Human–AI Collaboration in the Modernization of COBOL-Based Legacy Systems: The Case of the Department of Government Efficiency (DOGE)

Abstract

This paper aims to explore the challenges of maintaining and modernizing legacy systems, particularly COBOL-based platforms, the backbone of many financial and administrative systems. By exploring the DOGE team’s initiative to modernize government IT systems on a relevant case study, the author analyzes the pros and cons of AI and Agile methodologies in addressing the limitations of static and highly resilient legacy architectures. A systematic literature review was conducted to assess the state of the art about legacy system modernization, AI integration, and Agile methodologies. Then, the gray literature was analyzed to provide practical insights into how government agencies can modernize their IT infrastructures while addressing the growing shortage of COBOL experts. Findings suggest that AI may support interoperability, automation, and knowledge abstraction, but also introduce new risks related to cybersecurity, workforce disruption, and knowledge retention. Furthermore, the transition from Waterfall to Agile approaches poses significant epistemological and operational challenges. The results highlight the importance of adopting a hybrid human–AI model and structured governance strategies to ensure sustainable and secure system evolution. This study offers valuable insights for organizations that are facing the challenge of balancing the desire for modernization with the need to ensure their systems remain functional and manage tacit knowledge transfer.

1. Introduction

In the era of digital transformation, many sectors such as finance, telecommunications, healthcare, and public administration are under increasing pressure to modernize legacy systems to keep up with the incorporation of new advanced technologies and sustain a competitive advantage [1]. COBOL, developed in the 1960s, remains fundamental to these sectors, particularly public administration, where it supports critical systems such as tax processing, pension management, and social security [2].

However, over the past few decades, due to the retirement of the original developers and key stakeholders of these systems, coupled with a lack of available domain experts and the complexity of transitioning legacy systems to newer technologies [3], knowledge sharing has become a critical driver of innovation [4]. Recent studies have reinforced these findings, highlighting the need for substantial modifications to legacy IT infrastructures, particularly to address data privacy concerns and to mitigate the lack of specialized skills, which hinders the conversion of tacit knowledge into explicit knowledge—a fundamental process for sustaining long-term viability [5]. This is particularly relevant in sectors such as finance in developing countries, where legacy systems persist and cybersecurity resilience relies increasingly on the adoption of advanced technologies like AI and machine learning (ML) [6].

Empirical research suggests that migrating legacy software is a better approach than discarding it and developing entirely new systems. However, this migration does not guarantee the correctness and completeness of the legacy system, and it may take time to gain the confidence of users [7]. Recognizing these challenges, academics have explored strategies for migrating a legacy system, but outline that at the time of publication of their study, there was no structured approach to supporting developers in the complex activities of a migration process [8]. Some practitioners believe that cutting-edge technologies, such as generative AI, could help governments address the significant operating risks associated with the retirement of available workers with knowledge on outdated programming languages, such as COBOL, but this is not yet a consensual idea [9].

To encourage researchers to rethink solutions in AI-related areas and to generate fresh evidence and new insights that bridge the gap between theory and practice, there have been several calls for papers from top-tier journals, such as “Human-computer interaction” from Informatics (ISSN 2227-9709), which leads us to believe that AI tools can play an important role in shaping the knowledge ecosystem [10].

Despite the growing academic interest in AI applications and approaches to modernizing legacy IT systems, particularly COBOL-based platforms, there has been little research into how AI, integrated with Agile methodologies, can enhance the modernization of these systems. While renowned companies such as IBM are at the forefront of many current developments, including new mainframes, academic engagement with COBOL is declining as students and teachers lose interest in the language, pressing universities to sustain engagement through curriculum innovation [11]. Moreover, there appears to be greater need for real-world use cases and user histories than for theoretical debates [12]. However, theoretical studies remain essential to support new research.

Building on this foundation, this study positions itself at the intersection of those two needs, offering both a conceptual synthesis and a case-based perspective, responding to a clear research gap: the need for a deeper understanding of the pros and cons of AI and Agile methodologies in overcoming the limitations of static and highly resilient legacy architectures. It contributes by (i) analyzing how AI and Agile can be jointly used to modernize legacy systems, (ii) illustrating this through the DOGE case, and (iii) proposing actionable guidance for future public-sector transformations. The research question is therefore: “How can AI and Agile methodologies be applied to modernize COBOL-based government systems, while addressing automation goals, cybersecurity risks, and the loss of critical legacy knowledge?”

To answer this question, the study examines the landscape of COBOL-based legacy systems, with a particular focus on the real-world case of DOGE, an initiative of the second Donald Trump administration in the United States, whose goal is “modernising federal technology and software to maximize governmental efficiency and productivity“ [13].

Rather than relying solely on conceptual analysis, this research uses a dual-source literature approach to ensure a comprehensive understanding of both scientific and practical perspectives:

• The scientific literature: This study integrates peer-reviewed publications on Agile methods, particularly since the Agile Manifesto in 2001 [4], and on AI applications in legacy modernization.
• The gray literature and industry insights: To include practical perspectives, this research examines technology magazines, industry reports, and expert analysis, such as MIT Technology Review articles [14].

