people with cellphones

6 steps to increase online sales and minimize unpaid transactions rate (on the example of a telecommunication company)

How to increase sales in your online store? A lot of articles have already been written about SEO strategy or UX best practices – so today we are not talking about that.

We are going to focus on using artificial intelligence to automatically determine the risk of unpaid transactions and predict your customer’s behavior. Using TeleCom XYZ as an example, we will examine how to use machine learning to boost your online business profits in just 6 steps and with 1 piece of software – while it’s still the perfect time to do it.

Powerful e-commerce revenue growth: $111 bln in 2 years

The good news: despite the undeniable impact of the pandemic, the European e-commerce market is not (yet!) saturated. While a McKinsey report indicates that in 90 days the U.S. has seen e-commerce growth on par with the last 10 years…


E-commerce growth in the USA, source.

…In February this year, Statista summarized the situation in Europe as follows: “Recent forecasts are positive that the European e-commerce market will continue its fast and dynamic climb (…) in the next few years. This comes as no surprise as across the EU the share of e-commerce users is as high as 80 percent in the most mature e-commerce markets.”

Confirmation of these words is the report published in May 2021 – highlighting the growth of revenue from online sales channels in Europe by more than $111 billion in just 2 years (2019-2021).


Retail e-commerce revenue forecast 2017-2025, source.

It’s no wonder that companies are focusing on intensifying their online sales, but there is a real risk of fraudulent transactions. The question is: how to prevent such situations, even despite the galloping growth of popularity of online channels and lack of control over the flow of users? We found the answer in artificial intelligence algorithms.

How to increase online sales and minimize the risk of unpaid transactions… at the same time?

The data your customers leave behind is not just their first name or e-mail address for sending newsletters – the analysis of user behavior on the web and access to databases also provide opportunities to intensify sales. As an example, let’s take a fictional company TeleCom XYZ, which is currently developing an e-sales channel. By building and implementing appropriately defined scoring and recommendation models, this company is able to increase the number of transactions while reducing the risk of non-payment. With the help of artificial intelligence, it will achieve this goal in 6 steps while using:

  1. Anti-fraud rules during the form submission: based on the analysis of the user’s online behavior (tracking events)
  2. Anti-fraud rules after the application is submitted: on the basis of internal and external database queries;
  3. Anti-fraud rules after the application is submitted: based on the correlation between the analyzed application and historical applications (cross-checking);
  4. Analysis and mapping of feedback information from the location intelligence module: from geocoding to verifying whether a given address actually exists;
  5. Scoring models (AI) determining the risk of default and the highest transaction limit to be granted;
  6. Automatic recommendations of alternative offers to gain a loyal customer and maximize sales potential.

Let’s see these steps under a magnifying glass.

1. Anti-fraud rules as early as form filling – tracking event analysis

Everything the customer does while filling out the forms goes into the database, and after aggregation, the data is processed by a scoring engine. It takes into account not only the channel of access (type of device, IP, browser), but also all clicks on links, buttons, or even indicators of the location of the device used (e.g. whether the device is physical or virtual and how it moves while making a purchase). As a result of the collected data analysis, any defined anomalies are detected, e.g.

  • use of TOR browser;
  • pasting key data from the clipboard;
  • the way of entering data (biometrics) indicating the use of a machine (bot detection).

In the case of biometrics, into account is taken the time between keyboard strokes while typing the given characters. If each successive pair of digits or letters deviates markedly from the standard rate, the result time will be at a low level – thus indicating a bot.


Example of biometric scoring calculation, own source.

Data collected and analyzed in this way can protect TeleCom XYZ from a transaction with a high risk of fraud already at the form filling stage. The decision to stop credit verification goes hand in hand with process optimization – querying external and paid credit reports becomes redundant. Someone mentioned cost optimization? 😉

2. Anti-fraud rules after the form submission – based on internal and external database queries

If the customer has passed the stage of filling in the form with information necessary to submit an application, his/her data are subject to further assessment – in the following order:

  1. Validation of the provided information (e.g. checking the correctness of personal data, ID number);
  2. Downloading information from internal databases (list of customers with whom the operator already had a contract that ended with the cessation of payments);
  3. Downloading reports from external offices.

The occurrence of an event indicating insolvency at any stage results in the rejection of a risky credit application.

3. Anti-fraud rules after application – based on cross-check mechanism

A situation analogous to the previous one. However, what happens when basic customer data could not be found in internal databases or in reports of external offices? Cross-check rules using AI models are invoked. This means that with the help of additional identifiers (e.g. cell phone number, email, IP, cookie, device fingerprint, standardized main address, standardized delivery address) TeleCom XYZ is able to identify the same user even if the user changes the entered data. The artificial intelligence checks in this case:

  • Whether the database of previously submitted requests manages to find a trace of a person rejected for the reasons described in pt. 1 or pt. 2.


