Multidisciplinary Journal Epistemology of the Sciences
Volume 3, Issue 3, 2026, July–September, Special Edition
DOI: https://doi.org/10.71112/pahqyx95
THEORY AND PRACTICE OF AI-AUGMENTED SOFTWARE ENGINEERING IN
EUROPE IN 2025
TEORÍA Y PRÁCTICA DE LA INGENIERÍA DE SOFTWARE AUMENTADA POR
INTELIGENCIA ARTIFICIAL EN EUROPA EN 2025
Denis Svyatoslavovich Pashchenko
Russian Federation
DOI: https://doi.org/10.71112/pahqyx95
159 Multidisciplinary Journal Epistemology of the Sciences | Vol. 3, Issue 3, 2026, July–September, Special Edition
Theory and practice of ai-augmented software engineering in Europe in 2025
Teoría y práctica de la ingeniería de software aumentada por inteligencia artificial
en Europa en 2025
Denis Svyatoslavovich Pashchenko
a,*
denpas@rambler.ru
https://orcid.org/0000-0001-9089-8173
*
Corresponding author: denpas@rambler.ru,
a
Independent consultant and researcher, In
software technologies, Moscow, Russian Federation.
ABSTRACT
This article examines the practical application of artificial intelligence tools, specifically ChatGPT
versions 3.5 and 4, throughout the complete lifecycle of a real-world software development
project. The study is based on field research conducted in Europe (2022-2023) and evaluates
the effectiveness of AI in generating key software engineering artifacts, including functional
specifications, source code, automated test cases, user documentation, and application content.
The findings indicate that ChatGPT served as a highly effective assistant in the creation and
refinement of these deliverables. However, its limitations became apparent in tasks requiring
current technical knowledge, adaptation to rapidly evolving technologies, or the development of
innovative business logic. Overall, the use of ChatGPT substantially accelerated the software
development process, with project team members reporting significant improvements in
productivity. The study further identifies best practices for integrating AI into software
engineering workflows, emphasizing the importance of service-oriented architecture, iterative
prompt engineering, and continuous human oversight. The results support the research
DOI: https://doi.org/10.71112/pahqyx95
160 Multidisciplinary Journal Epistemology of the Sciences | Vol. 3, Issue 3, 2026, July–September, Special Edition
hypothesis that AI can reliably generate essential software development artifacts with minimal
human intervention, thereby representing a significant step toward the broader adoption of AI-
assisted software engineering practices.
Keywords: AI; software engineering; LLM; Android; ChatGPT.
RESUMEN
Este artículo explora la aplicación práctica de herramientas de IA (ChatGPT v3.5 y 4) en todo el
ciclo de vida de un proyecto real de desarrollo de software, basado en investigación de campo
realizada en Europa durante 2023–2024. La investigación evalúa la efectividad de la IA en la
generación de artefactos centrales de software, incluyendo especificaciones funcionales,
código fuente, pruebas automatizadas, documentación de usuario y contenido de aplicaciones.
Si bien la herramienta de IA demostró ser un asistente potente para crear y refinar estos
entregables, sus limitaciones se hicieron evidentes en áreas que requieren orientación técnica
actualizada o aportes creativos en el desarrollo de la lógica de negocio del producto. En
términos generales, ChatGPT aceleró significativamente el proceso de desarrollo, y los
miembros del equipo del proyecto reportaron un aumento sustancial en la productividad. El
estudio también destaca buenas prácticas para el uso de IA en ingeniería de software,
enfatizando la importancia del diseño orientado a servicios, el refinamiento iterativo de prompts
y la supervisión humana continua. A pesar de sus limitaciones en la generación de ideas
originales o en la adaptación a requisitos cambiantes de plataformas, ChatGPT demostró
sólidas capacidades para automatizar tareas repetitivas y mejorar la eficiencia global.
Los hallazgos confirman la hipótesis de investigación de que la IA puede producir de manera
confiable artefactos clave del desarrollo de software con mínima intervención humana,
marcando un paso decisivo hacia una integración más amplia de la IA en las prácticas de
ingeniería de software.
DOI: https://doi.org/10.71112/pahqyx95
161 Multidisciplinary Journal Epistemology of the Sciences | Vol. 3, Issue 3, 2026, July–September, Special Edition
Palabras clave: AI; ingeniería de software; LLM; Android; ChatGPT.
February 16, 2026 | Accepted: June 30, 2026 | Published: July 1, 2026
INTRODUCTION
The development of software using artificial intelligence (AI) tools has remained a
significant subject of both scientific inquiry and industrial practice for more than fifteen years.
Throughout this period, researchers, technology companies, and software engineering
practitioners have continuously investigated the potential of AI to transform the software
development lifecycle by automating complex engineering activities, improving software quality,
and increasing developer productivity. Numerous specialized and non-specialized organizations
have attempted to forecast the future evolution of software engineering, consistently identifying
AI technologies as one of the principal drivers of innovation and digital transformation (Gorban &
Grechuck, 2018). These studies have envisioned software development environments in which
intelligent systems actively support or automate a broad spectrum of engineering tasks, resulting
in highly efficient production processes characterized by extensive automation, data-driven
decision-making, and the widespread adoption of "intelligent" technologies (Kästner & Kang, 202).
Despite these long-standing expectations, the practical realization of AI-assisted software
engineering remained limited for many years. Earlier generations of AI-based development tools
typically exhibited constrained functionality, required substantial technical expertise, or were
applicable only to narrowly defined programming tasks. Consequently, although the theoretical
potential of AI in software engineering was widely acknowledged (Barenkamp et. al., 2020), its
adoption in everyday software development practices remained relatively modest. AI technologies
were primarily employed in isolated activities, such as static code analysis, defect prediction,
automated testing, or recommendation systems, rather than as comprehensive assistants
capable of supporting the entire software development lifecycle.
