How to Increase Average Purchase Amount Using AI

It’s hard to find an online retailer that has never used AI-based solutions, either ready-made or self-engineered ones. ML models help to automate processes, improve marketing effectiveness, and push up sales. How can you increase the average purchase amount using smart tools? Customization in customer service is a phenomenon that is highlighted in all lists of e-commerce trends: it is a must-have for any business that sells online. Before personalization tools, stores used to interact with customers largely at random: people often saw ads for products they were not interested in. For instance, a user who ordered mountain skis would immediately receive a notification about a sales promotion of snowboard boots and hockey sticks. To avoid such situations, there was no choice but to manually process data on each client. But if the store had several thousands of such customers, the task became almost impossible.Nowadays, thanks to automatic customization, we can predict customer needs without complex data manipulation. This is done by analyzing the user's purchase history, viewed pages, and search requests. Based on this information, the system selects appropriate products for advertising: either similar or complementary. From this perspective, if a consumer has just purchased a wine decanter from you, there is no point in showing them other decanters from the range; it is more logical to offer a set of glasses with a discount. During the analysis process, AI learns on its own: for example, if users often buy earphones or a power bank together with a smartphone, the system remembers this pattern and starts offering these accessories to other users who show interest in phones. Customization channels can vary, from tooltips on a website to Facebook ads, and there are a huge number of services available to set up any of them. For websites, these are, for example, RetailRocket, Mindbox, Convead, and REES46, and for email marketing, these are tools like MailChimp and SendPulse that can create personalized emails. If a website visitor has viewed goods for renovations, they will receive an email about a discount for a set of several items from this category. If the user has "forgotten" an item in their shopping cart, they will receive a corresponding reminder message.

"Immersion" in the client: from habits to a daily routine

The next step in communication with the customer is to take into account not only their preferences but also their daily routine, habits, and other characteristics. If a user usually looks through emails in the morning, then sending a notification about a promotion during these hours is critical for encouraging a purchase. The development of customer service technologies and fierce competition dictate the need to know your customer better than yourself. Detailed information about the customer will help you choose the right discount, which won’t exceed the discount amount a user is willing to accept. In other words, it makes no sense to offer the user a 20% discount on sneakers if they are ready to buy them with a 10% discount. The system, based on the LTV forecast (the total profit that a client made you) and the repeat purchase probability, divides your website visitors into segments, each of which has its own advertising and email marketing strategy. And if it is very unlikely that a particular user will order something from you, there is no point in spending your advertising budget on it, and vice versa.

Pricing algorithms

The store's pricing policy is influenced by demand dynamics, seasonality, inflation, competitors' actions, external economic events, and hundreds of other factors. You need a whole staff of analysts or just algorithms based on AI to take all factors into account. Amazon, Walmart, DNS-shop, and other retail giants are setting the prices on their products using machine algorithms that constantly analyze the market situation in order to recommend optimal prices, sales’ periods, the need for promotions, and so on. In addition, AI can forecast the supply and demand and suggest the best response to the expected changes. This is confirmed by the Competera platform case with a British retailer: the online store had been testing the price recommendation service on its platforms for a month. During that month, due to the season and the industry, the retailer's sales declined, but the algorithm predicted the drop and suggested an optimal pricing strategy. The result was really demonstrable: while the number of transactions in the control group fell by 17.2%, in the test group, it increased by 5.1%. Now machine price optimization is mainly used by large retailers or high-margin businesses: the more revenue an enterprise receives, the more noticeable the effect of "smart" pricing is. Moreover, such solutions are usually not cheap and require having a large database from the store. But with technology development, there is a gradual democratization of products, so it is likely that in a few years automatic price adjustment will become as common a practice for stores as CRM is today.

Widgets: smart and not so smart

Have you ever faced a situation when a couple of days after ordering a product you received an email with a promo code for purchasing the same product? The effectiveness of such late discounts is surely tending towards zero. In order to show a customer a special offer at the right time, stores use widgets that are generated for each page with a promotion. According to our statistics, requests via widgets on websites account for about 20% of all leads and, in some areas, this figure reaches 40%. But in this case, you should keep in mind that it is possible to obtain these numbers only if the widgets are properly configured. A sure way to make the user leave your website quickly lies in using dialog windows that constantly pop up, blocking the content that they really want to see. Stores use so-called "smart" widgets based on AI to avoid this. Their algorithms analyze the user's behavior during the session in order to calculate the optimal moment to demonstrate the offer and not cause a negative reaction.

"Your call is very important to us…”

"Unfortunately, all operators are busy at the moment. Please stay on the line. Your call is very important to us... " - you will have heard this when trying to contact a company. If a customer is not calling to return a product, then most likely, after these words, the person will hang up and open another store’s website. But since it is impossible to guarantee that the operators won't be busy (unless, of course, you hire a lot of employees), the optimal solution is to use a callback service. This service allows the customer to talk directly to a real person instead of having a long wait on the phone. Of course, this only works if they actually get a callback. Callback service difficulties are not so much related to the employees' organization skills but to the algorithm of their telecom equipment: it "does not see" the difference between the operator's voice and the answering machine message. To put it another way, if the client listened to the Interactive Voice Response (IVR) Greeting, the system will display such a call as successful, even if the person did not wait till the specialist responded. Solutions with DTMF support help you cope with this situation. For example, you can integrate an algorithm into a callback service that analyzes the audio track and determines which exact sounds are on the phone: beeps, automatic greetings, or the operator's voice. Based on this factor, the call is marked either as successful or not, and in the second case, the specialist is asked to contact the client. Data processing is something that a human will never be able to do as well as a machine. AI is one of the tools that allows you to analyze data more efficiently than your competitors. This is why, here at Andersen, we are actively using AI in our projects and successfully integrating AI algorithms into our clients' applications. Zhann Chubukov, Head of Data Science, Andersen: The role of human specialists in an online store may be insignificant compared to the role of automation quality and scale. I was involved in a project for a media holding with offices in 26 countries to develop an automated pipeline of recommendations for the subscription packages for customers. As a result, people did not work with the data or contact the customer at any step of the sale. A complex mechanism based on a combination of algorithmic steps and the use of ML models did all the work.