Integrating AI Tools Into the Continuous Testing Process.

The American Journal of Engineering and Technology

222

https://www.theamericanjournals.com/index.php/tajet

TYPE

Original Research

PAGE NO.

222-229

DOI

10.37547/tajet/Volume07Issue06-23

OPEN ACCESS

SUBMITED

18 April 2025

ACCEPTED

24 May 2025

PUBLISHED

30 June 2025

VOLUME

Vol.07 Issue 06 2025

CITATION

Vladyslav Korol. (2025). Integrating AI Tools Into the Continuous Testing
Process. The American Journal of Engineering and Technology, 7(06), 222

–

229. https://doi.org/10.37547/tajet/Volume07Issue06-23

COPYRIGHT

© 2025 Original content from this work may be used under the terms
of the creative commons attributes 4.0 License.

Integrating AI Tools Into
the Continuous Testing
Process.

Vladyslav Korol

Software Developer Engineer in Test Penn Entertainment
Miami, USA

Abstract:

This paper utilizes AI tools to enhance the

ongoing test cycle in a DevOps environment, thereby
creating Metabase. This data architecture is robust and
scalable, supporting a highly responsive release process.
The project is vital since the releases have become more
frequent; standard automation has already reached its
limit, increasing the costs of maintaining scripts and,
consequently, resulting in a significantly higher total cost
due to the discovery of many more defects. The novelty
of this work is grounded in an approach to choosing and
applying AI tools through comparative analysis over
available commercial and open-source platforms,
supported by content analysis of empirical use cases and
quantitative assessments from industry reports. Herein,
a methodology is presented that consists of
architectural solutions for data lake organization,
continuous model training scenarios, and ML endpoints
integrated into the CI/CD pipeline, which hosts
Predictive Test Selection, as well as self-healing and test-
case prioritization mechanisms. It narrates the creation
of the prompt-engineer position and the connections
between QA/ML experts and organizational facets. This
paper discusses the application of clear AI measures for
risk assessments. The final calls shown here indicate that
intelligent automation enables reducing a regression set
to a barely necessary size while maintaining 99.9% bug
detection and minimizing false alert failures. This, in
turn, leads to improvements in MTTR and TCO quality.
TestPi

lot and FlakeFlagger verify Meta’s practices;

furthermore, it is anticipated that forecasts will retain a
CAGR of 20.9% in the global AI testing market growth.
Solution maturity, encompassing both SaaS and on-
premises models, offers a flexible choice to regulated
industries. Metabase architecture is shown in which raw

The American Journal of Engineering and Technology

223

https://www.theamericanjournals.com/index.php/tajet

and processed data are kept separately to ensure timely
model retraining as well as to minimize computational
costs. This article will be helpful for software architects,
QA managers, DevOps teams, and ML engineering
specialists involved in building scalable and resilient
testing architectures.

Keywords:

artificial intelligence, continuous testing,

DevOps, self-healing, Predictive Test Selection, data
lake, ML integration, software quality, Metabase.

INTRODUCTION

The transition to a DevOps model has made delivery
speed the primary competitive factor, with continuous
testing serving as its built-in safeguard. Today, 83% of
developers already participate in DevOps practices, and
every build undergoes a chain of automated checks to
prevent defects from leaking into production at high
release velocities [1]. Traditional automation no longer
copes with the increased frequency of changes. Such
routine tasks consume up to half of the entire
automation budget, turning the test suite into a
bottleneck of the pipeline and forcing teams either to
slow down or to release code with reduced coverage.
This imbalance between development speed and test
inertia not only lengthens feedback loops but also
increases the total cost of defects detected late.

It is precisely at these points that a demand for smart
tools arises. The implementation of AI-oriented
platforms yields savings in the overall testing budget and
adds efficiency to test runs through more precise test-
case selection and extended coverage. Self-healing
mechanisms, capable of automatically correcting
selectors or API routes, reduce script maintenance costs,

freeing engineers’ resources for more valuable tasks.

