Generative AI and Potential for Augmentation: A Data-Driven Analysis of Labor Market in Russia PDF Free Download
1 / 15
/15
100%
Page 1/15
Generative AI and Potential for Augmentation: A Data-Driven Analysis of Labor
Market in Russia
Maksim Elisov
Lomonosov Moscow State University
Kirill Pshinnik
Innopolis University
Alexandra Bordunos
Graduate School of Management, Saint Petersburg State University
Oksana Zhirosh
Innopolis University
Article
Keywords: generative AI, GenAI, labor market, tasks automation, augmentation, Human-AI Collaboration
Posted Date: November 3rd, 2025
DOI: https://doi.org/10.21203/rs.3.rs-7173895/v1
License: This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License
Additional Declarations: No competing interests reported.
Page 2/15
Abstract
Drawing on the concept of Human-AI Collaboration (HAIC), this research analyzes the exposure of occupations to generative AI-driven automation in the
labor market of Russia using real vacancies data. The study addresses key research questions on the types of tasks and occupations amenable to GenAI
augmentation using GPT-4o for prediction of task automation potential and occupational variability with regard to augmentation prospects. Our ndings
contribute to the body of research revealing signicant disparities in AI potential adoption across tasks and occupations. We detected no tasks or
occupations prone to 100% automation; the highest automation potential of a task is 85% and that of occupation augmentation − 70%. Our major ndings are
threefold. First, occupations in culture, sport, leisure and entertainment, activities in the operation of real estate as well as information and communication
showed the highest augmentation potential, i.e., 63%, 58.8%, and 49.6%, respectively. Second, the potential for augmentation is positively associated with the
level of wages. Third, potential nancial impact by 2030 is predicted to reach 10.8 trillion rubles. The ndings underscore the urgency of reskilling initiatives
and ethical frameworks to mitigate inequality. By bridging theoretical and practical insights, this research informs organizational strategies for responsible AI
integration and highlights pathways to maximize human-AI synergy in the evolving workplace.
1. Introduction
The integration of AI technologies, particularly generative AI like ChatGPT or Midjourney, is increasingly recognized as a transformative force in the
workplace that improves decision-making process, increases productivity and creativity in the daily routine, and streamlines operations. While several waves
of concerns, a so-called “AI winter” [1], [2], witnessed skepticism, which primarily reduced interest and funding in articial intelligence projects, the current
situation can be described as “AI spring” [3], [4] with a growing body of reports and scholarly literature related to introducing generative AI tools in various
occupations, tasks, and efforts to evaluate their eciency. The potential nancial impact of AI adoption on the economy of Russia by 2030 is estimated as
high as 11.6 trillion Russian rubles [5]. These observations motivate further search for the possible means and grounds for proof of real shifts in workforce
eciency and for prediction of the potential for raising productivity
for each occupation due to generative AI technologies.
The emerging trend for the organizations is not just to experiment with generative AI, but to actively implement generative AI tools into daily routine.
Accenture report [6] highlights that generative AI has the potential to signicantly enhance productivity across industries while urging organizations to adopt
these technologies responsibly and strategically. Accenture outlines six essentials for organizations to embrace generative AI:
1. Business-driven mindset: foster experimentation with generative AI.
2. People-rst approach: prioritize reskilling and workforce transformation.
3. Data readiness: prepare proprietary data for effective model training.
4. Sustainable tech foundation: invest in infrastructure to support AI deployment.
5. Ecosystem innovation: collaborate across sectors to enhance AI applications.
. Responsible AI: ensure ethical practices in AI development and deployment.
While some companies are at the experimentation stage of generative AI adoption, others are already experiencing the benets of the technology's disruptive
potential. According to The Future of Jobs Report 2025 [7] by the World Economic Forum (WEF), 83% of the surveyed employers believe that advancement in
AI and
information processing technologies will be one of the main drivers of business transformation during 2025–2030. The Report concludes that AI adoption is
growing rapidly but unevenly, with the information technology sector leading the way while industries like construction lagging behind. This disparity reects
broader trends, as advanced and middle-income economies embrace generative AI more widely than low-income economies, where its use currently remains
minimal. The said Report claims that in response to AI adoption, companies manifest a strong focus on adapting existing workforces via reskilling and
upskilling (77%) and acquiring new talent to navigate the changing landscape brought about by AI. To maximize benets and avoid widening inequality, AI
development should focus on augmenting human capabilities rather than replacing them, supported by appropriate frameworks, incentives, and regulations.
For the purpose of this project, we use the term “automate” with regard to specic tasks and the term “augment” with regard to an occupation. The high
potential of augmentation of an occupation implies a high number of tasks that can be automated within the occupation.
Literature on the affordances of the use of AI in various industries and application for various tasks is gaining momentum. The current research was
motivated by the need to overcome inated expectations and subsequent disappointment of large-scale companies from investing into adoption of
generative AI technologies by searching for exact measures how increased eciency could be estimated and improved. Overall, the research ndings aim to
provide a basis for predictions of opportunities to transform particular job functions in order to foster eciency by the use of generative AI.
This paper reports a part of our broader research effort that aims to introduce generative AI-use training intervention for personnel and measure the impact
of the said intervention on the personnel eciency. To select and prioritise tasks and occupations for such a training intervention, we draw upon the
methodology presented in [8], which introduced an approach to evaluate task exposure and automation potential using GPT-4 LLM and measured overall
exposure in the contemporary U.S. labor market. Our project thus contributes to the growing body of literature on the methodologies and results of the
evaluation of potential use of generative AI tools and its impact on the labour markets, e.g., China [9], Latin America (Chile, Peru, Mexico) [10], EU and,
specically, Germany [11].
Our project contribution is threefold:
1. we analyze the data of the labour market in Russia and advise on the task prioritization for AI-use educational intervention;
Page 3/15
2. the data for the analysis were collected from real vacancies (as compared to expert description of tasks in O-NET base in 7)
3. we used a newer model, i.e., Chat GPT-4o, for data analysis.
Specically, we address the following research questions:
RQ1. What types of tasks in various occupations could be automated by generative AI?
RQ2. Is it possible to assess the potential of such tasks for automation with the GPT-4o model?
RQ3. How do the occupations range by the potential of augmentation?
RQ4. What is the potential nancial impact of generative AI augmentation for the labour market?
The above research questions motivated the following study design: rst, we parsed vacancies from the website of a staff recruitment service provider
Headhunter (https://hh.ru/); then, we uploaded the collected data to ChatGPT-4о LLM and prompted it to, rst, retrieve all tasks from vacancies descriptions,
and, second, build detailed predictions about the GenAI-driven automation potential of the retrieved tasks for particular occupations; further, we assigned
weights to the retrieved tasks based on the frequency of their occurrence in the resulting list of tasks; nally, we categorized the tasks by functional areas,
key competencies, and industrial elds, and completed the data analysis by estimation of the economic impact of the adoption of GenAI. As a result, we
identied no tasks or occupations that are susceptible to complete (100%) automation; the maximum automation potential observed for any individual task
is 85%, while the highest augmentation potential for an occupation is 70%. The sectors of culture, sport, leisure and entertainment; real estate operations;
and information and communication exhibited the greatest augmentation potential — 63%, 58.8%, and 49.6%, respectively. A positive correlation between
augmentation potential and wage levels was identied (Fig.2). The projected potential nancial impact by 2030 in Russia is estimated to reach 10.8 trillion
rubles.
The rest of this paper takes the form of ve sections. Section 2 presents an overview of the literature related to LLM potential for augmentation; Section 3,
details the methods of data collection and analysis; Section 4 presents the results, Section 5, interprets the results and offers the directions of future
research.
2. LLMs potential for augmentation.
The ways humans use AI is being investigated from various perspectives, and the conceptualizations of this phenomenon has been increasingly shifting
towards the idea of shared agency when humans and AI do or can leverage each other’s strengths, e.g., [12], [13], [14], [15], [16]. In this work we rely on the
concept of Human-AI Collaboration (HAIC), specically in the sense articulated in [17], i.e., when humans and AI “
perform interdependent actions”.