DOI: https://doi.org/10.71112/pahqyx95
162 Multidisciplinary Journal Epistemology of the Sciences | Vol. 3, Issue 3, 2026, July–September, Special Edition
A fundamental turning point occurred in the autumn of 2022 with the public release of large
language models (LLMs), particularly ChatGPT. For the first time, highly capable generative AI
systems became readily accessible to millions of software developers, analysts, testers, technical
writers, and project managers worldwide. This unprecedented accessibility, combined with
significant advances in natural language processing, reasoning capabilities, and code generation,
enabled a profound transformation in the way software projects could be designed, implemented,
tested, documented, and maintained. AI evolved from being a specialized auxiliary technology
into a general-purpose engineering assistant capable of participating in virtually every phase of
software development.
Since that milestone, software engineering has undergone rapid organizational and
technological change. Development teams have increasingly integrated AI tools into their daily
workflows to support requirements elicitation, functional specification development, software
architecture design, source code generation, automated test creation, technical documentation,
user support materials, and various project management activities. Consequently, the emergence
of modern generative AI has not merely introduced another software development tool but has
fundamentally altered the organizational and production paradigms of the software engineering
industry. The combination of advanced language models, widespread accessibility, and practical
usability has enabled a level of automation that had previously been anticipated primarily in
theoretical discussions.
Traditional twentieth-century software development centers, typically structured around
project-based teams managed by local leadership, are increasingly being replaced by
geographically distributed micro-teams (Boyko & Holoborodko, 2024). Within these new
configurations, each core production function (e.g., systems analysis, project management,
software design and development) is led by a domain expert fully equipped with automation
tools capable of handling all routine tasks. It is important to note that the transition from in-office
DOI: https://doi.org/10.71112/pahqyx95
163 Multidisciplinary Journal Epistemology of the Sciences | Vol. 3, Issue 3, 2026, July–September, Special Edition
to hybrid or fully remote work formats, a trend gaining significant momentum since 2020, has
also played a critical role in this paradigm shift (Pashchenko, 2024). These new work modalities
have profoundly altered established principles of team formation, hiring, and termination
practices in the IT sector.
Concurrently, the role of AI technologies underpinning the concept of AI-augmented
software engineering (Panetta, 2023) has become increasingly evident and transformative
within this evolving context. The ability of IT companies to successfully adapt to these rapidly
changing conditions has become essential for maintaining competitiveness (Cakmak, 2023).
While this shift in the organizational and production paradigm has already begun, investment in
its core trajectories demands a careful and balanced approach, thus necessitating further
applied research.
This study sets out to conduct a fundamental assessment of the potential of AI tools to
support the execution of a practical software development project from its earliest stages such
as the initial idea or concept - through to the product’s market release and consumer adoption.
The research hypothesis posits that all critical artifacts of a software product including the
technical specification, source code, and user documentation can be generated by AI tools with
a sufficient level of quality and minimal labor input from individual engineers within the project
team. Accordingly, the hypothesis assumes that the selected AI tools and the associated
practices employed by the team have reached a level of operational maturity that allows them to
function as comprehensive automation instruments for the everyday tasks of virtually every
project team member. The proposed research problem aims to demonstrate the real-world
capabilities and practical applications of AI tools in the core production functions of a software
development project.
To ensure an adequate user experience during the implementation and utilization of AI
technologies in software engineering, the findings of the author’s earlier scientific study, “Early
DOI: https://doi.org/10.71112/pahqyx95
164 Multidisciplinary Journal Epistemology of the Sciences | Vol. 3, Issue 3, 2026, July–September, Special Edition
Formalization of Large Language Model Utilization in Software Development” (mid-2023), were
thoroughly analyzed (Pashchenko, 2023). Additionally, insights from a subsequent author-led
study, “AI Tools in Software Production – Demands and Barriers,” completed in the end of 2024
(Pashchenko, 2025), were employed to refine AI-assisted software development practices in
2025.
Both studies encompassed 60+ teams from IT companies, system integrators, and
banks with robust in-house software development capabilities, spanning geographical regions
from Russia and Kazakhstan to the United Kingdom and Spain. The teams represented diverse
segments of the software development industry, including:
• Independent software vendors, including in-house and product development (e.g.,
Miro, Google, Finastra, Finshape, Sber, VTB, Playrix, OZON, PSB, Deutsche Bank);
• Custom software development and outsourcing firms (e.g., Atos, SOFTEC, First Line
Software, and EPAM); System integrators (e.g., ThoughtWorks and Auxo);
• Other IT companies with complex economic models (e.g., Capgemini Engineering,
ARM, ZEISS Digital, and Ericsson).
Both studies employed a mixed-methods approach, utilizing Google Forms for structured
surveys alongside video interviews. The structured findings were subsequently distributed to
domain experts, allowing them to provide commentary and contribute to the final interpretations
of the research. The extensive panel of experts – 35 software development teams (Pashchenko,
2023), together with the wide geographic representation, provides a solid foundation for
asserting that the resulting conclusions reflect at least the advanced practices currently
prevalent in Europe (see Table 1).
DOI: https://doi.org/10.71112/pahqyx95
165 Multidisciplinary Journal Epistemology of the Sciences | Vol. 3, Issue 3, 2026, July–September, Special Edition
Table 1.
Experts in theoretical studies
No.
The characteristic
Representation in the expert’s study
1
Level of professional
experience in software
engineering
Less than 10 years
10-15 years
More than 15
years
Kästner & Kang, 2020
22 %
19%
59%
Gorban et al., 2018
11%
22%
67%
2
Region of described the
experience in usage of AI-
tools
North and West.
