Page 119 - Azerbaijan State University of Economics

THE JOURNAL OF ECONOMIC SCIENCES: THEORY AND PRACTICE, V.77, # 1, 2020, pp. 113-132

Based on the assessment of the author, the main adoption and implementation
problems of AI is: The company culture does not recognize needs for AI in 23%;
Lack of data or data quality issues in 19%; Lack of skilled people or hiring
problems in 18%; Difficulties in identifying appropriate business use cases in
17%. Fewer hurdles with values below 10%: Technical infrastructure challenges;
Legal concerns, risks or compliance issues; Efficient tuning of hyperparameters;
Workflow reproducibility.

Based on these, the problems can be categorized as they are data, people, or business
issues. However, all companies are diverse and work in different ways, so they see
the occurrence of these problems otherwise and treat them differently. But in
general, the following can be said about these problems.

In the case of data problems, this may include difficulties with the quality and
quantity of the data. An AI system can work well if the data provided by the
company are adequate, as the company itself is responsible for providing the data.
For the AI to work, it is also needed to choose the right model for the system to
work. It should be taken into account that not all data is available even to the
company. In the case of hard-to-reach data, we can also talk about synthetic data,
which, based on existing information, means artificially generated data, in which AI
can also help. Another problem could be untagged data. Labeling is also possible
within the company, but a company can choose to outsource the work in the form of
data programming or use an existing data set. Another obstacle may be the explain-
ability of the data. If the system provides a result that is not clear to the user or has
not reached the same decision regarding the process, then AI technology may not
meet expectations. Therefore, the company should choose a model that it knows for
sure is transparent and also justifies the results generated by AI. In terms of bias, it is
sure that the AI is not biased and has no opinion. However, if only data is available
that man provides to the system out of bias, in his judgment, we already get a biased
result. However, this is not the failure of AI, but human judgment. Because the
system receives the wrong data in the event of bias, the results can also be
erroneous, which is sometimes diagnosed as a system error. The more advanced the
AI system, the more complex it is and so it can be difficult to determine and identify
in retrospect where the error slipped.

In terms of people issues, two possible problems are worth highlighting here. One is
that a lack of understanding of AI can hinder its proper functioning in many areas. If
people who use the application do not accept the technology or do not understand
how it works, it is impossible to achieve efficiency. The other hurdle is the lack of a
professional in the field, those, who know how to apply the technology and can react
119