How to teach artificial intelligence to sell
Robots [have not yet] learned human behavior even in text chat rooms, although they are trying with might and main. But there is a niche for the application of artificial intelligence. Machines do not know how to talk beautifully, but on the basis of big data they already make life easier for the business by automatically selecting a specific product for a particular client. The contact center can only contact the latter and with a large (or at least more) likely to complete the sale. Moreover, with much less preliminary efforts by people.
We already figured out what needs to be done before grabbing at work with models, and how to put together an intelligent sales optimization team using big data. How now to connect business products with customers?
Choosing an AI tool
What artificial intelligence has not yet learned is how to sell without customers. We need a list of potential buyers, according to which it will work.
Suppose we have such a list. How to understand who and what to offer?
The task of prediction and solves artificial intelligence - based on historical data. We take those who in the past bought some product, and build a model on them. Then we take a list of customers who have not yet bought this product, place it in the model, the model is trained and learns to predict those who would buy.
The disadvantage of the approach is that for each product you need to analyze, buy it or not buy. That is, for each product is built its own model. If we are a bank, then there are not so many products: for example, several options of plastic cards, a finite number of loans and deposits - a total of 15-20 products for sale.
But what if we are an online store with 1000 articles? Or an online movie theater with thousands of movies? For each of them, building a separate model is, to put it mildly, expensive. Such a thing as a recommender system comes to the rescue.
Recommender systems appeared just from online cinemas. Instead of hundreds of models, a “customers-products” matrix is built. The intersection shows which customer purchased which product. Further clients are compared, similarities and differences between them are sought, as a result, voids are filled in the matrix. Suppose two users watched 3 of some kind of film. And one of the users also looked at the 4th, and the second did not. Since they are similar in previous views, the system will offer the 4th movie to the second user.
The advantage of a recommender system is that for each client the product that the person most likely buys is automatically considered. You do not need to plant a staff of data scientists who will build a model for each of the hundreds or thousands of products. We have a recommended product for everyone. That is, we have automated the process of building a model.
The recommender system is especially good in the following case. As a company, we have active and passive channels. Active - where we communicate with the client on our own initiative (call, SMS, e-mail). Passive - where the customer comes to us (site, application, ATM). If you build a model for each product, it constantly narrows the list of customers for the offer, because it optimizes efforts and selects only those to whom this product is worth offering. But we have a situation where there are customers without any single offer at all. Just because each model chose its own - and left voids. That is, a person comes through a passive channel, and we have nothing to show him. A recommender system considers an offer for each client. And - the best offer.
But there remains a small problem. Suppose we are launching a new product and we need to sell it from our nose - we have a sales plan for this month. The recommendation system will not help - it honestly works and recommends to everyone exactly what is most relevant to it. She does not consider in any way our need to sell as much as possible of a specific product and fulfill its sales plan. It turns out that in this case the recommendation system is ineffective.
Therefore, in sales based on big data, a combination of methods is used: a model for a limited set of products, a recommender system for a general one.
Apply business rule
We taught the AI to select an offer for each client. But not every optimal product makes sense to offer. Filtering results is called a business rule.
Imagine that I use a premium bank card, which costs me 2,000 rubles a month. The model, built by the bank, considered that I would optimally offer another card, non-minima with servicing 300 rubles per month. Of course, I have a tendency to take it and save. But the bank does not make sense to offer me such a product, because it will lose in revenue. Such cases need to be cut off before the sentence. A similar situation with Internet service providers and telecom operators.
Therefore, a business rule is imposed on machine learning recommendations. So the client receives the relevant offer, and we do not drop revenue.
Selecting the channel offers
So, there is a client and filtered products that are optimal for him and us. It is necessary to calculate how much it will cost us to offer this product to the client. And is it worth it.
For example, a call is one of the most expensive options. If the product is high margin and the probability of its purchase is high, we can call without hesitation. If the product is low-margin or the probability of a purchase is very low, we will spend more time and money on notifying the customer than we earn from the sale. Then it is better to write an e-mail or SMS.
Some proposals do not make sense to drive through the active channel at all - it is more profitable to do nothing and wait until the client comes. For example, post products at an ATM or website. This money is not worth much, but there will be some kind of conversion from them.
Regarding the base of potential buyers. At the very beginning, we proceeded from the fact that we have a list of clients. It can be own and external. For example, we can broadcast offers of new products to existing customers, making the so-called cross-selling. We work freely with our base: we build models, we distribute customers into segments, we increase the average bill.
In the case of an external base, all the steps mentioned fall on a third-party partner. After all, firstly, none of the external sources will not give the data in a pure form. Secondly, in most countries it is legally prohibited. Therefore, in working with third-party databases, a method such as look alike - “find similar ones” is often used. Namely: a small sample of our own existing clients is taken, to which our offer fits, and their list is transferred in an impersonal form to the owner of the external base. He builds his model, selects the clients we need and shows them advertising.
Total, if we consider the whole cycle
- recommender system and models are taken;
- all of them are locked in the so-called business rules engines - an environment where business rules are applied;
- the results are locked into a system that optimizes channelization
At the exit, we get integrated communication with the client in terms of sales, where for each of them the optimal product and the optimal delivery channel are determined.
Yes, at the very beginning it is necessary to invest in the construction of the process. But then the costs of people are minimal. In contrast to the standard CRM, where people constantly invent campaigns, build models for them, make manual samples, load channels, and so on.
And we must not forget that no advanced methods of machine learning will help if the business is not ready to restructure business processes. A lot depends on the “last mile” contact center that works with machine learning results and reaches customers. Big data is not a panacea, but a good help - if used wisely.
Want to learn how to implement Big Data technologies - come to our course for analysts , starting on January 28th or to one of our corporate programs .
The post was prepared by the School of Data on the basis of the publication of the founder of the School in the Business HUB of PJSC "Kyivstar"