5 February, 2020
Even when artificial intelligence has not yet become an obvious trend, the Marketing Logic team has implemented these technologies in one of the largest banks of Russia
The analytical company and the Bank together began to build a full-fledged business process management system using AI first in marketing, then we connected network management, then HR. From the point of view of implementation of modern technologies it was one of the most impressive cases in Russia. More than four years have passed since the beginning of implementation, and this experience allows us to talk not only about the implementation stage, but also with 100% knowledge of the case to talk about which victories and problems arise next.
The implementation of systems with elements of artificial intelligence causes a storm of emotions, as it affects all levels of employees: from ordinary to senior management. The first emotional reaction is often rejection. For most, artificial intelligence is a black box. A person does not always understand why a bad or good result was revealed, how the regularity was determined, which factors were taken into account to a greater extent, which — to a lesser extent.
Even at the start of technology development, we consulted with experts, including psychologists, how to increase human confidence in the machine. The answer was simple — the system must earn credibility, and this requires an understanding of the majority of employees of the AI working principles. The task is quite difficult. Objectively, we could not make all the staff of programmers, analysts and machine learning specialists, so next to the main core of AI, we created a "little brother", which clearly explained the decision in understandable metrics based on the facts that do not cause confusion or rejection of humans. It significantly reduced the implementation time and removed many objections, as most employees saw the main stages of the system, and it fit into their usual logic and decision-making algorithms. At the same time, the developers understood that these are not the only conditions that the system takes into account. Decisions are often so accurate precisely because of the huge number of factors taken into account, including those that a person does not take into account because of the seeming insignificance. Machine learning still processes them, and these patterns also have an impact on the final result and make it as accurate as possible
The worst variant of AI implementation is a revolutionary method when the team has neither its understanding nor acceptance. This is a matter of trust in the implementation product and the team that implement it. We understood that we could not instantly change the practice that had developed over the years and make experienced professionals listen to the decisions of the machine. This would create conflict and become a significant barrier to implementation. Therefore, a smooth, phased implementation scheme was developed. In the first year, for example, we introduced a "system of recommendations" and very clearly positioned them as "not instead, but to help", so as not to form a relationship "man against machine" and come to a more productive model "machine helps man, and Vice versa." In this case, people begin to get used to AI, watch how it offers recommendations, begin to find convenience for themselves.
When employees begin to work in a fully digital environment, the entire set of their actions is available for AI observation and forms a unique data window for training. The quality of such a base is incomparably higher than the usual data uploaded for the construction of statistical models. This makes it much faster and more accurate to highlight best practices among employees and include them in recommendations.
Another question that arises after the implementation of artificial intelligence systems in business processes and is directly related to things that we discussed above: "From whom and on whom should the system learn?". In the first stages after implementation, the answer is obvious-it is the history of the organization’s work for a year or two or three before implementation. Next —are the best employees. But if the system of recommendations has been working for two or three years, the following situation arises:
At this stage, there is a fairly well — known situation in the statistics-the sample displacement. New data for training is no longer available. The system is closed in itself and could potentially miss out on new opportunities. The usual method is to allocate a certain proportion of activities in which employees have to go against the system. But as you learn the system, you see that it is increasingly difficult for ordinary employees to formulate and implement new successful practices. Therefore, in this case, it was recommended to create special teams of employees who are experts in the process and at the same time capable of finding new opportunities.
Another important thing that creates difficulties in the implementation of artificial intelligence is that employees see it as a competitor, and not without reason. Any automation leads to the fact that the recommendation system works better over time than employees, which saves the company money, but creates a risk for certain groups of employees.
Sooner or later, they will need to enrich their knowledge with the ability to work with artificial intelligence, and the labor market itself will also be transformed: those who can be easily replaced will leave, but new personnel will be needed who will work with the latest technologies, engage in their implementation, training, support. This is already happening: not so long ago, German Gref said that Sberbank ceases to hire lawyers who do not have an understanding of the principles of working with the neural network. And this is just one particular example of the profession.
Why is this happening? The answer is banal-economy. The introduction of such systems is beneficial to employers. For example, according to WEF data, four years ago, three of Detroit's largest companies and three of Silicon valley's largest companies had comparable revenues, but the latter employed 10 times fewer employees. It is easy to guess which way the business will develop. Even now, in banks according to our data, the majority of decisions in the framework of credit agreements are made with the help of artificial intelligence recommendations. Credit institutions more often give risk management to machines that process a lot of documents, determine the solvency of the client, the risk of bankruptcy, late payments, as well as fraud risks.
By 2030, according to PwC, artificial intelligence will provide 14% growth in global GDP, which is about 15.7 trillion dollars, that is, It will become an integral part of the economy, and we all need to take this into account now. The winner is the one who will join the game earlier and accumulate internal experience of employees’ work in conjunction with AI.
Most likely, in competition with the ubiquitous artificial intelligence, and this, as we already understand, is a matter of time, there will be a number of factors influencing further success: the winner will be the one who will keep the expert group capable of generating new solutions for him, who will be able to digitize a greater number of events, build the most complete data showcase, who will be able to form the strongest team of specialists.