Europe (Spain,
Sweden, France,
Germany, Swiss,
UK, etc)
Central and
South Europe
(Poland, Czechia,
Hungary, Serbia
Bulgaria, Cyprus,
etc)
Eastern Europe
and CIS (Ukraine,
Russia, Armenia,
Turkey
Kazakhstan,
Georgia, etc)
Kästner & Kang, 2020
11%
33%
56%
Gorban et al., 2018
29%
20%
51%
3
Types of IT-business
Independent
software vendors,
including in-house
product
development
Custom
development and
outsourcing
software services
Other type of IT
business
Kästner & Kang, 2020
56%
33%
11%
Gorban et al., 2018
46%
29%
25%
Principal results and key findings of those studies (Pashchenko, 2025) had been used in
practical implementation of AI-augmented software development paradigm in 2023-2025 in
Android development teams of software company in Spain. Summary of the study’s results and
result of AI implementation (AI-Augmented Software Engineering paradigm) are given in
following sections of the article.
DOI: https://doi.org/10.71112/pahqyx95
166 Multidisciplinary Journal Epistemology of the Sciences | Vol. 3, Issue 3, 2026, July–September, Special Edition
METHODOLOGY
This study shows how theoretical result are forming the frame for practical
implementation of AI-Augmented Software Engineering paradigm in real software industry. The
goal of the study is defining the best practices of using AI instruments in real project of product’s
software development. Those best practices had been defined in mentioned above research
and made a solid base for practical implementation. The methodology of the research is based
on the following provisions:
1. Decomposition, system analysis, synthesis - for processing the results of theoretical
research and collecting promising methods and approaches for implementing
innovation;
2. Change management and project management for implementing AI tools in software
development processes;
3. Empirical assessment and system analysis of production indicators for analyzing
intermediate and final results of implementing innovation.
The description of the process of the implementation of AI tools in the practice of
implementing a software project requires delving into the details of the study:
1. Timeframe: The software development project was executed from July 2023 to April
2025, and involved a geographically distributed production team of five people from
the IT company Slavasoft. The development was carried out in the Scrum production
paradigm (Misra et al., 2012).
2. Software product: Android-app for smartwatch branded as «AI Alter Ego»,
implementing the functions of a digital personal assistant in a smartwatch. During
the project, three major releases of the system were released. Three releases of the
system in 2024-2025 fully implemented all the intended functionality of the system for
the Google Wear 3, 4 and 5 operating system installed on smartwatches of all
DOI: https://doi.org/10.71112/pahqyx95
167 Multidisciplinary Journal Epistemology of the Sciences | Vol. 3, Issue 3, 2026, July–September, Special Edition
leading global manufacturers (except Apple) (Barai & Mutreja, 2023). The Alter Ego
software product is available for download from the Google.Market app store for free
and without geographic restrictions.
3. Artifacts which are necessary for the release of this software product:
• System vision (business requirements level),
• Functional specifications,
• Architectural diagrams, such as component and Archimate-TOGAF (Wierda,
2021),
• Software code,
• Test plan with test cases,
• User and service documentation for the software product.
4. As an AI tool that implements the tasks set for the development team, the ChatGPT
service version 3.5 (2023) and version 4.0 (2024-2025) from OpenAI, which has
gained enormous worldwide popularity since the fall of 2022 (Ahlgren, 2023).
Thus, this study sets a relevant scientific task of fundamentally assessing the capabilities
of an engineering team to create all key artifacts of a software project using an artificial
intelligence tool that is available on a permanent free basis. The results of two applied studies
were used as a theoretical basis necessary for managing the implementation of AI tools. A brief
summary of these results is given in the next section of the article.
RESULTS
By the end of 2024 there were clearly observed: innovators, early adopters, early
majority, late majority, and laggards (Rogers, 2003). An interesting observation is that the
introduction of AI tools into the actual practice of software production follows this classification:
DOI: https://doi.org/10.71112/pahqyx95
168 Multidisciplinary Journal Epistemology of the Sciences | Vol. 3, Issue 3, 2026, July–September, Special Edition
• Categories with "innovators" and "early adopters" according to E. Rogers were
formed and the formalization of the use of AI tools began (project teams carry out a
corporate plan for introducing AI into software development, create and use
centralized corporate policies and/or recommendations);
• The process of forming the category of "early majority" according to E. Rogers is
finished to 2025 - teams finished the studying AI tools in various ways (R&D,
individual / team experiment, etc.) and started its implementation in software
development.
Around 23% of experts are using AI tools in their regular job with the high frequency and
it has a strong impact on their personal tasks in IT projects. Then in the end of 2024 the study
shown the rising of the share (to 35%) of teams and organizations already has started the
implementation of AI-tools in real software production (in different process areas like coding,
testing, etc) and it’s planned to do in near future for 48% of teams. Development in the real
practice allows:
• Automation of the routine operations and time saving;
• Speed up of the operations in the team \ organization;
• Software product excellence, including software quality, UX and documentation.
• Also both studies had shown that AI-tools are not in high demand for business and
system analysis, but there were much more popular types of tasks for AI-tools:
coding, code reviews, fast software prototyping.
About 63% of experts in study used the AI-tools in making software code and were
satisfied with the reached result. Both studies demonstrated that impact of AI-tools on software
quality management is rising and there were defined most popular types of tasks for LLMs in
software quality management: writing the auto-tests, managing the defects and reports
DOI: https://doi.org/10.71112/pahqyx95
169 Multidisciplinary Journal Epistemology of the Sciences | Vol. 3, Issue 3, 2026, July–September, Special Edition
analysis;searching of the errors and vulnerabilities in the code. But still around the half of
experts in study (Pashchenko, 2025) does not use AI-tool in software quality management.
Expert panel estimated the value and the role of LLMs in learning and in the excellence
of the software development skills in the end of 2024:
• The impact of usage LLMs in professional learning is very high – 41% of experts;
• It’s just one more useful tool on the board – 44% of experts.