In the
analyzed literature, HAIC is seen as a transformative aspect of the labor process; with regard to AI potential for augmentation, the main themes are related to
improved eciency and productivity, creativity and innovation, decision-making, and quality.
2.1. Eciency and productivity
LLMs have a signicant impact on operational eciency. Experimental studies demonstrate that the application of generative AI can substantially reduce the
time required to perform routine or standardized tasks while simultaneously enhancing result quality. For example, the use of ChatGPT in professional writing
leads to reduced time expenditures and higher quality assessments of the nal product. In practical terms, when composing press releases or commercial
proposals, ChatGPT helps structure the text, correct style and grammar, and thus accelerates document preparation while minimizing errors [18]. In the IT
sector, GitHub Copilot considerably speeds up code development processes by enabling programmers to complete tasks 55.8% faster compared to
traditional methods. Moreover, the Copilot not only generates templated solutions but also suggests optimizations, identies potential errors, and assists in
improving software architecture – an especially valuable feature under tight deadlines and intense competition [19].
Furthermore, LLMs are nding application in market research, where their ability to generate plausible responses to consumer inquiries allows for cost-
effective data collection. In the study by Brand et al. [20], it is shown that modeling consumer preferences with the help of LLMs yields results comparable to
those obtained from traditional surveys, but with considerably less time and nancial investment. For instance, companies can use LLMs to simulate
audience reactions to a new product line, enabling them to quickly adjust marketing strategies and identify potential strengths and weaknesses in the
offering at early stages of product development.
Besides, signicant changes are observed at the macroeconomic and labor market levels. Impact assessments indicate that the adoption of LLM
technologies can not only automate a substantial portion of work tasks but also markedly boost overall labor productivity. Eloundou et al. [8] demonstrate
that, thanks to LLMs, approximately 15% of all work tasks can be completed signicantly faster; when supported by LLM-powered software, this gure
increases to between 47% and 56%. Such improvements facilitate the optimization of internal processes within organizations, the reallocation of human
resources, and the emergence of new professional roles associated with coordinating human–AI interactions. Similar conclusions are corroborated by
Dell’Acqua et al. [21], consultants who used AI for tasks within AI capability frontier completed 12.2% more tasks, 25.1% faster, with over 40% higher quality
results; lower-performing consultants improved by 43% and higher-performing ones by 17% compared to their baselines.
2.2. Creativity and Innovation
Kaufman and Beghetto proposed a 4C model of creativity: mini-c, little-c, Pro-c, Big-C [22]. Here we mostly address Pro-c, i.e., professional creativity, a
domain-specic socially-recognised form of creativity, as the most relevant for this study. An example of the positive impact of HAIC on human innovative
thinking is freeing up time and attention of humans for more creative tasks by assigning routine, automatable, tasks to AI. In [23] AI assistance led to a
Page 4/15
signicant increase in creativity among higher-skilled employees, enabling them to generate more innovative solutions and improve performance. When AI
handled the repetitive and codied parts of a task, employees could focus on more complex, creative problem-solving. Another example of such an impact
on team work is described in [24], which, among other research questions, investigated how team expertise is leveraged in new product development tasks
with the use of AI. During ideation stage, without AI assistance, technical and commercial specialists tended to produce ideas within their professional
background while overseeing other perspectives; with AI assistance, however, the said specialists engaged in more holistic, interdisciplinary thinking, which
enabled them to produce more balanced ideas without compromising the effectiveness of the suggested solutions.
From a theoretical perspective, a systematic literature review [25] suggests that HAIC has the potential to transform organisational innovation practices via
seven dimensions: Predicting, Assessing, Decision-making, Problem- solving, Adjusting, Cultivating, and Absorbing. Further, [26] systematically investigate
how the latest developments in AI platforms and technologies inuence innovation ecosystems enabling new ways to generate value. They extend the
existing theories of innovation ecosystems and propose a new conceptualized framework, an “AI innovation ecosystem” consisting of three essential
elements, i.e., actors and roles, data- driven decision-making process impacted by AI, and value generated by AI. They also introduce an “AI-Collaboration
Matrix within Innovation Ecosystems” that illustrates how different levels of AI adoption, combined with varying degrees of collaboration, impact innovation
results. Another conceptualization of AI adoption in innovative rms is presented in [27]. Drawing upon a systematic analysis of empirical studies, they
proposed a framework that represents the inuence of AI adoption on innovation capabilities. The framework consists of two sets of capabilities: six
enabling capabilities
, and seven
enhancing capabilities
, both inuenced by the technological, organizational, and environmental context. The
enabling
capabilities
are necessary to allow for AI adoption, while the
enhancing capabilities
are experienced by innovating rms as a result of AI adoption and allow
for the transformation of innovation practices. Besides, the authors suggest a taxonomy of AI applications, i.e.,
replace, reinforce, reveal
, with the latter
capable to “unveil hidden technological opportunities and unshadow unforeseeable external situations” [p.91, 27].
2.3. Decision-making
Due to a shift from rapid, heuristic-based processes performed by LLMs (System 1) to slower, more analytical and structured methods of information
processing (System 2) [28]. This evolution enables the models not only to generate text but also to perform complex logical reasoning and planning. For
instance, the authors in [29] illustrate the application of these models in solving complex mathematical and optimization problems that require step-by-step
analysis and iterative renement of intermediate solutions.
At the cognitive level, there is an enhancement of "slow thinking", wherein LLMs are employed to support multi-stage reasoning and decision-making. The
models are capable of thoroughly analyzing tasks, highlighting key aspects, and offering several solution alternatives, thereby reducing the likelihood of
errors and improving the overall quality of the nal output. Such capabilities are particularly crucial in areas where precision is critical—such as in the
development of complex software products or in legal research, where detailed analysis and logical justication are required [29], [18].
Study [30] aimed at helping to overcome data scarcity for predictive modeling of ERP adoption. For this, GANs and VAEs were used to create synthetic ERP
adoption data; the generated data closely matched real data, as conrmed by statistical tests; the system allows for more informed decision-making. In [31]
the suggested system combined Digital Twins and Generative AI (like ChatGPT), which enabled it to quickly and accurately detect issues in the test scenarios
as well as learn from past data. This predictive capability can transform decision-making from reactive to proactive, reducing costs, minimizing downtime,
and optimizing performance. A review in [32] explores the application of AI in Finance, and concludes that integrating Generative AI models like GPT-4 with
Big Data enhances, among others, predictive accuracy and creates high-quality synthetic data, which leads to major improvements in data engineering and
enterprise analytics.
2.4. Quality
Noy and Zhang [18] demonstrate that the use of ChatGPT in professional writing tasks signicantly enhances output quality, with treatment group
participants achieving scores 0.45 standard deviations higher than those in the control group, reecting improvements in overall quality as well as in specic
dimensions such as writing quality, content, and originality. The grade distribution shifted upward across the board, indicating a robust quality enhancement
that was particularly pronounced among individuals with lower baseline skills, thereby reducing productivity inequality. This effect appears to stem primarily
from ChatGPT substituting for routine drafting efforts, which allows users to reallocate time towards editing and renement; notably, even when participants
submitted ChatGPT’s raw output without signicant modication, the quality was substantially superior to that achieved without AI assistance. Dell'Acqua et
al. in [21], however, assert that the quality of outputs depends on whether the tasks assigned to AI are within or outside of the “jagged technological frontier”,
i.e., AI capabilities for particular types of tasks. In their experiment, the quality of the outputs obtained with the use of GPT-4 for tasks within AI capabilities
was graded 43% higher compared to the control group (no AI used) for the lower-skilled consultants and 17% higher for the higher-skilled consultants. For
the tasks outside of AI capabilities, though, the outputs of the consultants relying on AI showed 19% lower quality than the outputs of those working without
AI.
In conclusion, the integration of LLMs across various professional domains is likely to be accompanied by complex processes that merge the automation of
routine operations, the improvement of complex analytical tasks, and the emergence of new forms of interaction between humans and AI.