The case studies on regression testing with AI
frameworks demonstrated a reduction in cycle duration
of up to 30%, directly accelerating release delivery
without compromising quality [2]. This, in turn, gives
one the ability to leverage artificial intelligence not as a
trendy extra, but as the natural and upcoming way in the
evolution of continuous testing, which would maintain
the balance between DevOps pipeline velocity and
product reliability.

MATERIALS AND METHODOLOGY

The study of AI tool assimilation for ongoing testing was

conducted through a review of 19 works, comprising
three academic papers [4

–

6], market reports [1, 3, 19],

and technical overviews [13-16]. Studies on Predictive
Test Selection [4] and LLM-driven test generation [5]
form the theoretical base. The research on self-healing
approaches to reducing flaky failure complements the
framework [6]. Industry reports have shown that
traditional automation loses effectiveness as release
frequency increases [1, 3] and that the AI-testing market
is rapidly developing [19]. Additional data were provided
on the functional capabilities of market leaders

—

Applitools, Testim, Functionize, Mabl, and Tricentis [8

–

12]

—

while the TestRail report [13] offered quantitative

evaluations of tool usage.

Methodologically, the work includes four components.
First, a comparative analysis of platforms: evaluation of
the depth of the AI layer (self-healing, generation, visual
regression) and licensing schemes (SaaS vs. on-
premises) for Applitools, Mabl, Testim, Functionize,
Tricentis, as well as open-source projects Retecs and
DeepOrder [8

–

12, 14]. Second, a Grey Literature Review

[7] allowed the identification of the prevalence of self-
healing and test-prioritization patterns among
commercial solutions. Third, a content analysis of
empirical cases reveals the effectiveness of Predictive
Test Selection at Meta [4], the median coverage of
70.2% by LLM-generated tests at TestPilot [5], and the
reduction of false failures using FlakeFlagger [6]. Fourth,
an analysis of the numbers in the industry reports: how
often test-runs are done and changing amounts of tool
shares from TestRail [13], predictions for how well the
AI-testing market will do [19], and a look at GPU
infrastructure costs for keeping up training and
inference [15

–

16, 18].

The technical portion of the study is supported by
architectural overviews of data lakes [14] and
continuous model training methodologies [15, 16].
Scenarios for integrating AI modules into CI/CD are
considered, including sending a hash and test list to an
ML endpoint, ranking a subset for execution, and self-
healing requests upon UI step failures [15]. Risk
assessment accounted for explainable AI metrics [17]
and economic factors (growth in GPU-instance
expenses) [18]. Organizational aspects

—

such as the

roles of prompt engineers and interactions between QA
and ML engineers

—

are described based on pilot

implementation practices and the monitoring of key

The American Journal of Engineering and Technology

224

https://www.theamericanjournals.com/index.php/tajet

metrics (MTTR, code-to-production velocity, and the
share of defects reaching production) [17].

Thus, the combination of platform comparative analysis,
grey-source review, content analysis of cases, and
quantitative industry data has enabled the development
of a concise yet comprehensive methodology for
implementing AI tools in continuous testing,
encompassing technology selection, organizational
preparation, and technical integration.

RESULTS AND DISCUSSION

The transition from mechanical automation to
intelligent automation is explained by two interrelated
factors. First, companies have already exhausted cost
savings on scripts: 68% of organizations in the recent
World Quality Report 2024 stated that without
generative AI further expansion of automation becomes
economically unfeasible, whereas implementation of
GenAI promises to compensate for the growth of the
regression suite through intelligent selection and repair
of tests [3]. Second, the test data itself has become a rich
source of defect correlations, and machine learning
models can extract patterns that are not readily
available to manual analysis.