For sure, the implementation of LLMs has their specific features and risks, that we need
to estimate before the implementation of AI-tools in software production. Expert’s panel in study
figured the main barriers in the implementation of any AI-instruments in the software
development in their teams and organizations:
• High level of different risks - from legal aspects to ethical – 44% of experts;
• Lack of the resources (money, time, knowledge, HR-capital) – 30% of experts;
• Organizational resistance of engineers and managers – 18% of experts.
But those risks are not a solid barrier – more and more IT companies are in the active
process of AI-tools implementation in the software engineering practices. From 12% of teams in
study to 26% of teams in study are executing an official plan in how to use AI-instruments in IT
projects, and in around 30% of teams its usage is continuing in test mode in their teams without
centralized management.
So, study (Pashchenko, 2025) confirmed the earlier results from:
• Innovator’s and early adopter’s groups are formed in software development domain;
they had started the AI-tools implementation in software production.
• Current AI-instruments usage is focused on the working with the software code (in
different kind of ways), on the quality assurance and on the tasks about
documentation.
DOI: https://doi.org/10.71112/pahqyx95
170 Multidisciplinary Journal Epistemology of the Sciences | Vol. 3, Issue 3, 2026, July–September, Special Edition
• For some IT companies the lack of centralized efforts on the corporate level might
lead to the missing of the competition advantage in the software production,
connected with AI tools.
• Software engineers need new skills in AI-human interaction: as earlier company will
start educate them as more effective they will be in future production process.
This leads to some considerations about best practices in AI tools implementation. First,
focus AI usage in software engineering on main process areas in the production, start with all
activities where software code and product documentations are the expected result. Second,
formalization of the process of the implementation of AI into software engineering is a key factor
in overcoming all barriers. And third, rising of interest to AI from industry and governments lead
to the need of complex risk management in internal projects of implementation of AI.
Moreover, the above results have identified a set of practical questions for the current
study, which are formulated in Table 2. This set of questions allows us to solve the scientific
problem posed above and test the proposed research hypothesis for the full software life cycle:
from the earliest stages (idea, concept) to business and system requirements collected in the
technical task, to system design and creation of program code, then to product implementation
and to its release into industrial environments accessible to end users.
Table 2.
Areas of software development and corresponding questions
№
Software Engineering
Process Area
Current Research Question
1
System Analysis
How can an AI tool be practically useful in creating a technical
specification at all stages: ¿from decomposition and analysis of
requirements to its documentation?
DOI: https://doi.org/10.71112/pahqyx95
171 Multidisciplinary Journal Epistemology of the Sciences | Vol. 3, Issue 3, 2026, July–September, Special Edition
2
Software Design and
Construction
How can an AI tool be practically useful in designing a product
architecture, quickly prototyping functions and creating stable working
software code?
How quickly does an AI tool fail when the logic is significantly
complicated and there is a need to support the related software code
of several modules or services?
3
Software Quality
Management
How can an AI tool be practically useful in software quality
management: ¿from creating automated tests to generating test
cases for manual testing?
4
Software Content
Creation
¿Can an AI tool create diverse content, high-quality and expert in the
subject area of automation?
5
Software
Documentation
Can an AI tool create detailed and accurate project documentation
based on software code and requirements - user manuals, service
documents, ¿etc.?
Also, to solve the scientific problem of this study, it is necessary to take into account the
results of studies, namely, what difficulties and barriers exist in scenarios for the implementation
of AI tools in the activities of a software development company (Sattelberg, 2023; Glinka &
Raed, 2009). Of course, the implementation of AI tools has its own characteristics and risks that
must be managed on an ongoing basis.
Practical case of AI implementation in software engineering
This section of the study, focused on the practical case of AI tool implementation, details
the process and outcomes of the standardized use of ChatGPT 3.5 (in 2023) and 4.0 (in 2024)
within a software development workflow based on the Scrum methodology. The software
company SlavaSoft initiated the development of the “Alter Ego” application for Android-based
smartwatches from scratch, beginning with a product "vision" document. Since late 2023, a
dedicated team of five engineers has employed the AI tool to design software and generate all
related project documentation.
DOI: https://doi.org/10.71112/pahqyx95
172 Multidisciplinary Journal Epistemology of the Sciences | Vol. 3, Issue 3, 2026, July–September, Special Edition
The adoption of AI tools within the company began with this team and, starting in 2023,
all new innovation initiatives have been structured as “pilot projects”. The implementation was
guided by an official change management plan and aligned with the company’s product
roadmap. Even prior to the project launch, the team conducted several meetings to discuss AI
tool usage for software coding and testing in light of previous research findings (Pashchenko,
2023) y (Pashchenko, 2025), with the objective of developing internal best practices. In the end
of 2023 were built official policy – how to use AI tools for solving typical tasks in software
production processes, including code, tests and documents generation.
Best practices in particular software project are inherently dependent on the type of
software product under development. Accordingly, understanding the complexity and associated
risks of the project, as well as estimating development and quality assurance efforts, is critical.
“Alter Ego” is a sophisticated Android application with over 50 features, designed to deliver
immediate AI-based care and assistance via a smartwatch interface. As stated in the product
specification: “Alter Ego” is the first deeply human-centered AI assistant built exclusively for
Wear OS smartwatches”. This positions “Alter Ego” as a feature-rich and complex application,
with strong emphasis on smartwatch-specific usability and interface design.
It is important to note that the selection of ChatGPT as the primary AI tool for this project
followed a careful evaluation process. The main considerations included:
• The availability of a free subscription account enabling full-day work on code and
documentation generation (for version 3.5);
• The rapid advancement of the tool, along with its versatility across nearly all tasks
associated with software development.
Other AI tools, such as Microsoft Copilot (in 2023) and Cursor (in 2025), which are
integrated into software development environments, were also considered. However, the
DOI: https://doi.org/10.71112/pahqyx95
173 Multidisciplinary Journal Epistemology of the Sciences | Vol. 3, Issue 3, 2026, July–September, Special Edition
aforementioned advantages of ChatGPT proved more compelling for the project team than
factors such as ease of integration or response speed.