3. Methods
This section details data collection and analysis processes we used to address our research questions.
3.1 Data collection
Page 5/15
To accumulate the data, we rst made a Python-based program. The program collects job vacancies data in HeadHunter (HH.ru) API. We consider the HH.ru
platform representative enough for the purpose of data collection (see https://stats.hh.ru). The program provides functionality to search for vacancies by
occupation title and to save the results in an Excel le with the .xlsx format. We collected such data for 78 most popular occupations; the criterion of
popularity implies the highest numbers of vacancies posted on the said website for a particular occupation; the said 78 occupations represent more than
85% of all the open vacancies in HH.ru. We took 100 vacancies for each occupation from the January 2025 data list. On HH.ru the vacancies are
automatically deactivated and become unavailable for open access 30 days after activation; hence the retrieved list might fail to respond to the changes
related to, e.g., seasonality. Additional analysis revealed that increasing the sample size to over 100 vacancies did not lead to a signicant improvement in
estimate accuracy or produce substantial changes in the model's responses. Initially, we compiled a list of vacancy IDs matching specic parameters. The
program works as follows: a GET request is sent to the HH API with the following parameters: search text (occupation titles), region code (search can be
restricted to specic cities, regions, or countries, we used data for the country). The API response is checked for success, and the data are returned in JSON
format. Each suitable vacancy is added to a table that includes its title and URL. The function stops as soon as the required number of vacancies has been
collected or when there are no more vacancies on the current page. Next, the program processes the vacancy links obtained from the HH.ru website, extracts
data for each vacancy using the HH API, cleans and structures the information, and then saves the results to a new Excel le. Standard Python libraries were
used throughout this procedure.
3.2 Data pre-processing
Upon collection, the textual data were preprocessed by removing HTML tags, decoding HTML entities, and eliminating special characters from the vacancy
descriptions, thereby preserving only plain text. After that, the vacancies were retrieved using the identiers collected in the previous stage. For each
identier, a GET request was sent to the HH.ru API in order to obtain structured vacancy data, including the job title, employer name, location (city), salary,
and description. The program accepts an Excel le containing a list of vacancy URLs, from which it extracts and validates each link. If a URL matches the
standard HH.ru vacancy format, the corresponding vacancy ID is extracted and used to query the API. The retrieved data are appended to a list, and the nal
output is organized in tabular form.
3.3 AI processing
For the neural network-based processing of vacancies, a dedicated table was constructed containing only the textual descriptions of the vacancies. The
program performs asynchronous POST requests to the OpenAI API, utilizing parameters such as the model specication, prompt, and user input. For each
row in the table, an individual task is generated to query the API. These tasks are executed concurrently, which signicantly improves the processing speed
for large datasets. Once the API responses are received, they are written to the resulting DataFrame on a row-by-row basis, with the processing status
recorded for each entry.
Following this, all tasks are processed using the
ChatGPT-o1-mini
model. At this stage, a different model was employed due to its better performance in
processing textual data. Although this model is more effective for natural language understanding tasks, it also incurs a higher usage cost. An expandable
list of unique tasks is initialized. For each incoming task, the model receives both the current list of unique tasks and a new task from the complete set. If the
task can be categorized as matching one of the existing unique tasks, the corresponding task rating is incremented by one. If no match is found (or if the list
is initially empty), the task is appended to the list as a new unique task with an initial rating of one. Here, the rating reects the frequency with which the task
appears in the dataset.
The resulting set of unique tasks is subsequently analyzed by the neural network to assess their potential for automation via large language models (LLMs).
Automation potential is dened as the extent to which the use of an LLM may accelerate the execution of a given task. Tasks deemed unsuitable for
automation receive a score of 0%.
The structured response returned by the API is then parsed and decoded as JSON. If the response contains keys and examples, these are incorporated into
the nal dictionary. For each task, a distinct row is created in the output, containing the original task description, its evaluated characteristics, and, where
available, examples. The nal results are saved to a new Excel le.
Finally, a metric termed “task weight” is introduced. It is dened as the product of the number of relevant vacancies and the automation potential score
(expressed as a percentage). This value reects each task’s contribution to the overall automation potential of a vacancy. Tasks assigned an automation
score of 20% or lower are considered economically inecient to automate, and their contribution is therefore treated as zero. The overall automation score
for a vacancy is calculated as the ratio of the cumulative task weights (with the threshold applied) to the total number of tasks.
The threshold is justied by the observation that the neural network does not consistently assign a 0% automation score, even for tasks that are not
automatable. This behavior can be attributed to the fact that large language models (LLMs) may still be applicable in auxiliary roles—such as suggesting
routes, identifying recreational activities, or commenting on user actions. While these applications may theoretically support certain tasks, they are often of
marginal utility or entail higher implementation costs than the benets they yield.
Initially we generated a workow with the two alternative scenarios: with and without the extended list (Fig.1). Current research is based on the extended
scenario, because it provides task weights.
4. Results
4.1 The potential of augmentation within occupations.
Page 6/15
Table A-1 (Appendix A) generated within the above framework represents an analytical tool for the quantitative evaluation of the potential for automating
professional tasks using large language models (LLMs) based on data obtained via the HH.ru API. The table reects two key indicators: the total number of
unique tasks identied in the job listings, and the average automation percentage calculated for all tasks as well as for the subset of tasks with ratings
exceeding the established threshold. This approach allows not only for the assessment of the overall potential of applying LLMs but also for highlighting
those aspects of professional activities in which the implementation of neural network technologies may lead to a signicant increase in productivity. For the
purpose of this project we use the term “automate” with regard to specic tasks and the term “augment” with regard to an occupation. The high potential of
augmentation of an occupation implies a high number of tasks that can be automated within the occupation.
Data analysis showed that occupations oriented toward information processing, analytics, and managerial functions demonstrate a high level of automation
(Table 1). This is due to the fact that the nature of their tasks entails the active use of textual information, which aligns closely with the functional capabilities
of modern language models (e.g., Table 2). The high automation metrics in these areas indicate a signicant potential for improving the eciency of work
processes through the implementation of LLMs.
Table 1
Top Ten Occupations with the Highest Potential for Augmentation.
Occupations Average Automation % (≥ 20%)
Presentation Designer 70.6%
Dispatcher 67.4%
Product Analyst 66.4%
Writer 65.7%
Data Engineer 65.3%
Marketplace Manager 64.8%
Financial Analyst 64.6%
Head of Analytics Department 63.3%
Procurement Manager 62.1%
Table 2
Potential of Tasks for Automation within Presentation Designer Occupation
Original Task Automatability
Creating visual content for a presentation based on provided text. 85%
Designing an appealing presentation to introduce a new product. 80%
Creating a unique design for a new product launch. 75%
Developing a presentation for a business meeting, including charts and graphs to illustrate key data. 75%
Creating a visual concept for a presentation, including selecting a color palette, fonts, and graphic elements. 75%
Designing a presentation for a corporate event, including theme selection, slide design, and visual content creation. 75%
Creating slide animations to make information more visual and retain audience attention. 75%
Conversely, our study results reveal that professions associated with manual labor or direct interaction with machinery, such as drivers, exhibit a considerably
lower potential for automation (Table 3). In these elds, work processes depend on physical actions and specic skills that are not amenable to direct
processing by language models (e. g., Table 4). Thus, the observed imbalance in automation indicators suggests that the effectiveness of LLM applications
varies signicantly depending on the nature of the tasks performed.