The qualitative difference between automating routine
steps and intelligent automation is evident at the task
level. Suppose the former forces the machine to repeat
pre-described scenarios. In that case, the latter tasks AI
with decision-making: which part of the suite to run,
how to fix a broken selector, or which code fragments
appear potentially defective. DevOps business metrics
measure the practical effect. Thus, the Predictive Test
Selection strategy deployed in the Meta monorepository
reduced the volume of executed tests to one-third of the
original, preserving 99.9% defect coverage and halving
infrastructure costs, which directly accelerated the
code-to-production cycle [4]. Simultaneously, large
language model

–

driven test generation demonstrates a

median coverage of 70.2% of code lines in the TestPilot
project, outperforming classical heuristics with no copy-
paste from existing tests [5].

This is also accomplished by lowering false-positive
breakages. The FlakeFlagger tool increased the precision
of identifying flakes in 16 out of 23 industrial projects;
consequently, teams could detect actual defects caused
by environmental instability at a faster rate [6]. All these

impacts together form a new Total Cost of Ownership
(TCO): the part of the budget that was previously
allocated for script upkeep and long runs is now
redistributed to feature creation.

The particular situations of AI use in constant testing
encompass the entire journey of ensuring quality is
achieved. With the rise of intelligent automation, large
language models examine how features are signed or
user records and create unit as well as complete end-to-
end (E2E) tests. This progress has been particularly
significant in recent efforts, such as GPTTestGen and
TestPilot. Then comes self-healing: models check the
current DOM against the baseline one and automatically
pick new locators. According to a multi-year review [7],
self-healing is the most common AI pattern among one
hundred commercial tools.

Next is prioritization: reinforcement learning algorithms,
such as Retecs or stochastic classifiers from the Meta
approach, select a minimal yet bug-rich subset of the
regression suite, which allows for keeping run duration
within minutes even with a sharp increase in code.
Predictive testing fills remaining gaps: an analysis of
coverage metrics and failure history forms a map of
hotspots, where additional checks are directed even
before code review. The fight against flaky tests is built
on clustering logs and identifying unstable steps.
Generative analytics automatically produces an
executive summary for management, thereby reducing
the cognitive load required for interpreting results.

Thus, AI shifts focus from manual script support to
semantic risk management. The combination of
intelligent generation, self-healing, prediction, and
prioritization simultaneously accelerates release flow. It
increases the reliability of feedback, thus directly
improving speed-to-release, reducing Mean Time To
Resolution (MTTR), and optimizing the total cost of
quality ownership.

The global market for AI-oriented testing tools has
already moved beyond the niche novelty phase and
entered the saturation phase. A meta-analysis [7]
counted over one hundred commercial and academic
solutions, with Applitools, Testim, Functionize, AccelQ,
and Mabl being the five most mentioned in the industry.
According to [19], the global AI-enabled testing market
size was estimated at $856.7 million in 2024 and is

The American Journal of Engineering and Technology

225

https://www.theamericanjournals.com/index.php/tajet

projected to increase from $1,010.9 million in 2025 to
$3,824.0 million by 2032, showing a compound annual
growth rate (CAGR) of 20.9% during the forecast period.

North America dominated the global AI-enabled testing
market, accounting for a 35.27% share in 2024, as shown
in Fig. 1.

Fig. 1. Projected Market Size of AI-Enabled Testing Tools, 2024

–

2032 (USD Million) [19]

Consulting confirms the same trend: Gartner forecasts
that the share of enterprises using AI add-ons to testing
pipelines will grow from 15% in 2023 to 80% by 2027 [8],
and the latest Forrester Wave evaluated fifteen vendors,
categorizing them into leaders, strong performers, and
niche players [9]. These figures demonstrate that
platform selection has ceased to be an innovative
experiment and has become a matter of technology
strategy.