It should also be emphasized that ChatGPT was not integrated into the development
environment (e.g., Android Studio for code or Atlassian Jira for task and requirements
management). All interactions with the AI tool occurred within its native interface, with dialogues
stored on the platform’s side. Key outputs were manually transferred into project artifacts.
The general algorithm for AI tool usage was consistent across tasks and included the
following steps:
1. Construction of a precise and detailed prompt describing the task in terms of
system requirements;
2. Verification of the response and, where necessary, refinement through iterative
prompting in dialogue mode;
3. For code generation tasks, prompts were often decomposed to ensure each
response was limited to 300–400 lines of code in 2023 and 600-700 lines of code in 2025,
supporting consistency and logical coherence throughout the engineer-AI interaction;
4. In some instances, entire AI-generated responses were re-submitted as input for
further refinement or correction.
Once verified and finalized, AI-generated outputs were incorporated into project artifacts
such as technical specifications, working system code, user documentation, and test cases. For
software code in Java and XML, the finalized outputs were integrated into the corresponding
modules of the Android project within the Android Studio environment. Over the course of
development, both the codebase and the software service structure evolved. Nevertheless, by
2025, only two of the more than 50 system services (specifically those comprising the core
business logic) did not contain any AI-generated code.
DOI: https://doi.org/10.71112/pahqyx95
174 Multidisciplinary Journal Epistemology of the Sciences | Vol. 3, Issue 3, 2026, July–September, Special Edition
One notable limitation identified was the AI tool’s restricted ability to maintain logical
coherence across large codebases. Empirical observations indicated that version 4.0 could
successfully interpret and manage interdependencies across services sharing common data,
with a practical upper limit of approximately 700 lines of code per response. Attempts to
generate abstract-level inter-service logic (i.e., based on semantic rather than structural or data
relationships) consistently failed to meet the project team's quality expectations. It is also
relevant that the team did not utilize collaborative AI tools. Each engineer independently
interacted with the AI tool on a prompt-response basis. However, the team regularly reviewed
and discussed generated artifacts, especially documentation, and to a lesser extent, source
code. The outputs from ChatGPT were manually processed and refined by engineers and
served as the foundation for all major project artifacts, including technical specifications,
program code (Java, XML), test cases, and product documentation.
Simultaneously, certain critical artifacts - such as the business vision document (defining
high-level business requirements) and architectural diagrams (covering component, application,
and data views in accordance with the ArchiMate (TOGAF) framework (Wierda, 2021) - were
created exclusively by the engineering team, without the usage of AI tools.
Although the initial technical specification was fully aligned with the business vision at
the project's inception, new ideas and requirements emerged throughout the development
cycle. These were rapidly prototyped and incrementally incorporated into the evolving technical
documentation. This approach aligns with established industry practices wherein requirements
management continues throughout the software development lifecycle (Sommerville, 2009).
The overall results of using ChatGPT in the project are shown in Table 3 (collected
from project documentation and interviews with engineers):
Table 3
AI-tools usage findings from project Alter Ego
DOI: https://doi.org/10.71112/pahqyx95
175 Multidisciplinary Journal Epistemology of the Sciences | Vol. 3, Issue 3, 2026, July–September, Special Edition
№
Software
Engineering
Process Area
Current Study
Question (from
Table 2)
ChatGPT 3.5 Usage
Results (2023-2024)
ChatGPT 4 Usage
Results (2024-2025)
1
System
analysis
How can an AI tool
be practically useful
in creating a
technical
specification at all
stages: from
decomposition and
analysis of
requirements to its
documentation?
Exceeds team
expectations: AI creates a
technical task of any
complexity, based on
business requirements from
the project team. The
result: a structured
document with a sufficient
level of nesting and detail
for the practical
development of a software
product.
Exceeds team
expectations: AI creates a
technical task of any
complexity based on
business requirements
from the project team.
The result: a structured
document with a sufficient
level of nesting and detail.
AI itself suggests
additional sections and
finds contradictions in
system requirements.
2
Software
product design
and
construction
How can an AI tool
be practically useful
in designing a
product
architecture, quickly
prototyping
functions and
creating stable
working software
code?
How quickly does
an AI tool fail when
the logic is
significantly
complicated and
there is a need to
support the related
software code of
several modules or
services?
Architecture design – below
the team's expectations,
since the tool does not
support graphical
visualization of the
component diagram,
however, at the text level, it
describes the composition
of classes and their
interaction with each other
and, for example, with the
database - correctly and in
detail.
Rapid prototyping (under
conditions of unclear or
incomplete requirements) -
above expectations, in
about 80% of cases, the AI
tool built a correct prototype
of the future function at the
Architecture design –
below the team’s
expectations, still no
visualization support. At
the text level, it
decomposes in detail and
in a coherent manner,
selects technological
priorities for
implementation.
Rapid prototyping (under
conditions of unclear or
incomplete requirements)
– above expectations, in
about 95% of cases, the
AI tool built a correct
prototype of the future
function at the program
code level on the first try.
DOI: https://doi.org/10.71112/pahqyx95
176 Multidisciplinary Journal Epistemology of the Sciences | Vol. 3, Issue 3, 2026, July–September, Special Edition
program code level on the
first try.
Working versions of the
program code - meets the
team's expectations: the
more detailed the
requirements for the code
were described, the more
accurate and logically
consistent the result was.
All code was verified
multiple times in the
project, both by engineers
and with the help of
ChatGPT.
The ChatGPT tool supports
the logical structure and
consistency of the program
code only for closely
related services
(integrated) and if all the
service code is loaded into
the dialog with the tool.
Working versions of the
program code – meets
the team’s expectations:
the more detailed the
requirements for the code
were described, the more
accurate and logically
consistent the result was.