Page 7/15
Table 3
Ten Occupations with the Lowest Potential for
Augmentation
Profession Average Automation %
Driver 7.60%
Nanny 6.50%
Locksmith 3.80%
Security Guard 2.10%
Electrician/Installer 1.00%
Courier 0.20%
Turner (Lathe Operator) 0.10%
Welder 0.00%
Tractor Operator 0.00%
Cleaner 0.00%
Table 4
Potential of Tasks for Automation within the Occupation of Driver
Original Task Automatability
Accompanying the executive on trips around the city, region, and other areas. 5%
Ensuring the safety and well-being of passengers on public transportation. 10%
Car maintenance, including keeping it clean and in working condition. 10%
Providing safe and comfortable transportation for the company executive and their family. 10%
Delivering goods and materials to retail locations within the city. 10%
Compliance with trac regulations. 10%
Managing the process of renting or servicing the vehicle. 30%
Preparing documents. 60%
Expense reporting. 70%
The occupations were clustered based on three distinct principles: by Functional focus, by Industry sector and by Key skills and Competencies. For each
resulting cluster, the average automation level was calculated within each category, representing the proportion of tasks within that category that are
susceptible to automation. This value reects the estimated automation potential for each category under the respective clustering scheme. The data in
Table 5 suggest that AI augmentation will thrive in knowledge-intensive, creative, and analytical elds, while its impact will be more limited in manual or
highly regulated industries. Tailoring AI solutions to sector-specic needs (e.g., automated analytics for nance, design aids for creatives) will maximize
adoption.
Table 5
Prospects of Augmentation in Different Industries
A. By Functional Focus
Functional Area Average Potential Automation %
Creativity and Communication 65.5%
IT and Digital Technologies 46.1%
Analytics and Business Management 44.7%
Education, HR, and Development 41.5%
Medicine and Healthcare 35.7%
Specialized Services and Support 34.5%
Engineering and Technical Manufacturing 20.9%
B. By Industry Sector
Page 8/15
Industry Sector Average Potential Automation %
Digital Technologies and Communications 49.0%
Finance, Analytics, and Management 44.8%
Education and Workforce Development 41.5%
Healthcare and Social Services 35.7%
Specialized Services and Transportation 34.5%
Engineering, Manufacturing, and Construction 20.9%
C. By Key Skills and Competencies
Key Skills and Competencies Average Potential Automation %
Creative Thinking and Communication 49.5%
Analytical and Managerial Skills 44.8%
Professional Services and Support 34.8%
Technical and Software Skills 31.5%
4.2. Potential nancial impact of augmentation with AI on the labor market.
The presented data (Table 6) evaluate the potential impact of generative neural networks, particularly large language models (LLMs), on the labor market in
Russia across various economic sectors. The sectors are identical to the ones used by Rosstat [33]. The analysis focuses on the extent to which LLMs can
enhance worker productivity, the average monthly salaries by sector, the proportion of the national workforce employed in each area, and the resulting
macroeconomic effects. The metric labeled “automation” reects the estimated productivity increase from the integration of LLMs. Automation rates for
each sector were computed as the mean of the individual automation percentages of the occupations comprising that sector. By multiplying this metric with
employment distribution and salary data, the study estimates the annual economic benet that could be realized through eciency gains. Notably, sectors
with broad employment coverage and moderate automation potential—such as Wholesale and Retail Trade, as well as education—demonstrate the highest
estimated yearly savings, exceeding 2.28 and 1.32 trillion rubles respectively. Conversely, areas requiring deep domain expertise, such as scientic research
and development, are associated with lower automation potential and correspondingly smaller economic gains. These ndings emphasize not the
replacement of workers, but rather the opportunity for task redistribution and productivity enhancement, which may lead to GDP growth, labor market
transformation, and the emergence of new professional pathways. We estimate the maximum effect of the introduction of neural networks in the Russian
labor market by 2030 at 10.79 trillion rubles, which is close to the estimate from the Higher School of Economics [5].
Table 6
Statistics based on Rosstat data [33]
Activity Sector Automation Average Salary (RUB) Share of Population Annual Savings, Trillion RUB
Wholesale and Retail Trade 27.7% ₽66,226 14.2% 2.28
Information and Communication 49.6% ₽136,988 2.4% 1.43
Education 37.7% ₽54,315 7.4% 1.33
Agriculture, Forestry, Hunting, Fishing and Aquaculture 46.3% ₽54,158 6.0% 1.32
Construction 22.3% ₽71,707 9.3% 1.30
Professional, Scientic, and Technical Activities 29.1% ₽108,253 4.1% 1.20
Scientic Research and Development 37.4% ₽120,790
Transportation and Storage 21.4% ₽76,223 8.1% 1.16
Healthcare and Social Services 34.5% ₽61,651 6.2% 1.15
Financial and Insurance Activities 36.5% ₽170,600 1.8% 0.98
Real Estate Operations 58.8% ₽55,443 2.6% 0.74
Manufacturing 9.0% ₽71,855 14.2% 0.80
Arts, Sports, Entertainment, and Recreation 63.0% ₽65,702 1.6% 0.58
Administrative and Support Service Activities 40.7% ₽50,573 3.1% 0.56
Total 81.0% 13.49
By 2030 10.79
Page 9/15
We also divided the occupations into four categories depending on the level of automation (Table 7). The data on wages were retrieved from the HH.ru
website. The results indicate a statistically signicant positive correlation between the level of wages and automation potential across occupations (Fig. 2).
These results are consistent with those in [8] and [9].
Table 7
Division by automation sectors
Criterion, % of automation Average Automation Average Salary (thousand RUB)
60%+ 64.7% 136.1
40–59.9% 49.4% 117.8
20–39.9% 32.9% 113.0
0–19.9% 6.8% 104.9
In recent years, the Russian labor market has witnessed a marked shift in wage dynamics, particularly among traditionally manual labor occupations. These
professions — historically associated with lower compensation and limited potential for automation by large language models (LLMs) — have experienced
substantial wage increases, in some cases earning several times more than in prior years [34]. This structural realignment may have attenuated the observed
relationship between income and augmentation. The correlation analysis across 78 professions yielded a Pearson coecient of r = 0.381, with a statistically
signicant p-value of 0.0006, indicating a moderate positive association. However, the 95% condence interval for r, ranging from 0.172 to 0.557, suggests
considerable variability. The ination of wages among less automatable occupations could be contributing to this broader interval and the dilution of a
stronger trend, thereby partially masking the extent to which higher salaries may otherwise correlate with greater exposure to automation.
5. Discussion and conclusion
5.1 Summary of Findings
This study set out to evaluate the possible transformative impact of generative AI (specically GPT-4o) on workplace tasks and productivity in the Russian
labor market. Overall, the results provide robust evidence that the potential for AI-driven task augmentation is signicant but highly uneven across task types
and occupations. Tasks heavily involving information processing, data analysis, and other text-centric activities exhibit the greatest augmentation potential.
For example, occupations in elds like data science, marketing, and management — where daily work revolves around generating or analyzing text-based
information — showed markedly high automation scores, indicating that a large portion of their routine tasks could be accelerated or enhanced with GPT-4o
assistance. Automation is also high for certain professions, such as dispatchers, due to the ability of large language models (LLMs) to effectively receive,
process, and relay information to the appropriate recipients. In contrast, roles that require physical labor, direct manipulation of the environment, or face-to-
face interaction (e.g. drivers, machine operators, and similar manual-intensive jobs) consistently scored low on automatability, as their core tasks are not
readily handled by current language models. This clear disparity supports our rst hypothesis (RQ1) that the amenability of tasks to generative AI depends on
task nature: cognitively intensive and textual tasks are far more augmentable than physically intensive ones.
Crucially, our methodology leveraged GPT-4o to assess task automation potential using real-world job vacancy data, rather than relying solely on expert
judgments or static occupational databases. The GPT-4o-based assessment proved to be feasible and insightful (addressing RQ2). The model was able to
parse thousands of job listings, identify the tasks involved, and evaluate each task’s likelihood of being automated or enhanced by AI. The resulting estimates
were not only intuitively plausible but also exhibited convergent validity when compared with external indicators and studies. Notably, occupations that GPT-
4o identied as highly augmentable tend to be those with higher average wages, and indeed we observed a positive correlation between an occupation’s
automation percentage and its salary level. This aligns with prior research from the University of Pennsylvania, [8] which found that roles commanding higher
wages often have greater exposure to AI technologies. In essence, GPT-4o’s predictions mirrored known patterns (e.g., that well-paid, knowledge-intensive
jobs contain many tasks AI can assist with), lending support to the accuracy of our approach. The third and fourth research questions (RQ3 and RQ4),
concerning how augmentation potential varies across occupations and which occupations rank highest, were also armatively answered. We found
substantial variability across the occupational spectrum: a handful of professions emerged as clear front-runners for AI augmentation (with task automation
potentials well above the chosen threshold), while others lagged far behind. Indeed, we were able to stratify jobs into four broad categories of automation
readiness (from low to high), a categorization that could guide where AI interventions might be most impactful rst.