Visual and cognitive testing products lead the
commercial SaaS segment. Applitools, specializing in
Visual AI, achieved an approximate annual revenue of
USD 35 million and expanded its distribution through
AWS Marketplace, indicating a mature client base and a
cloud-optimized business [10]. Mabl, in turn, operates
on a full-pipeline-as-a-service model and has raised a
total of $76.1 million in venture capital, including
investments from GV and CRV, demonstrating
institutional investor confidence in the plat

form’s long

-

term growth [11]. Testim and Functionize compete in
the same price range, focusing on self-healing and test
generation. Both services were included in the Forrester

report’s Strong Performer category, balancing their

relatively lower revenue with high technical scores [9].
Tricentis, occupying the enterprise segment, is noted by
Gartner as a representative vendor in the 2024 AI
automation guide, and its Tosca platform remains the

only one in the review offering end‐to‐

end codeless

automation for ERP landscapes, which is critical for large
enterprises that are not ready to rewrite tests manually
with every SAP release [12].

Technically, these solutions differ in the depth of their AI
layer and licensing scheme. Applitools, Mabl, and Testim
are sold on a SaaS subscription with monthly billing for
parallel sessions. Functionize and Tricentis complement
the subscription with volume licensing of connectors to
private clouds, which is particularly important for the
banking and government sectors. In terms of functional
coverage, the visual engines of Applitools/Functionize
lead in detecting UI regression, whereas Mabl and
Testim are stronger in generative recording of scenarios
and automatic locator remapping. All five platforms
support REST API for embedding predictions into
existing CI/CD scripts. Still, only Tricentis and Functionize

provide an on‐premises variant, which increase

s their

maturity for regulated industries.

The American Journal of Engineering and Technology

226

https://www.theamericanjournals.com/index.php/tajet

The open segment is formed by research projects that
more often serve as intelligent inserts rather than fully
functional systems. DeepOrder, built on deep recurrent
networks, demonstrated better early bug detection
efficiency compared to industrial heuristics. Both tools

are distributed under MIT‐type licenses, require self‐

training on run history, and currently lack commercial
SLAs, which limits their use by teams with insufficient
ML competencies.

Comparing maturity, one can consider SaaS platforms
ready for plug-and-play implementation, as they offer
certified integrations, user support, and key

performance indicators (KPIs) for reporting. Open‐

source solutions offer finer algorithm configuration and
no licensing costs, but transfer operational risks to the
team. The decision between them ultimately comes
down to the ratio of the three metrics above: release
speed, incident resolution time, and total cost of
ownership. In places where cost predictability and

regulatory reporting are critical, Applitools, Tricentis, or
Functionize lead; in areas where flexibility and
experimentation are more vital, Retecs or DeepOrder
give the team control over model and data.

The safe operation of AI parts begins with obtaining the
correct data for each attempt, including details about
the change, setup of the surroundings, start-up events,
step-by-step notes, coverage counts, and time

measurements. According to TestRail’s 2023

report,

automated tests are run more than one hundred times
per day by 62% of teams, generating tens of gigabytes of
logs every week. Therefore, information must be
presented in a simple form, along with smaller, concise
pieces for quick review [13]. Compared to 2021, fewer
teams were in the 0

–

100 range (42% to 38%) while more

teams were in the 1,001

–

10,000 range (17% to 20%), an

overall movement towards heavier automated testing
as shown in Fig.2.

Fig. 2. On average, how many automated tests does your team run per day? [13]

According to Fig. 3, Selenium remains the most popular
automation tool (approximately 41% in 2022, although
it was 45% in 2020), whereas the JavaScript-oriented
Cypress grew from 12% in 2020 to 19% in 2022. Java-
based tools (JUnit, TestNG) remain at roughly the same
level of about 16

–

17%, while the mobile solution

Appium, conversely, gradually loses share (from 21% in

2020 and 2021 to 15% in 2022), as does the behavior-
driven BDD framework Cucumber (from 18

–

20% in

2020

–

2021 to 14% in 2022). Cloud services gain notable

prominence (BrowserStack

–

11% in 2022), and new

players emerge: Playwright entered the sample for the
first time at 8%. Approximately one-fifth of respondents

in 2022 indicated 'Don’t Know', which suggests a

The American Journal of Engineering and Technology

227

https://www.theamericanjournals.com/index.php/tajet

particular gap in awareness of specific tools within teams.

