pattern of customer orders, the impact of external data, and the general behavior of customers. Other functions were also created after a trial and error phase in creating and testing machine learning models.

Various machine learning models were used to compare the quality of the results:

• Linear models (linear regression)
• Tree models (extra tree and decision tree)
• Ensemble models (random forest, extra trees, and bagging regressor)
• Neighbors models (neighbors regressor) Neural networks/long short-term memory networks

Just a few exchanges between experts and data scientists led to an exact model and misinterpretations could be avoided.

The example shows that data quality, process efficiency and the systems of an organizational unit determine which approach makes the most sense. Often, already known measures and tools such as an effective sales system (CRM) or the clever use of available data (e.g., big data) solve the most important problems.

However, if you want to benefit from the most reliable prediction or discover new business areas, you need to use artificial intelligence (AI). You also need to be aware that the demands on the organization are more complex. The above example shows that new technologies such as AI complement a marketer's toolbox. This does not change the fact that the basis for a successful business relationship is still the trust of the customer. In order not to jeopardize this trust, data protection and cyber security must have priority and be communicated accordingly.