A high level of automation was observed in the sectors of
Real Estate Operations
and
Arts, Sports, Entertainment, and Recreation
. In the real estate domain,
this can be attributed to occupations such as realtors, which involve tasks like analyzing large volumes of data, generating listings, and interacting with
online platforms—all of which can be signicantly enhanced by large language models (LLMs). In the creative sector, tasks such as writing poetry or
generating visual artwork can also be substantially accelerated through the use of neural networks, highlighting the growing applicability of generative AI in
domains traditionally viewed as human-centric.
Zooming out to the macroeconomic perspective (RQ5), the ndings indicate that widespread adoption of generative AI in the workplace could yield
signicant productivity and eciency boosts and economic gains, albeit distributed unevenly across sectors. By combining each sector's automation
potential (as estimated by GPT-4o) with employment and wage data, we estimated the potential annual eciency gains in monetary terms. The results
suggest that certain large-employment sectors with moderate AI amenability stand to gain the most in absolute terms. For instance, wholesale and retail
trade and education — sectors that employ a substantial share of the workforce — could each realize yearly productivity benets on the order of ₽1–2 trillion
through task augmentation. In contrast, sectors requiring highly specialized human expertise, such as scientic R&D, showed lower augmentation
percentages and consequently smaller aggregate gains. Summing across industries, the upper-bound estimate for economy-wide impact is considerable: our
Page 10/15
analysis suggests that by 2030, generative AI integration could contribute up to roughly ₽10.79 trillion in eciency-related gains annually in Russia’s labor
market. This gure is in line with independent projections by national research bodies (e.g., a similar estimate by the Higher School of Economics [5]). It is
important to note that these gains reect improved productivity and task eciency rather than outright replacement of workers. In fact, a key insight is that
generative AI’s value in this context lies in augmenting human labor — freeing workers from tedious tasks and thereby enabling labor redistribution towards
more complex or creative activities — which can catalyze GDP growth, labor market transformation, and the emergence of new roles.
5.2 Interpretation of Results
The above ndings carry several important interpretations for theory and practice. First, they strongly reinforce the notion that task characteristics are a
decisive factor in determining AI’s impact. This was anticipated by our theoretical framing: according to the Technology Acceptance Model (TAM) [35], the
likelihood of adopting a new technology depends on its perceived usefulness and ease of use for a given job function. GPT-4o effectively has a higher
“perceived usefulness” for tasks that are already digital and information-centric, since it can be readily applied to generate text, analyze language, or expedite
information workows. Our empirical evidence supports this TAM-based expectation – roles where an AI like GPT-4o can be easily applied showed far
greater productivity uplift than roles where it cannot. In practical terms, jobs involving routine cognitive processing were most conducive to AI augmentation,
because GPT-4o could seamlessly slot into those workows (e.g. drafting reports, writing code, summarizing data). By contrast, in occupations centered on
physical skills or interpersonal interaction, the model’s utility is inherently limited, explaining the low augmentation indices observed there. This divergence
underscores a critical point: current generative AI excels at
substituting or speeding up information processing sub-tasks
, but it struggles with tasks requiring
embodiment, physical manipulation, or complex social intelligence. From a human–AI collaboration (HAIC) standpoint, this means the optimal division of
labor is one where AI handles the text-based or procedural elements while humans focus on the physical, empathic, and judgment-based components of
work. Such a synergy was noted in our results as a “synergistic effect,” particularly in elds like marketing, strategic planning, or innovation, where human
creativity coupled with AI-driven analytical support can enhance overall outcomes.
Another key interpretation is the validation of GPT-4o as a predictive tool for workforce analytics. One of our research questions (RQ2) probed whether a
large language model could reliably estimate task augmentation potential. The ndings are encouraging: GPT-4o task exposure estimates correlated with
independent benchmarks (e.g., wage levels, known patterns from prior studies) and yielded plausible sector-wise projections. This suggests that advanced
generative models can serve not just as productivity aids, but also as analytical instruments to forecast technological impacts. In our case, using the model
to analyze real job postings provided a data-driven way to quantify AI exposure at scale, complementing or even accelerating traditional expert surveys. This
approach is a novel contribution of our work, demonstrating how AI can help map out its own impact on labor markets in a more dynamic fashion. It is worth
noting, however, that while the model’s estimates were broadly consistent with external data and conclusions on task exposure in [8], they should be
interpreted as indicative rather than denitive. Generative AI can sometimes overestimate its capabilities or overlook tacit job requirements, so human
validation remains important. Nonetheless, the correspondence we observed bolsters condence in leveraging models like GPT-4o for preliminary
assessments of automation potential in rapidly evolving job landscapes.
The fact that higher-wage occupations showed greater AI augmentation potential is an intriguing outcome with labor economics implications. Historically,
automation in earlier industrial eras often threatened lower-skill, routine jobs; by contrast, generative AI appears to target many higher-skill professions (e.g.,
lawyers, analysts, developers) because those jobs involve abundant information work that AI can optimize. Our data conrmed that industries with higher
average salaries tend to have a larger share of tasks that are automatable by GPT-4o. This might initially raise concern about possible disruption of well-paid
professional roles. However, our interpretation, consistent with the concept of augmentation, is that these roles are more likely to be transformed than
eliminated. Professionals in high-exposure elds stand to become more productive by ooading routine aspects of their job to AI, potentially increasing the
value of their creative and supervisory skills. In fact, evidence from related studies indicates that generative AI can help level the playing eld within such
occupations: for example, an experiment [18] found that when oce workers used ChatGPT for writing tasks, even those with lower prior skills saw
substantial quality improvements, narrowing performance gaps. This hints that AI augmentation, if accessible, could reduce certain skill disparities within
high-skill jobs by allowing a broader range of workers to achieve high-quality outputs. From a theoretical view, this resonates with the HAIC framework –
rather than a zero-sum replacement, we are observing a
complementary enhancement
where human strengths and AI strengths together yield better
productivity and quality than either could alone. It also underscores the importance of human capital: those workers and organizations that effectively adapt
and incorporate AI are likely to reap disproportionate benets (higher output, new innovations), which could widen gaps between rms or individuals if others
lag in adoption. This dynamic invites careful consideration of how to ensure broad-based gains from AI, a topic we address below.
Finally, the macro-level ndings provide a strategic interpretation for economic planning. The projection of trillions of rubles in potential eciency gains
highlights that generative AI could become a signicant driver of productivity growth in the coming decade. However, these gains will not materialize
automatically; they depend on the cumulative choices of enterprises and workers across many sectors. The fact that the largest gains accrue in sectors like
retail and education (which are not traditionally seen as tech-heavy) is insightful – it suggests that even moderate technological improvements, when applied
to very large labor pools, can yield huge aggregate benets. Therefore, a broad-based diffusion of AI tools (even for relatively simple augmentations in day-to-
day tasks like documentation, reporting, or scheduling) could have outsized economic effects. On the other hand, the lower gains projected for specialized
domains (e.g., scientic R&D or nance) imply that in those elds, either the technology has less of a foothold or the work is already highly optimized. It may
also reect that in expert domains, AI is currently used more for quality enhancement than for labor saving. In all cases, our interpretation aligns with the
view that human–AI collaboration is key to unlocking these macro benets – the emphasis is on productivity enhancement and task reallocation, not
straightforward job replacement. The emergence of new professional pathways and the reallocation of human effort from mundane to higher-level tasks
could, if managed well, lead to positive-sum outcomes (e.g., improved services, new industries, and growth in demand for AI-savvy talent). This paints a
picture of the future workforce that is augmented by AI: many jobs will evolve to incorporate AI oversight or co-working, new roles (such as AI workow
coordinators, prompt engineers, or AI ethicists) will become commonplace, and overall economic productivity may surge during the “GenAI era” as these
technologies diffuse.