Fig. 3. What test automation tools, suites, or frameworks do you use? [13]

A suitable destination is the enterprise data lake, which
is a repository: an object-based storage system that is S3
compatible and able to accept binary artifact archives
and structured events. Informed by a systematic review
of data-lake architectures, it is confirmed that keeping
separate the raw and prepared views reduces latency
between new test data arriving and being ready for
analytics, as complete dataset copies are not needed
[14].

The ML service has been developed as a standalone
microservice, consisting of two parts. A background
trainer retrieves new examples from the lake layer and
updates models, while the inference part provides REST
and gRPC interfaces for CI agents. The practice of
continuous training is described in Springer Data
Engineering research, where active, proactive training
updates weights incrementally with new data and
reduces parameter transfer volume by two to four
orders of magnitude without loss of quality, thus making
retraining within a day technically and economically

feasible [15].

Integration into the CI/CD pipeline itself seems
minimalist. After the artifact build, the pipeline script
sends the change hash and the list of available tests to
the ML endpoint and receives a ranked subset in return.
Upon initiation of UI scenarios, the driver wrapper, upon
each step failure, makes a secondary self-healing
request that returns alternative selectors. The final
mandatory element is a closed feedback loop. A
reporting script publishes actual results (success, error
type, duration) back to the lake, and the retraining
scheduler monitors data volume and staleness to launch
incremental model updates before concept drift occurs.
Modern reviews of online learning emphasize that this
strategy preserves metric stability amid code and
environment

changes

while

minimizing

the

computational costs of full retraining [16].

The motive for actual tests is a preexisting agreement:
the majority of firms utilize or deploy generative AI for
QA jobs,

yet admit that its worth must be demonstrated

The American Journal of Engineering and Technology

228

https://www.theamericanjournals.com/index.php/tajet

with numerical indicators. Therefore, the pilot begins
with a narrow, well-isolated microservice where the
change flow is relatively stable and historical runs are
easily collectible. In choosing success metrics, DORA-like
indicators are used, including mean time from commit
to green build, mean time to resolution (MTTR), and the
share of defects caught before production. The baseline
for comparison is formed from existing logs. After
calibration, the model is transferred from the lab to
night builds, and coverage is gradually expanded: first,
the regression of critical user flows, then the service
layer, and finally component tests.

Process

evolution

also

requires

organizational

restructuring. The testing team assumes the role of
prompt engineers, describing the domain logic for the
generative model, whereas ML engineers ensure
continuous training and data quality. Change
management is based on transparent visualization of
model impact: a board with daily sparklines of MTTR and
machine-time savings is shown to all participants. It
serves as a trigger for further suite expansion.

Simultaneously, a risk control system is built. The main
technological vulnerability is model output opacity:
without explanations, an engineer does not know why a
test was excluded or healed. Explainable AI research
recommends using interpretability and trust metrics as
mandatory complements to classical AUC/Accuracy
metrics. The Pandora framework demonstrates that
hybrid human-machine defect analysis accelerates the
localization of systemic failures by nearly three times
[17]. Lastly, the infrastructure costs: Datadog notes a
40% year-over-year increase in organizational spending
on cloud GPU instances used for inference and model
training for test analytics, so when scaling, planning for
spot or on-premises modes is critical [18]. Therefore, a
well-scoped pilot, gradual rollout, specialist cross-
training, and steady monitoring of metrics make it
possible to implement AI tools without compromising
process stability. When these practices are followed,
corroborated by industry research, the economic effect

–

reducing MTTR, accelerating release, and lowering

TCO

–

becomes reproducible, and risks shift from the

undefined zone to being managed.