All code was verified
multiple times in the
project, both by engineers
and with the help of
ChatGPT.
As before, the ChatGPT
tool supports the logical
structure and consistency
of the program code only
for closely related
services (integrated) and
if all the service code is
loaded into the dialog with
the tool.
3
Software quality
management
How can an AI tool
be practically useful
in software quality
management: from
creating automated
tests to generating
test cases for
manual testing?
The team had a
controversial impression of
the automated tests; in
some cases, the automated
test logic was not
implemented in the best
way even with multiple
adjustments.
Test cases for manual
testing – meet the team's
expectations, although a
specialist in the team
manually significantly
Autotests - meet the
team's expectations, are
created correctly and
logically consistent.
Test cases for manual
testing - meet the team's
expectations, although a
specialist in the team still
manually significantly
supplemented them to the
working version.
DOI: https://doi.org/10.71112/pahqyx95
177 Multidisciplinary Journal Epistemology of the Sciences | Vol. 3, Issue 3, 2026, July–September, Special Edition
supplemented them to the
working version.
4
Software
content creation
Can an AI tool
create diverse
content, high-
quality and expert
in the subject area
of automation?
The content of the
application (expert data on
the field of automation, lists
of standard values,
dictionaries, etc.) is beyond
any expectations. When
formulating precise queries,
the content was created in
any required volume while
maintaining high data
quality.
Exceeds all expectations
of the team. The
emergence of the visual
content generation
function further expanded
the capabilities of the
development team in
creating the application.
When formulating precise
requests, content was
created in the required
volume while maintaining
high data quality.
5
Software
documentation
Can an AI tool
create detailed and
accurate project
documentation
based on software
code and
requirements - user
manuals, service
documents, etc.?
User and service
documents - meet the
team's expectations.
Moreover, the AI tool was
equally good at creating
voluminous instructions for
users and short service
documents on ready-made
java code.
Any user documents -
exceed any team
expectations. AI itself
suggests missing sections
and corrects the
document structure.
It is also worth describing the use of the AI tool for other tasks that turned out to be relevant
for this software development project. Thus, when significant engineering problems arose for the
development and testing environments (including those involving real smartwatches from various
manufacturers (Chuah et. al., 2016)), the team turned to ChartGPT for technical advice. However,
due to the limitation of this tool in having the most up-to-date information, all the answers were
template-based and significantly inferior to the team's current knowledge. The experience of
consulting on placing an application in the Google Play Market store was also unsuccessful:
Google's tightening of policies on placing and updating applications for smartwatches since 2021
DOI: https://doi.org/10.71112/pahqyx95
178 Multidisciplinary Journal Epistemology of the Sciences | Vol. 3, Issue 3, 2026, July–September, Special Edition
made the publication process itself more complex (Sattelberg, 2023). Attempts to find a solution
to constantly emerging operational issues using the AI tool were also unsuccessful - the data
inside ChatGPT turned out to be outdated and irrelevant in this area.
Thus, the capabilities of AI tools (using ChatGPT 3.5 and 4 as an example) in
implementing a practical software development project should be assessed as sufficient. The
research hypothesis has been confirmed: all significant software product artifacts – functional
specifications, software code, user documentation in a real software development project can be
generated by AI tools with minimal effort from the relevant engineers in the project team. This
means that the scientific task has been solved, and the AI tool chosen by the team - ChatGPT
from OpenAI has already reached the required high level of efficiency for the fundamental tasks
of software engineering and is an effective tool for automating project activities.
In the end of the description of the results of the implementation of AI tools, several
conclusions should be made about their most effective usage. These provisions were formulated
from the results of retrospectives (meetings of engineers to discuss the work done and the
problems solved) after development sprints and they are categorized by areas of application in
software engineering.
1) Creating a technical specification based on business-level requirements (e.g.
based on a business vision document) is most effectively done for system services, which implies
the need to identify individual services during design at the earliest design stages. The ChatGPT
tool generates system and technical (non-functional) requirements much more effectively if a
service oriented architecture is applied (Glinka & Raed, 2009) and the business sense of its
operation is specified for each service (e.g. in the form of value for the end user).
2) Generating system code (both during rapid prototyping and for the working version
after analyzing and coordination of the requirements) is an iterative process, it’s aligned with other
research (Ridi, 2024). Persistence in clarifying prompt requests and high structuring and detailing
DOI: https://doi.org/10.71112/pahqyx95
179 Multidisciplinary Journal Epistemology of the Sciences | Vol. 3, Issue 3, 2026, July–September, Special Edition
of system requirements are of decisive importance. In some cases, the ChatGPT tool behaves
non-deterministically: sometimes it misses some of the requirements, sometimes it “invents”
additional business logic. Persistence in repeating precise and detailed requests leads to success
in 95% of cases, even if the first answer of the AI tool was completely inaccurate. The approach
of sending the generated code back to the ChatGPT dialog with additional instructions worked
well; within 1-3 additional iterations, the tool produced fully working code that met the initial system
requirements.
3) Best working prompts for code and documents generation had been done in formal
structure:
a. Role and Context (for AI agent in task of prompt),
b. Step-by-step algorithm (how AI should do the task);
c. Output format (language, format, structure, etc);
d. Final instruction (additional remarks, including level of creativity for AI).
4) Auto-tests and test cases for manual testing as part of ensuring high quality of the
software product were generated based on the existing system code. It’s aligned with modern
research (Peng et. al., 2023) and (Petrovic et. al., 2024)), but not very common in software
development. Of course, these artifacts were verified and refined (expanded and clarified) by the
software quality assurance engineer, but in his opinion, the results could have been immediately
used in the project if there had been such a production need. This is an obvious paradox, since
the purpose of testing in software engineering is to check the code for compliance with
requirements, but the team managed to obtain good quality testing artifacts immediately based
on the code and even identify code sections (without testing) that had previously been
implemented not quite accurately. This conclusion is in deep contradiction with traditional
approaches to software quality assurance and requires further reflection: if valid (in the opinion of
DOI: https://doi.org/10.71112/pahqyx95
180 Multidisciplinary Journal Epistemology of the Sciences | Vol. 3, Issue 3, 2026, July–September, Special Edition
the team) test artifacts are obtained based on the code (and not on system requirements and the
work of the corresponding engineer), can they help to find defects at all? According to the author,
the reason for this paradox in a specific project is the high atomicity of test cases and the simplicity
of the business logic of the software product.