Page 11/15
5.3 Implications for Policy and Practice
The uneven yet signicant impact of generative AI on work tasks carries important implications for business leaders, workers, and policymakers. At the
forefront, organizations should approach AI integration strategically and humanely. Rather than indiscriminately automating tasks, employers are advised to
adopt a “people-rst” augmentation strategy – identifying which tasks can be reliably handed off to AI and retraining employees to focus on the
complementary aspects of their roles. This aligns with industry guidance such as Accenture call for a
People-First Approach
that prioritizes reskilling and
workforce transformation as companies embrace generative AI [6]. Our ndings highlight which occupations and task types should be prioritized for such
interventions. For example, roles in analytics, marketing, and other high-exposure areas could be early targets for deploying GPT-based tools to handle
routine data processing or content generation. By doing so, organizations can boost eciency in these functions while simultaneously freeing employees to
concentrate on creative strategy, complex decision-making, and interpersonal responsibilities that AI cannot fulll. However, realizing these gains requires
substantial investment in training and change management. Employers should invest in upskilling programs that make staff procient in using AI tools (AI
literacy), and cultivate an organizational culture that views AI as a collaborative partner rather than a threat. Notably, resistance or “AI anxiety” among
employees is a real obstacle; to overcome it, leaders should emphasize success stories, involve employees in AI adoption plans, and ensure transparency
about how the technology works and what data it uses. When workers understand AI’s limitations and strengths, and see it as augmenting their work rather
than spying on or replacing them, they are more likely to embrace it, leading to better outcomes.
Policymakers and regulators likewise have a crucial role to play in guiding the GenAI-driven transition in labor markets. Education and vocational training
policies must be updated to reect the changing skill demands: curricula should integrate data literacy, prompt engineering, and human–AI collaboration
skills, preparing new entrants for AI-enhanced workplaces. Governments could partner with industry to create reskilling initiatives for mid-career workers in
at-risk occupations, ensuring that those whose tasks are highly automatable are given pathways to move into more secure roles. The evidence that high-
wage, high-skill jobs are also heavily exposed to AI means that continuous learning is imperative even for well-educated professionals; thus, policy support
for lifelong learning and professional development in AI-related competencies will be benecial across the board. Furthermore, labor regulations and social
safety nets may need updating. As tasks shift, job descriptions and classications might need revision to accurately capture new AI-in-the-loop
responsibilities. Policymakers should also monitor for any emergent inequalities: for instance, if certain groups or regions adopt AI more slowly, targeted
support or incentives might be required to prevent widening productivity gaps. On the ip side, if AI drastically increases output in certain sectors, there may
be a case for sharing the gains (through higher wages or reduced working hours) to ensure workers benet from the productivity dividend. Responsible AI
governance is another policy implication — as organizations deploy GPT-4o-like systems, concerns around data privacy, algorithmic bias, and accountability
for AI-generated errors will grow. Regulators should establish clear guidelines that encourage innovation while protecting workers’ rights and societal values.
This includes enforcing transparency in AI decision-making (especially in high-stakes domains), mandating human oversight for critical decisions, and
perhaps dening ethical standards for human–AI workplace interactions.
From a broader perspective, our study supports the view that maximizing the benets of generative AI requires focusing on augmentation over replacement.
This principle should be enshrined in both company practice and policy frameworks. International labor studies (such as [7]) echo this emphasis: a majority
of employers plan to focus on adapting their workforce through reskilling (77% of surveyed companies) and stress that AI should be used to complement
human workers rather than replace them, accompanied by appropriate supportive measures and regulations. In practical terms, companies should establish
internal guidelines for human-AI collaboration, delineating which decisions or tasks must remain under human control and how AI outputs are to be veried.
They should also invest in what Accenture termed a
sustainable tech foundation
and
ecosystem innovation
: upgrading IT infrastructure to safely deploy AI at
scale and collaborating with other organizations (e.g., through industry consortia or public-private partnerships) to share best practices and resources for AI
adoption. By fostering an innovation ecosystem, even sectors or rms that lag in AI expertise (such as traditional industries) can catch up through knowledge
transfer and shared platforms. Finally, ethical practices must guide this transition. Responsible AI use — encompassing fairness, accountability, transparency,
and security — is not just a slogan but a necessity for long-term trust and ecacy. For instance, if generative AI is used to screen job candidates or evaluate
performance (a possible extension of our work on task analysis), checks should be in place to prevent bias or unjust outcomes. Similarly, when AI aids in
content creation, organizations should set standards to avoid the spread of misinformation or to clearly attribute AI-generated material. In summary, the
implication is that human-centric policies and practices will determine whether the productivity boosts quantied in our study translate into sustainable
economic and social benets. With proactive reskilling programs, supportive governance, and a commitment to using AI as a tool for empowerment, the
workforce can not only weather the AI revolution but thrive alongside it.
5.4 Limitations
While this research provides valuable insights, several limitations must be acknowledged:
Scope of Data (Russia-specic): Our analysis was conned to the Russian labor market and relied on job vacancy data from a single platform (HH.ru).
Labor dynamics, task compositions, and AI adoption rates can differ in other countries or even within non-digital segments of the Russian economy.
Thus, caution is needed in generalizing the quantitative results beyond the studied context. Future studies should expand to diverse data sources and
geographic contexts to verify if similar patterns hold.
Use of GPT-4o for Task Evaluation: We utilized the GPT-4o model to estimate task automation potential, which introduces uncertainties inherent to the
AI’s judgments. While we found GPT-4o’s assessments credible and consistent with external benchmarks, they are ultimately
model-generated
estimates
. The model might misinterpret certain task descriptions or lack up-to-date knowledge of niche job requirements, potentially leading to biased
or inaccurate automation scores. There was no direct human validation of each task rating in this study. This limitation suggests our ndings are best
viewed as indicative estimates of automation potential, not precise predictions of actual outcomes in every workplace.
Task Denition and Thresholding: The way tasks were parsed from job postings and the threshold for “automatable” tasks could affect results. Job
listings may not enumerate all tasks comprehensively, and they often focus on current needs, possibly underrepresenting infrequent but important
Page 12/15
duties. Moreover, we applied a certain cutoff (automation probability) to decide if a task is counted as automatable. Different threshold choices or
weighting schemes might change the measured percentages. We partially addressed this by introducing a “task weight” metric (tasks × vacancies) to
highlight economically signicant tasks, but the approach still simplies the complex spectrum of task automation feasibility into binary or average
metrics. In reality, many tasks lie in a gray area of partial automation, which our summary metrics might not fully capture.
Temporal and Technological Evolution: The study provides a snapshot based on GPT-4 generation technology as of now. AI capabilities are rapidly
advancing; future models might overcome some limitations (for instance, better handling of multimodal inputs or physical reasoning) that currently
constrain automation potential. Conversely, the regulatory environment and public sentiment towards AI could shift, affecting adoption. Our projection
up to 2030 assumes a certain trajectory of technology improvement and uptake, but unforeseen breakthroughs or setbacks could alter that path. Thus,
the long-term macroeconomic estimates carry uncertainty – they represent a maximum potential if AI is adopted extensively, rather than a guaranteed
outcome.
Focus on LLMs (Generative AI) Only: We specically examined generative text-based AI. However, workplace automation can also come from other AI
domains (computer vision, robotics, expert systems) and from broader process innovations. Some occupations currently showing low LLM
augmentation potential (e.g., drivers or trades) might eventually be signicantly impacted by non-LLM AI such as autonomous vehicles or robotic
automation. Our study does not encompass these technologies. Likewise, even in high-LLM occupations, there are non-textual tasks (like creating visual
designs, or performing hands-on experiments) that GPT-4o cannot assist with. The total automation potential of an occupation might be underestimated
if other AI tools are considered, or overestimated if we assume GPT-4o alone can tackle everything. We focused on the generative text aspect, so our
conclusions should be integrated with analyses of other AI tools for a complete picture.