CONCLUSION

The study demonstrates that integrating AI tools into the

continuous testing process is not merely a fashionable
overlay but an almost inevitable step in the evolution of
DevOps ecosystems. Transition from simple repetition
of pre-described scenarios to intelligent management of
the test suite and self-diagnosis of infrastructure
establishes a qualitatively new paradigm in QA. The use
of machine learning algorithms for log clustering,
identification of flaky failures, and creation of a map of
hot spots for testing enables a rapid response to changes
in the codebase and minimizes false-positive triggers.
This, in turn, not only reduces mean time to recovery
(MTTR) but also increases trust in test results, making
feedback genuinely timely and reliable.

Moreover, the presented market research indicates that
AI-oriented tools have evolved beyond niche solutions
into a mature commercial ecosystem. Industry leaders

—

from Applitools and Testim to Tricentis and
Functionize

—

offer various deployment models (SaaS

and on-premises), allowing platform selection according
to specific regulatory and infrastructural requirements.
Considerable growth in investment and market
expansion forecasts already confirms that the choice of
an AI platform has become an integral part of a

company’s technological strategy, and above all, a key

factor in achieving a competitive advantage for those
who want to balance release speed with product
stability. This approach, supported by explainable AI
practices under organizational restructuring (which
acquires prompt engineering skills where QA and ML
engineers work collaboratively), minimizes both
technical and human risks related to implementation.

Thus, integration of AI tools into the continuous testing
process represents an evolution from mechanical
automation to risk management at a semantic level.
Comprehensive

application

of

intelligent

test

generation, self-healing, prioritization, and predictive
selection establishes feedback speed and quality
previously unattainable with traditional approaches.
When described practices are followed and appropriate
infrastructure for data storage and processing is
established, the economic benefits

—

accelerated

release, reduced Mean Time To Resolution (MTTR), and
optimized

total

cost

of

ownership

—

become

reproducible and manageable.

REFERENCES

1.

“State of CI/CD Report 2024: The Evolution of

The American Journal of Engineering and Technology

229

https://www.theamericanjournals.com/index.php/tajet

Software Delivery Performance,” CD Foundation.

https://cd.foundation/state-of-cicd-2024/
(accessed May 01, 2025).

2.

M. S. Bari, A. Sarkar, and S. A. M. Islam, “Ai

-

augmented Self-healing Automation Frameworks:
Revolutionizing Qa Testing with Adaptive and

Resilient Automation,” Advanced International

Journal of Multidisciplinary Research, vol. 2, no. 6,
Dec.

2024,

doi:

https://doi.org/10.62127/aijmr.2024.v02i06.1118.

3.

Open Text Corporation, “World Quality Report 2024

shows 68% of Organizations Now Utilizing Gen AI to

Advance Quality Engineering,” PR Newswire, Oct.

22, 2024. https://www.prnewswire.com/news-
releases/world-quality-report-2024-shows-68-of-
organizations-now-utilizing-gen-ai-to-advance--
quality-engineering-302282709.html (accessed May
03, 2025).

4.

M. Machalica, A. Samylkin, M. Porth, and S.

Chandra, “Predictive Test Selection,” arXiv, Jan.

2018,

doi:

https://doi.org/10.48550/arxiv.1810.05286.

5.

M. Schäfer, S. Nadi, A. Eghbali, and F. Tip, “An

Empirical Evaluation of Using Large Language

Models for Automated Unit Test Generation,” arXiv,

Sep.

2023,

doi:

https://doi.org/10.48550/arXiv.2302.06527.

6.

A. Alshammari, C. Morris, M. Hilton, and J. Bell,

“FlakeFlagger:

Predicting

Flakiness

Without

Rerunning

Tests,”

2021

IEEE/ACM

43rd

International Conference on Software Engineering
(ICSE),

May

2021,

doi:

https://doi.org/10.1109/icse43902.2021.00140.

7.

