How to carry out an Artificial Intelligence project

Massimo Fubini
Founder &CEO @ContactLab

What is Artificial Intelligence, besides a buzzword on everyone's lips? It is related with Big Data, a huge amount of apparently disconnected data that, thanks to the computational capacity of algorithms, become talking data, guiding towards the best business decisions.

But many are still unprepared for it. Like sex for teenagers, everyone is talking about it, but nobody really knows how to do it.

First off, what types of data can be useful to a company?

  • Personal data: socio-demographic data (sex, geolocation, job, etc.)
  • Purchasing events: transactional data, i.e. the records of the purchases made by customers, associated with the profiles emerged from the segmentation previously carried out.
  • Dynamic events: dynamic data referring to all the interactions that customers have had on the different digital channels. It is typically structured data, even if scattered across different systems, silos-like (Customer Care, Preference Center, Social Login, CRM, E-commerce, Retail, Analytics etc.).


What problems may arise?

  • Vertical data to be aggregated

Data alone doesn’t tell much, it’s the relationship between different data that brings value. For this reason, it is fundamental to create an intelligent relation that aggregates the various data collected: the most important one to be discovered is the cause-effect relationships. But if we multiply the data of each individual customer for all customers, we end up with a huge amount of material to analyze.

  • Customers present themselves with different identities depending on the system

Actually, we’re not talking about customers, but about identities. When a customer visits a physical store and buys, he is recognized as such. But if the same customer shows up on social media, where they go under a different name, they will be registered with another identity and you won’t be able to recognize that it’s the same profile of the customer who came to the store. A first step then should be to bring all the various identities back to a single profile.

What it takes to do that? Engagement intelligence combined with a human touch. For this to happen, you need a Customer Data Platform, which is a structured database to manage people, identities and the events related to them.

In this context, AI can support us:

1) enriches contact profiles through predictive models that finds correlations between users

2) generates dynamic clusters by identifying related behavioral patterns

3) precisely customizes content to generate in-store traffic and increase sales

But it is not just a black box to feed with data: to get results you’ll also need human intervention.


A human touch applied to technology

As Eddison used to say, Genius is 1% inspiration and 99% perspiration. This can be applied to AI: 1% represents the creative part, while 99% is pure sweating, i.e. <strong>data analysis</strong>. The human touch comes in to understand the situations, the market and the needs of a specific company. After all, data is useful only if we know how to gain a competitive edge out of it.

Let's see the real case of a customer called Alpha. It is a fashion retail with one million registered users, an average ticket of 250 euros, and an average purchase frequency of 1-2 per year. But are customers buying twice choosing products from the same season or products from different seasons? By leveraging insights, the communication with them can be improved: it's easier to convince a customer to buy again, when we know they’re already attracted to products from the same season.

The same concept can be applied to discount budget: can we predict who would buy our products regardless of the discount, and maximize efforts so that there is no loss of margin? AI and data can help us understand the price elasticity of a customer. In this way, you can allocate accurately prices and discounts, and make personalized offers.


What are the steps for an AI project?


stepsData and AI should lead to business optimization. Let's see the steps for their practical application.

  • Objective

The first thing to do is define the objective to be achieved, which must be clear, shared and achievable.

  • Data

You must ask yourself whether the data available can be used to answer the questions that may have arisen when defining the objective. If not, it must be reformulated. Data can be of various types: personal data, retail events, e-commerce events, e-mail feedback.

  • Exploratory&Cleaning

80% of the project is dedicated to exploring and cleaning data. In this phase, we might notice anomalous trends and outliers in the purchases that give us a distorted view; that could be caused by a lack of integration of the customer information flows. They must be cleaned up, because if you feed the technology with distorted data, the results will also be distorted.

  • Model building

Once the data has been cleaned, a model can be built. The question could be: which customers are at risk of abandonment, and which ones will definitely buy? In this case we could apply the model of Customer Behavior BTYD, which provides for each customer a probability of abandonment and the expected number of future transactions.

It takes into account the customer's transactions and their history, from the first purchase to the latest one, the frequency, regularity, etc. Trained in predicting a time span of one year with an accuracy of 74%, after 6 months it was able to correctly predict 83% of real buyers.

To better understand the customers who have purchased, another clustering algorithm can then be applied, to assign each customer a level of engagement on the e-mail channel. There are highly-engaged customers who follow the newsletters and click on the links, others who are less interested in them, and inactive users, who no longer follow the channel.

  • Insight

Thanks to the E-mail Engagement Index, it is possible to discover that the customers who have purchased are very active on the e-mail channel. Inactive customers mainly abandon the business. This data set can be added into the model to increase its accuracy (fine tuning).

  • Action&Measures

At this point, you can create targeted marketing strategies on users with high risk of abandonment, like:

- improve customer relationship

- upselling

- customized discounts/promotions

At this stage, the monitoring of KPIs is very important.

As for the insights collected thanks to the E-mail Engagement Index, we can understand who the future buyers are, and then engage them at the lowest cost possible, avoiding waste of media budgets, and those customers who will likely leave the business, focusing on their recovery (customer match/custom audience/DMP).

The next steps could be to predict the best content/product for each customer or the best time to contact them.