No Direct Measurement of Productivity Gains: We inferred potential productivity improvements from task automation rates and prior research, but we did
not measure actual productivity changes in workplaces using GPT-4o. Factors like the learning curve in adopting AI, the quality of AI outputs, and human
oversight required could reduce the realized productivity vs. the theoretical maximum. Additionally, productivity gains do not automatically translate to
economic gains if, for example, organizational or market constraints prevent output from increasing. Our macroeconomic benet calculations assume
eciency translates proportionally into cost savings or output – an assumption that may not hold in every scenario (e.g., if demand for the
product/service is xed). Empirical studies tracking companies that implement GPT assistance would complement our approach by revealing how much
of the estimated potential is captured in practice.
In summary, these limitations suggest that our ndings should be viewed as exploratory and illustrative of broad trends rather than precise forecasts. They
open several avenues for renement, as discussed next, and highlight the need for ongoing research and validation as generative AI technologies and their
workplace applications continue to evolve.
5.5 Future Research Directions
Building on this study insights and limitations, we identify several directions for future research:
Granular Task Analysis: Future work should examine which specic task attributes (e.g. complexity, standardization, creativity level) most inuence
automatability. Our results hinted at task complexity and workow standardization as factors (simple, standardized tasks were easier to automate).
Rigorous analysis could involve classifying tasks by complexity or interaction level and seeing how AI performance varies, helping to ne-tune the
selection of tasks for automation.
Longitudinal and Experimental Studies: To gauge the real-world impact of GPT-4 and similar AI on productivity, longitudinal case studies or controlled
experiments within organizations are needed. Researchers could deploy generative AI tools in certain teams and measure productivity, quality, job
satisfaction, and skill change over time compared to control groups. Such studies would validate (or adjust) the projected gains and identify any
unintended consequences (e.g., over-reliance on AI, changes in collaboration patterns).
Multi-Modal and Integrated AI Systems: Since many jobs involve non-textual tasks, future research should extend analysis to multi-modal AI (combining
language, vision, and robotics). For instance, investigating how language models can cooperate with robotic process automation or computer vision
systems would provide a fuller estimate of automation potential in occupations that require both cognitive and physical activities. This includes
exploring solutions for professions with low LLM suitability by integrating other AI technologies , thereby moving towards a more holistic human-AI
workow integration.
Cross-Country and Cross-Industry Comparisons: It would be valuable to replicate this study’s methodology in different labor markets (e.g., countries with
varying income levels or different industry structures) and across more industries. Such comparisons could reveal how cultural, economic, or regulatory
differences mediate AI’s impact on work. They might also validate whether the correlation between wages and AI exposure holds universally or is
context-dependent. Additionally, sectors like healthcare, law, or public services warrant focused studies, as they have unique professional norms and
data sensitivity issues that affect AI adoption.
Human Capital and Inequality Impacts: Further research is needed on the labor economics aspects – particularly how AI augmentation affects
employment levels, wage distributions, and skill demands over time. Will AI augmentation lead to higher wages for those who master it and stagnation
for those who don’t, thereby widening income inequality? Or will it democratize expertise and reduce skill premiums (since AI can assist less-skilled
workers)? Economic modeling combined with empirical labor data as AI diffusion progresses can help answer these questions. This includes examining
secondary effects, such as job creation in AI oversight and maintenance, and the redeployment of labor into new tasks that AI cannot do (yet).
Policy and Ethical Frameworks: Interdisciplinary research involving law, ethics, and public policy is crucial to accompany technical and economic
analysis. Studies could propose and evaluate frameworks for responsible AI integration in workplaces, drawing on real-world pilots. For example, what is
the effect of implementing an “AI ethics audit” in a company or giving employees a formal role in governance of AI tools? What kinds of regulatory
incentives or standards most effectively encourage rms to invest in worker retraining alongside AI investments? Such research would inform guidelines
to ensure that AI’s productivity gains are achieved in a fair, transparent, and socially benecial manner .
Page 13/15
Evolution of Human-AI Collaboration Paradigms: Finally, research should continue to explore how human-AI collaboration can be optimized. As AI
systems become more capable, the design of workows will need to dynamically allocate tasks between humans and AI agents. Future studies might
draw on organizational psychology and design thinking to develop new collaboration models, prototyping how teams that include AI “colleagues”
operate. This line of inquiry is aligned with our broader research agenda and could yield best practices on managing “hybrid” teams, mitigating AI-related
stress, and maximizing the
synergistic effect
noted in our conclusion.
By pursuing these future directions, scholars and practitioners can deepen understanding of generative AI’s evolving role in the workplace. The goal is to
continually rene both the
predictive analytics
(how the anticipated impact of AI is measured) and the
prescriptive guidance
(the way to positively shape
such an impact). This opens a possibility to better ensure that the GenAI era leads to augmented human capabilities, sustainable productivity growth, and
broadly shared prosperity, rather than unintended disruption.
Declarations
Funding:
The authors declare that no funds, grants, or other support were received during the preparation of this manuscript.
Author Contribution
Maksim Elisov contributed to the research design, created the new software used in the work, performed data collection and analysis, prepared gures and
tables, reviewed literature, drafted and revised the report.Kirill Pshinnik conceived the study idea, guided the research design, data collection and analysis.
Aleksandra Bordunos contributed to the research design, data analysis, literature review, drafted and revised the report.Oksana Zhirosh contributed to the
research design, data analysis, literature review, drafted and revised the report.All the authors reviewed the manuscript.
Data Availability
Data is provided within the manuscript.
References
1. S. Samoili, M. Lopez Cobo, B. Delipetrev, F. Martinez-Plumed, E. Gomez Gutierrez, and G. De Prato,
AI Watch. Dening Articial Intelligence 2.0
, EUR
30873 EN, Luxembourg: Publications Oce of the European Union, 2021. ISBN 978-92-76-42648-6. doi:10.2760/019901.
2. Y. Jiang, X. Li, H. Luo, S. Yin, and O. Kaynak, “Quo vadis articial intelligence?,”
Discover Artif. Intell.
, vol. 2, art. 4, Mar. 2022, doi: 10.1007/s44163-022-
00022-8.
3. S. Hajkowicz, C. Sanderson, S. Karimi, A. Bratanova, and C. Naughtin, (2023). Articial intelligence adoption in the physical sciences, natural sciences,
life sciences, social sciences and the arts and humanities: A bibliometric analysis of research publications from 1960-2021.
Technology in Society
,
74
,
doi.org/10.1016/j.techsoc.2023.102260
4. KPMG, “Trust, attitudes, and use of articial intelligence,” 2025. [Online]. Available: https://kpmg.com/us/en/articles/2025/trust-attitudes-and-use-of-
articial-intelligence.html
5. Yu. Ya. Dranev, I. I. Kuchin, and M. I. Miryakov, “Economic effect of the implementation of articial intelligence technologies in Russia,” Institute for
Statistical Studies and Economics of Knowledge, NRU Higher School of Economics, Moscow, 2025. [Online]. Available:
https://issek.hse.ru/news/1022068478.html
. Accenture,
A New Era of Generative AI for Everyone
, Accenture, Mar. 2023. [Online]. Available:
https://www.accenture.com/content/dam/accenture/nal/accenture-com/document/Accenture-A-New-Era-of-Generative-AI-for-Everyone.pdf.
7. World Economic Forum, Future of Jobs Report 2025, Insight Report, Jan. 2025. Accessed: Apr. 26th, 2025, [Online]. Available:
https://www.weforum.org/publications/the-future-of-jobs-report-2025/
. T. T. Eloundou, S. Manning, P. Mishkin, and D. Rock, “GPTs are GPTs: An early look at the labor market impact potential of large language models,”
Working Paper
, OpenAI, OpenResearch, Univ. of Pennsylvania, Aug. 2023. [Online]. Available:https://arxiv.org/abs/2303.10130
9. Q. Chen, J. Ge, H. Xie, X. Xu, and Y. Yang, “Large language models at work in China’s labor market,”
China Econ. Rev.