F. Ricca, A. Marchetto, and A. Stocco, “A Multi

-Year

Grey Literature Review on AI-assisted Test

Automation,” arXiv (Cornell University), Aug. 2024,

doi: https://doi.org/10.48550/arxiv.2408.06224.

8.

“Gartner Market Guide for AI

-Augmented

Software-

Testing Tools 2024,” Tricentis, Mar. 04,

2024.
https://www.tricentis.com/resources/gartner-ai-
market-guide (accessed May 07, 2025).

9.

D. Giudice, C. Gardner, A. Cornwall, A. Lozada, and

K. Hartig, “The Forrester WaveTM: Continuous
Automation Testing Platforms, Q4 2022,” 2022.

Accessed: May 10, 2025. [Online]. Available:
https://www.tekmarkgroup.com/wp-
content/uploads/2023/10/The-Forrester-
Wave%E2%84%A2-Continuous-Automation-
Testing-Platforms-Q4-2022.pdf

10.

“Applitools Company Overview, Contact Details &

Competitors,”

LeadIQ,

2024.

https://leadiq.com/c/applitools/5a1d9d422300005

b008d1c07 (accessed May 14, 2025).

11.

“Mabl company profile,” Tracxn, Aug. 17, 2021.

https://tracxn.com/d/companies/mabl/__b9wljjy7
_SAZSPVCj34bz9RedSOIojaJFYuWPFn4iBk (accessed
May 16, 2025).

12.

“Tricentis Recognized in the February 2024

Gartner® Market Guide for AI-Augmented Software-

Testing Tools,” Tricentis, Mar. 06, 2024.

https://www.tricentis.com/news/tricentis-2024-
gartner-market-guide-for-ai-augmented-software-
testing-tools (accessed May 17, 2025).

13.

“2023 Software Testing and Quality Report,”

Testrail, 2023. Accessed: May 16, 2025. [Online].
Available:
https://content.testrail.com/hubfs/Downloadables
/TestRail-2023-Software-Testing-and-Quality-
Report.pdf

14.

S. Azzabi, Z. Alfughi, and A. Ouda, “Data Lakes: A
Survey of Concepts and Architectures,” Computers,

vol. 13, no. 7, p. 183, Jul. 2024, doi:
https://doi.org/10.3390/computers13070183.

15.

I. Prapas, B. Derakhshan, A. R. Mahdiraji, and V.

Markl, “Continuous Training and Deployment of
Deep Learning Models,” Datenbank

-Spektrum, vol.

21, no. 3, pp. 203

–

212, Nov. 2021, doi:

https://doi.org/10.1007/s13222-021-00386-8.

16.

M. Sulaiman, M. Farmanbar, S. Kagami, A. N.

Belbachir, and C. Rong, “Online deep learning’s role

in conquering the challenges of streaming data: a

survey,” Knowledge and Information Systems, Feb.

2025, doi: https://doi.org/10.1007/s10115-025-
02351-3.

17.

B. Nushi, E. Kamar, and E. Horvitz, “Towards

Accountable AI: Hybrid Human-Machine Analyses

for Characterizing System Failure,” arXiv (Cornell

University),

Sep.

2018,

doi:

https://doi.org/10.48550/arxiv.1809.07424.

18.

Datadog Inc., “Datadog’s State of Cloud Costs 2024

Report Finds Spending on GPU Instances Growing

40% as Organizations Experiment with AI,” PR

Newswire,

Jun.

13,

2024.

https://www.prnewswire.com/news-
releases/datadogs-state-of-cloud-costs-2024-
report-finds-spending-on-gpu-instances-growing-
40-as-organizations-experiment-with-ai-
302172281.html (accessed May 25, 2025).

19.

“AI

-enabled Testing Market Size, Share & COVID-19

Impact Analysis,” Fortune Business Insights, May 12,

2025. https://www.fortunebusinessinsights.com/ai-
enabled-testing-market-108825 (accessed May 26,
2025).