    The wife tried to apply for a new phone, but her ID number is listed as a debtor in the operator’s internal database. When the application is submitted, all the identifiers mentioned above are registered. After some time (a minute, an hour, a day…), the husband submits the application – he uses e.g. the same main address or the same device. In this situation the cross-check mechanism will link both events and the application will be rejected, and in the future this information will significantly affect the finally granted credit limit.

  • Whether there are too many requests in the database of previously submitted requests that are related (by any of the analyzed identifiers) to the currently analyzed request.


    Cross-checking rules will reject a transaction if someone tries to buy a particular phone using, say, 2 different devices but the same email address.

4. Analysis and mapping of feedback from location intelligence module

The key factor here is Location Intelligence – it allows to detect and understand the relationship between geographical space and people, events, or transactions. This is possible, among others, by displaying information in the form of a map. For this purpose, geocoding and analysis of the entered information in correlation with other business data will be used. When a customer enters data, TeleCom XYZ finds out, among other things:

  • Whether the address given actually exists;
  • If given address indicates a block of apartments but no apartment number is given;
  • Whether a non-existent apartment number is given;
  • If the main address is not a building in which no one is registered (e.g. an office, a vacant building).

Moreover, based on geocoded addresses from the entire history of contracts held by the operator, TeleCom XYZ can build and map a risk model for a given area/region.


Example of mapped scoring for Poland, own source.

The model created this way is also used in the main risk assessment model. How?

5. Scoring models to determine default risk and the highest possible transaction limit

The scoring model is built on all retrievable data, which includes:

  • Application data (including age, gender, address and associated geocoding score);
  • Detailed data from external bureau reports;
  • Behavioral data – e.g.: how the customer paid the invoice (for customers who currently or previously had a contract with the operator).

The obtained index allows for a precise ranking of customers from the least to the most risky, which can then be used to grant a specific credit limit. The worst rated customers (e.g. the worst decile according to the scoring):

  • Have the option to buy a contract without a phone (SIM card only);
  • Get limit = 0 (and a message about the possibility to buy in the “full price” formula).

The customer can purchase the phone “in installments”, “for 1 USD”, etc., but only if the value of the device is paid off in future invoices. – However, this is on the condition that the value of the device is repaid in future invoices. In case the scoring model assesses the customer as not creditworthy, the operator may allow the customer to sign a contract with full payment or a possible deposit, e.g. 800 USD.

transaction_limit transaction_limit_legend_ENG

Amount of transaction limit depending on risk level, own source.

6. Recommendation model of an alternative offer, or how not to lose a customer

So let’s assume that a new customer is granted a limit of 1000 USD, but he applied for a phone, which is valued by the operator at 1200 UDS. This situation does not mean the loss of business and potential, long-term cooperation (and profits) – in this case, it is worth to propose an alternative offer. How to do it effectively? It is enough to automate the process.

An auto-model of recommendations is activated. From the database of available phones, artificial intelligence selects the alternative with the most similar parameters, which is within the granted transaction limit (e.g. smartphone for 3800 PLN). For example, three phones will be displayed and the one that guarantees the highest margin to the operator will be highlighted. XYZ TeleCom can also promote a specific brand, if this brand was included in the customer basket.

In this way, the leaders on the online sales market are more and more willing to rely on the technological golden mean to intensify online business: meeting customer expectations while increasing the security of transactions. Especially since the implementation of artificial intelligence algorithms is a direction that not only the “biggest players” are looking at – complex artificial intelligence algorithms are no longer a glass ceiling that cannot be broken.

All 6 steps in 1 tool to safely increase online sales profits

This is because the implementation of innovative technologies does not require investment in your own IT department and infrastructure. Today, fully automating the “6 steps” process with 1 tool becomes quick and easy.

Let’s take one of Algolytics solutions as an example: although the application hides complex machine learning algorithms, its intuitive interface allows users to build predictive models on their own in just a few minutes. If necessary, support is provided by Algolytics consultant: from flexible deployment (cloud or on-premise?), through help with integration with already existing databases, to consulting during use of the tool.

And yes, the potential for increased profits from online channels applies not only to the telecommunications industry, but also to the entire e-commerce market. And yes, you can strengthen the competitiveness of your online business already. We provide you with ready-to-use technology to safely intensify your profits. Take advantage of the free trial version and… see for yourself how it works.

Pin It on Pinterest