, vol. 40, art. 102413, Apr. 2025, doi:
10.1016/j.chieco.2025.102413.
10. O. A. Herrera, L. Ripani, and E. T. Ramirez, "AI and the Increase of Productivity and Labor Inequality in Latin America: Potential Impact of Large Language
Models on Latin American Workforce," 2024. [Online]. Available: https://publications.iadb.org/en/ai-and-increase-productivity-and-labor-inequality-latin-
america-potential-impact-large-language
11. M. Oschinski, “Assessing the Impact of Articial Intelligence on Germany's Labor Market: Insights from a ChatGPT Analysis,
”
2023. [Online]. Available:
https://mpra.ub.uni-muenchen.de/118300/
12. R. Chaves et al., “Charting a course at the human–AI frontier: A paradigm matrix informed by social sciences and humanities,”
AI Soc.
, vol. 40, no. 2, pp.
123–134, Mar. 2025, doi:10.1007/s00146-025-02295-w.
Page 14/15
13. M. Raees, I. Meijerink, I. Lykourentzou, V. J. Khan, and K. Papangelis, "From explainable to interactive AI: A literature review on current trends in human-AI
interaction,"
Int. J. Hum.-Comput. Stud.
, Art. no. 103301, 2024. [Online]. Available:
https://www.sciencedirect.com/science/article/pii/S1071581924000855
14. H. Cui and T. Yasseri, "AI-enhanced collective intelligence," Patterns, vol. 5, no. 11, 2024. [Online] Available: https://www.cell.com/patterns/fulltext/S2666-
3899(24)00233-2
15. J. Li, J. Huang, J. Liu, and T. Zheng, “Human-AI cooperation: Modes and their effects on attitudes,”
Telemat. Inform.
, vol. 73, Art. no. 101862, 2022.
[Online] Available: https://www.sciencedirect.com/science/article/pii/S0736585322000958
1. V. Kolbjørnsrud, “Designing the Intelligent Organization: Six Principles for Human-AI Collaboration,”
California Management Review
, vol. 66, no. 2, pp. 44–
64, Nov. 2023, doi: 10.1177/00081256231211020.
17. D. Koutrintzes, C. Spatharis, and M. Dagioglou, “Human-Aware Design for Transferring Knowledge During Human-AI Co-Learning,” in
Proceedings of the
Workshop on Embracing Human-Aware AI in Industry 5.0 (HAII5.0 2024)
, CEUR Workshop Proceedings, vol. 3765, pp. 67–73. [Online]. Available:
https://ceur-ws.org/Vol-3765/Camera_Ready_Paper-05.pdf
1. S. Noy and W. Zhang, “Experimental evidence on the productivity effects of generative articial intelligence,”
Science
, vol. 381, pp. 187–192, 2023.
[Online]. Available: https://www.science.org/doi/abs/10.1126/science.adh2586
19. S. Peng, E. Kalliamvakou, P. Cihon, and M. Demirer, "The impact of AI on developer productivity: Evidence from GitHub Copilot," arXiv preprint
arXiv:2302.06590, 2023. [Online]. Available: https://arxiv.org/abs/2302.06590
20. J. Brand, A. Israeli, and D. Ngwe, "Using LLMs for Market Research," Working Paper 23-062, Jul. 29, 2024. [Online]. Available:
https://www.hbs.edu/ris/Publication%20Files/23-062_ed720ebc-ec4d-4bc3-a6ba-bad8cfbd9d51.pdf
21. F. Dell’Acqua et al, "Navigating the jagged technological frontier: Field experimental evidence of the effects of AI on knowledge worker productivity and
quality," Working Paper 24-013, Sep. 2023. [Online]. Available: https://www.hbs.edu/ris/Publication%20Files/24-013_d9b45b68-9e74-42d6-a1c6-
c72fb70c7282.pdf
22. J. C. Kaufman and R. A. Beghetto, “Beyond Big and Little: The Four C Model of Creativity,”
Review of General Psychology
, vol. 13, no. 1, pp. 1–12, Mar.
2009, doi: 10.1037/a0013688. [Online]. Available: https://psycnet.apa.org/record/2018-70050-001
23. N. Jia, X. Luo, Z. Fang, and C. Liao, “When and How Articial Intelligence Augments Employee Creativity,”
Academy of Management Journal
, vol. 67, no.
1, pp. 5–32, Feb. 2024, doi: 10.5465/amj.2022.0426.
24. F. Dell’Acqua et al., “The Cybernetic Teammate: A Field Experiment on Generative AI Reshaping Teamwork and Expertise,” National Bureau of Economic
Research, Working Paper No. 33641, Apr. 2025. doi: 10.3386/w33641.
25. J. Zheng, Y. Hong, and A. Richter, “Articial Trailblazing: How Human-AI Collaboration Transforms Organizational Innovation Practices,” in
Proc. 58th
Hawaii Int. Conf. on System Sciences (HICSS 2025)
, Jan. 2025, pp. 307–316, doi: 10.24251/HICSS.2025.038.
2. G. Secundo, C. Spilotro, J. Gast, and V. Corvello, “The transformative power of articial intelligence within innovation ecosystems: A review and a
conceptual framework,”
Rev. Manag. Sci.
, 2024, doi: 10.1007/s11846-024-00828-z.
27. F. Gama and S. Magistretti, “Articial Intelligence in Innovation Management: A Review of Innovation Capabilities and a Taxonomy of AI Applications,”
J.
of Product Innov. Manage.
, vol. 42, no. 1, pp. 76–111, Jan. 2025, doi: 10.1111/jpim.12698.
2. Y. Bengio, “The Consciousness Prior,”
arXiv preprint arXiv:1709.08568
, Sep. 2017. [Online]. Available: https://arxiv.org/abs/1709.08568
29. S. Yao et al., “Tree of Thoughts: Deliberate Problem Solving with Large Language Models,” in
Advances in Neural Inf. Process. Syst.
36 (NeurIPS 2023),
2023. [Online]. Available: https://arxiv.org/abs/2305.1060
30. K. C. Hong, A. S. Shibghatullah, T. C. Ling, and S. Sarsam, “Generative AI-powered predictive analytics model: Leveraging synthetic datasets to determine
ERP adoption success through critical success factors,”
Int. J. Adv. Comput. Sci. Appl.
, vol. 15, no. 5, pp. 469–482, 2024.. [Online].
31. M. Mateev, “Predictive analytics based on digital twins, generative AI, and ChatGPT,” in
Proc. 27th World Multi-Conf. Syst., Cybern. Inform. (WMSCI)
,
2023. [Online]. Available: https://www.iiis.org/DOI2023/SA437PH/
32. S. Joshi, "The Synergy of Generative AI and Big Data for Financial Risk: Review of Recent Developments,"
Int. J. Multidiscip. Res. (IJFMR)
, vol. 7, no. 1,
Jan.-Feb. 2025. [Online]. Available:
https://www.researchgate.net/publication/388398425_The_Synergy_of_Generative_AI_and_Big_Data_for_Financial_Risk_Review_of_Recent_Developmen
33. Federal State Statistics Service (Rosstat), "Russian Statistical Yearbook 2024," [Online]. Available: Russian Statistical Yearbook.
34. SberIndex, “Median Monthly Wages Dashboard,” Sberbank, Moscow, Russia, Feb. 2025. [Online]. Available: https://sberindex.ru/en/dashboards/median-
wages
35. F. D. Davis, R. P. Bagozzi, and P. R. Warshaw, “User acceptance of computer technology: A comparison of two theoretical models,”
Management Science
,
vol. 35, no. 8, pp. 982–1003, Aug. 1989, doi: 10.1287/mnsc.35.8.982.
Figures
