Customer service. Every day, over 30,000 customers call a call center when they encounter a problem. It takes 5 minutes on average to connect a client to an operator, listen to a problem, switch to a specialist and listen to the problem again. At the same time, an organization collects lots of data: transactions, operations, communications, online activity on the customer level until the moment of the call. In 80% cases the reason of a call can be predicted from data. It is required to predict the reason for the call within 5 seconds to reduce the waiting time and improve the quality of service.

Results The waiting time has been reduced from 5 minutes to 3 seconds. The savings in the work time of call-center operators amounted to more than 800 hours for every 10,000 calls. An infrastructure has been created to monitor the quality of customer service in real time, predict and respond to problems.

https://vc.ru/services/353119-prinyatie-resheniy-so-skorostyu-potokov-sobytiy-na-zhivom-primere

Quality control. Every 90 seconds, 1 unit of finished product leaves the conveyor. The technological process is complex. Quality control takes more than 1 day and is carried out after the fact. High risk of low-quality output during the day. In 90 seconds, it is required to: collect and process data from different sensors and systems, make a prediction about the quality, determine the place in the process where the quality is deteriorating.

Results Prevention of the risk of producing 1000 low-quality finished goods and losing 1 day of production downtime.

Online scoring. Credit conveyor. A bank processes more than 1 million applications every day. 1% increase in delinquency due to algorithms losing predictive power is costing millions of dollars per month. It is required to reduce the time for detecting the loss of predictive power by models from 1 month to 1 day, to increase the predictive power of models, and to reduce time to market of new models from 3 months to 1 week.

Results Loss of predictive power of algorithms is detected in real time. The anomaly detection in data quality is provided online. Time to market of the new versions of algorithms has been reduced to 1 day. The lifecycle of algorithms has been reduced from 3 months to 10 days. The increase in the predictive power of algorithms is about 30%, which led to an increase in approval rate at a given level of delay.

https://moskva.mts.ru/about/media-centr/soobshheniya-kompanii/novosti-mts-v-rossii-i-mire/2020-07-24/mts-bank-v-tri-raza-sokratil-vremya-rascheta-ocenki-kreditosposobnosti

Personalization. A visitor wants to place an order. Each visitor has his own preferences: dishes, ingredients, calories, drinks. Placing an order takes up to 5 minutes. This leads to queues, congestion of visitors and dissatisfaction with the speed of service. It is required to reduce checkout time by personalizing recommendations.

Results Data streams from cash desks, mobile applications, wi-fi points are connected. For each visitor, a history of orders, changes in preferences is accumulated. Trends are continuously predicted taking into account the time of day, day of the week, location of the point of sale, actual weather. At the moment of opening the application, a personal recommendation is calculated for the visitor, taking into account personal preferences, trends and the current context. Formation of the recommendation takes 1 second. The checkout time is reduced from 5 to 2 minutes. +4% average check growth due to referrals.

Demand forecasting. Assortment management. A pharmacy network has 2,000+ points of sale and 3,000 SKUs. Demand at each location is different and changes rapidly over time. The lack of some SKUs in the assortment and in stock, as well as an excess of stocks, lead to a decrease in the efficiency of the network.

Results The system is connected to data streams from payment terminals in real time. The facts of lack of drugs are recorded. For each SKU, a demand forecast for 1, 3, 7 days is built every 5 minutes. An assortment optimization is calculated continuously. An auto-order for the supply of drugs is being formed. Inventory optimization allowed to release working capital by 11%. The expansion of the assortment, taking into account local demand, allowed to increase sales by 6%.

Fraud prevention. The bank processes about 10,000 transactions every second. Preparation for the implementation of fraudulent operations is manifested in various information systems. It is required to combine data from different systems, continuously detect behavior that has signs of preparation, block risky transactions in real time, and manage business logic without the involvement of developers.

Results 96% of fraudulent transactions are blocked. Configuring business logic of the new patterns doesn’t required programming and takes minutes.

News categorization. A bank receives about 500,000 news daily in Russian and English regarding corporate clients. It is required to automate the processing of the news flow in real time; identify organizations, people, locations; categorize news by given topics; match organizations with customers; predict the degree of importance of news for customers; integrate the solution with other bank systems.

Results The categorization accuracy exceeded 75%. The news categorization takes 2 seconds. The commissioning time for new AI algorithms has been reduced to 1 hour. An infrastructure has been created to allow the internal data team to continuously improve the accuracy of predictions and reduce commissioning time.

Scoring conveyor. The customer scoring is based on the analysis of data from dozens of internal and external systems. The customer assessment process includes both fully automated and expert procedures.

Results The system allowed to reduce the time of data processing from 1 day to online. The time to make of the conveyor changes has been reduced from 1 month to 1 day.