5) User documentation for the already generated code of the working system was
completely created in ChatGPT in 2024. The result of the work was so good that the technical
writer involved in the team part-time only had to add illustrations (screenshots of the system by
functions).
6) Creating application content using AI using the Alter Ego software product as an
example turned out to be the fastest of all existing methods. The team did not need any help from
external specialists to create 100% of the application content in the AI tool. The emergence of the
ability to generate visual objects in ChatGPT version 4 also made it possible to automate some
of the marketing functions for the application (marketing leaflets, graphics for the website). It’s
fully aligned with results of studies (Pashchenko, 2023) and (Naimi, 2024) ;
7) Creating a software product is the implementation of a certain business logic (and
not just a set of loosely coupled application functions). ChatGPT understands the most complex
logic (both at the level of the requirements text and at the level of software code) and can offer its
improvement according to the specified requirements of the project team as it was defined earlier
in (Kurnianingrum, 2024). At the same time, the generation of new ideas for AI tools in terms of
business logic is template-based and clearly uncreative, i.e. seriously inferior to the capabilities
of any specialist in the field of software engineering. The team doubts that the tool used could
create the original business logic of the new software product without serious and significant
human involvement. Those doubts in practical project are aligned with results of study (Shiye &
Chien-Ming 2022).
DOI: https://doi.org/10.71112/pahqyx95
181 Multidisciplinary Journal Epistemology of the Sciences | Vol. 3, Issue 3, 2026, July–September, Special Edition
8) No one in the project team had any doubts that the selected AI tool allows
performing personal tasks in the areas of software engineering much faster and more efficiently
than without this automation. There was no practical sense for the project team to calculate the
percentage of increasing efficiency or economic feasibility, because after only two months of using
the selected tool, the team's work had changed significantly: not a single member of the project
team wanted to work on the project without this tool anymore.
CONCLUSIONS
Applied research has identified the implementation features and key aspects of using
artificial intelligence tools in the practice of software development companies. Research in
2023-2024 have shown that a significant proportion of organizations have already begun
implementing AI tools in real software production. The use of AI in software development
automates typical and routine operations and is in demand due to the need to speed up
software development and improve the quality of the final product. Although experts from
(Pashchenko, 2023), (Pashchenko, 2025) noted that by the end of 2024, AI tools were not in
great demand for solving business and system analysis problems, they were nevertheless in
significant demand in writing working code, rapid software prototyping, ensuring software quality
management and some other areas.
As part of this study, the scientific problem of determining the fundamental capabilities of
AI tools in creating key artifacts of a software project at all stages of its life cycle was solved -
from technical specifications to working code and accurate user documentation for the software
product. The process of using the AI tool by the development team was algorithmic and
standardized in corporate policy: from formulating precise requests to the selected ChatGPT
tool to applying the received and verified answers in practice. The development team engineers
DOI: https://doi.org/10.71112/pahqyx95
182 Multidisciplinary Journal Epistemology of the Sciences | Vol. 3, Issue 3, 2026, July–September, Special Edition
worked with the AI tool individually, but the final results were regularly discussed and processed
by the entire team.
The ChatGPT tool demonstrated high efficiency in creating technical specifications, rapid
prototyping of functions and generating software code. Importantly, requirements management
was continued throughout the project, which allowed for flexible adaptation to new ideas and
constant (albeit small) changes in the area of business requirements. The AI tool was used for
engineering consultations, creating technical specifications, generating system code, developing
automated tests and test cases, compiling user documentation and creating application content.
As a result, the following conclusions were made:
• Using AI to create technical specifications and user documentation proved to be
effective, especially when working with service architecture and detailed description
of business value for the user when using each service;
• Generating system code using AI was successful, but it was iterative, in which
persistence in interaction with AI is important (constant repetition of context and
requirements management in prompts). There are noticeable limitations in
“remembering” by the AI tool of the logical connectivity of individual services in the
system;
• AI-powered generation of application content was fast and efficient, requiring no
additional assistance from external specialists. At the same time, AI's capabilities in
creating the product's business logic were limited and required significant human
intervention to generate original ideas.
• Overall, the use of ChatGPT met the project team's expectations in most respects,
although there were moments that required additional attention, intellectual effort,
and adjustments. Based on the results obtained, it should also be concluded that the
DOI: https://doi.org/10.71112/pahqyx95
183 Multidisciplinary Journal Epistemology of the Sciences | Vol. 3, Issue 3, 2026, July–September, Special Edition
use of AI in software product development is an effective tool, although it requires the
use of certain skills, control, and regular verification of results by a software engineer.
Conflict of interest declaration
The author declares no conflict of interest for this research.
Author Contributions Statement
Denis Svyatoslavovich Pashchenko: Conceptualization, writing – original draft.
AI use declaration
The author declares that artificial intelligence was used as a support tool for this article
and for data analysis, given the nature of the topic. However, this tool does not in any way
replace the intellectual task or process. After rigorous reviews using different tools, which
verified that no plagiarism exists as evidenced in the supporting documentation, the author
states and acknowledges that this work is the result of his own intellectual effort and has not
been written or published on any electronic or AI platform.
REFERENCES
Gorban, A., Grechuk, B., & Tyukin, I. (2018). Augmented artificial intelligence: A conceptual
framework. arXiv. arXiv:1802.02172 [cs.AI].
Kästner, C., & Kang, E. (2020). Teaching software engineering for AI-enabled systems. En
Proceedings of the ACM/IEEE 42nd International Conference on Software Engineering:
Software Engineering Education and Training (pp. 45–48).
https://doi.org/10.1145/3377814.3381714
Barenkamp, M., Rebstadt, J., & Thomas, O. (2020). Applications of AI in classical software
engineering. AI Perspectives, 2, 1. https://doi.org/10.1186/s42467-020-00005-4
DOI: https://doi.org/10.71112/pahqyx95
184 Multidisciplinary Journal Epistemology of the Sciences | Vol. 3, Issue 3, 2026, July–September, Special Edition
Boyko, O., & Holoborodko, Y. (2024). Distributed software development in 2025: All you should
know. SPD Tech Blog. https://spd.tech/dedicated-development-teams/distributed-
software-development-team/
Pashchenko, D. S. (2024). Consolidation of a new organizational and production paradigm in
software development projects. Information Technologies, (3), 150–158.
https://doi.org/10.17587/it.30.150-158
Panetta, K. (2023). Set up now for AI to augment software development. Gartner Report.
https://www.gartner.com/en/articles/set-up-now-for-ai-to-augment-software-development
Cakmak, Z. (2023). Adapting to environmental change: The importance of organizational agility
in the business landscape. Florya Chronicles of Political Economy, 9, 42–53.
https://doi.org/10.17932/IAU.FCPE.2015.010/fcpe_v09i1004
Pashchenko, D. S. (2023). Early formalization of AI-tools usage in software engineering in
Europe: Study of 2023. International Journal of Information Technology and Computer
Science, 15(6), 29–36. https://doi.org/10.5815/ijitcs.2023.06.03
Pashchenko, D. S. (2025). Crecimiento en la demanda de herramientas de inteligencia artificial
en ingeniería de software: Resultados de un estudio paneuropeo 2024. Revista de
Investigación en Tecnologías de la Información, 13(29), 82–91.
https://doi.org/10.36825/RITI.13.29.008
Misra, S. C., Kumar, V., Kumar, U., Fantazy, K. A., & Akhter, M. (2012). Agile software
development practices: Evolution, principles, and criticisms. International Journal of
Quality & Reliability Management, 29, 972–980.
DOI: https://doi.org/10.71112/pahqyx95
185 Multidisciplinary Journal Epistemology of the Sciences | Vol. 3, Issue 3, 2026, July–September, Special Edition
Barai, T. V., & Mutreja, S. (2023). Smartwatch market: Global opportunity analysis and industry
forecast, 2023–2032. Allied Market Research.
https://www.alliedmarketresearch.com/smartwatch-market
Ahlgren, M. (2023). OpenAI statistics & facts for 2023 (DALL·E, ChatGPT & GPT-3.5).
https://www.websiterating.com/research/openai-statistics/
Rogers, E. (2003). Diffusion of innovations (5th ed.). Simon & Schuster.
Wierda, G. (2021). Mastering ArchiMate Edition 3.1: A serious introduction to the ArchiMate
enterprise architecture modeling language. R&A.
Sommerville, I. (2009). Software engineering (9th ed.). Addison-Wesley.
Chuah, S., Rauschnabel, P. A., et al. (2016). Wearable technologies: The role of usefulness and
visibility in smartwatch adoption. Computers in Human Behavior, 65, 276–284.
https://doi.org/10.1016/j.chb.2016.07.047
Sattelberg, W. (2023). Google is updating app guidelines for smartwatches just in time for Wear
OS. Android Police. https://www.androidpolice.com/2021/09/03/google-is-updating-app-
guidelines-for-smartwatches-just-in-time-for-wear-os-3/
Glinka, F., & Raed, A. (2009). A service-oriented interface for highly interactive distributed
applications. En European Conference on Parallel Processing (Lecture Notes in
Computer Science, Vol. 6043, pp. 266–277). https://doi.org/10.1007/978-3-642-14122-
5_31
Ferdiana, R. (2024). The impact of artificial intelligence on programmer productivity. En
International Conference on Software Engineering and Information Technology
(ICOSEIT) 2024 (Vol. 2).
DOI: https://doi.org/10.71112/pahqyx95
186 Multidisciplinary Journal Epistemology of the Sciences | Vol. 3, Issue 3, 2026, July–September, Special Edition
https://www.researchgate.net/publication/378962192_The_Impact_of_Artificial_Intelligen
ce_on_Programmer_Productivity
Finio, M., & Downie, A. (2024). La IA en el desarrollo de software. IBM Review.
https://www.ibm.com/es-es/think/topics/ai-in-software-development
Peng, S., Kalliamvakou, E., Cihon, P., & Demirer, M. (2023). The impact of AI on developer
productivity: Evidence from GitHub Copilot. https://arxiv.org/pdf/2302.06590.pdf
Petrovic, N., Lebioda, K., Zolfaghari, V., Schamschurko, A., Kirchner, S., & Purschke, N. (2024).
LLM-driven testing for autonomous driving scenarios. En Proceedings of the 2024 2nd
International Conference on Foundation and Large Language Models (FLLM) (pp. 173–
178).
Naimi, L., Bouziane, E., et al. (2024). Automating software documentation: Employing LLMs for
precise use case description. Procedia Computer Science, 246, 1346–1354.
https://doi.org/10.1016/j.procs.2024.09.568
Kurnianingrum, D., Jumbri, I., et al. (2024). Exploring the Chat GPT's impact and prospects for
business research purposes. 1–6.
https://doi.org/10.1109/INOCON60754.2024.10511483
Cao, S., & Huang, C.-M. (2022). Understanding user reliance on AI in assisted decision-making.
Proceedings of the ACM on Human-Computer Interaction, 6(CSCW2), 1–23.