This combined approach enables a critical analysis of a politically sensitive yet operationally relevant government initiative, providing a grounded reflection on public-sector IT transformation. Through an examination of this case, the study evaluates the trade-offs of legacy systems management and the potential of AI and Agile methodologies, while outlining future research directions to inform actionable strategies.

The article is structured into five sections: Firstly, an overview of the topic, including the main studies that support the research, research gaps, and research objectives. Next, core concepts of legacy modernization are defined. This is followed by the methodology section, which details the systematic literature review and the inclusion of the gray literature to provide practical insights. The results of the literature review and a case study are then discussed. Finally, theoretical and practical implications, recommendations for organizations, research limitations, and avenues for future research are presented.

2. State of the Art

In this section, the state of the art in legacy systems and COBOL-based infrastructures is examined, with a particular focus on modernization challenges and the integration of AI tools and Agile methodologies, based on a systematic literature review. Key frameworks such as AI-driven decision support systems [15], modular AI architecture for anomalies detection [16], and Agile-based iterative modernization models will be explored to highlight structured approaches to legacy system transformation [17]. Finally, the gray literature will be used to analyze the DOGE case.

2.1. Defining Legacy Systems and COBOL-Based Software

A legacy system is defined as an “older software application that uses an obsolete hardware and software platform which is hard and expensive to maintain and integrate with modern systems” [18]. Despite their increasing obsolescence, legacy systems continue to play an important role in multiple sectors, particularly in financial and governmental institutions [19]. As recently highlighted by [20], the maintenance of legacy software, often developed in COBOL, is a critical and necessary effort as it contains the important business logic. The authors emphasize that this modernization can be challenging and requires a deep understanding of system behavior, stating that “it is necessary to completely understand the behaviour of components in relation to their interfaces, i.e., their interface protocols, and to preserve this behaviour during the maintenance activities of the components” [20].

While some modernization strategies focus on replacing legacy systems, others advocate for extending their lifespan through incremental improvements. Authors such as [21] argue that legacy systems do not necessarily need to be replaced; instead, their effective life can be extended by integrating new technologies such as the Internet of Things (IoT) and AI. However, there are several approaches to modernizing legacy systems in the literature. The study [18] explore the use of artificial neural networks to enhance data processing and fraud detection while preserving the core business logic embedded in financial legacy systems. In contrast, [19] propose a system re-engineering approach, transitioning a monolithic architecture to a new microservice architecture, aiming to reduce complexity, achieve lower coupling and higher cohesion, and simplify integration.

Legacy systems present different levels of complexity and integration challenges, often categorized according to their adaptability and compatibility with modern technologies. Fully monolithic systems pose the greatest modernization challenges due to their rigid structure, whereas postmodern systems have been incrementally updated through enterprise-wide coordination, allowing them to evolve with internal and external requirements while integrating newer technologies [22]. In this regard, Ref. [23] refer that in 2019, many agencies that supported the COVID-19 response relied on outdated legacy systems (stand-alone electronic systems), which struggled to handle the increased demand for digital services and rapid data processing, so they had to accelerate their efforts to modernize these systems.

COBOL remains one of the most enduring programming languages in legacy systems, with at least 20 billion lines of code still in use worldwide, representing 43% of all banking systems in the US in 2017 [24]. However, its monolithic structure and outdated design make modernization complex, requiring program abstraction and Agile approaches to support the redevelopment and improvement of the system [4].

2.2. Legacy Systems, AI Tools, and Agile Methodologies

Based on a comprehensive review, the study [15] highlight the extensive applications of AI techniques in revolutionizing traditional approaches in IT infrastructures, particularly in (a) improving accuracy [25], (b) reducing operational costs [26], and (c) enhancing decision-making [16]. According to Elahi [15], they also characterize modernization strategies for legacy systems, characterize AI-driven data processing, predictive maintenance, and automated decision support with AI-generated insights. In this regard, [26] highlight how AI-enhanced real-time monitoring improves the efficiency of legacy system operations by reducing response times to failures and optimizing resource allocation. However, [27] explain that despite these advances, the biggest challenge remains the integration of new technologies with the legacy infrastructure, as most organizations are required to completely restructure their classical systems or find a way to ensure successful integration. To address this concern, the study [28] use a real-world case study of a legacy banking system to describe how the challenge previously highlighted by Asif, Ghanem, and Irvine [27] can be addressed by leveraging Agile methodologies so that organizations can make incremental adjustments and continuously assess requirements to mitigate technical risks.

A central challenge is the epistemological gap between COBOL’s structured Waterfall approach and Agile’s constructivist paradigm [29]. COBOL development is rooted in positivist epistemology, emphasizing linear processes, rigid documentation, and pre-defined requirements. This results in codified but static knowledge, making it difficult to adapt to evolving business needs. Agile methodologies, on the other hand, operate within a constructivist epistemology, where knowledge is continuously refined through collaborative problem-solving, iteration, and team-based learning.

The shift from Waterfall to Agile is not just a technical transition but a fundamental change in how knowledge is created, shared, and retained. Catherine et al. [30] show a real case of modernizing COBOL-based legacy banking systems that COBOL experts rely on for extensive documentation and structured workflows, while Agile promotes tacit knowledge transfer through practices like pair programming and cross-functional collaboration. This collaboration is important for preventing the loss of institutional knowledge, particularly as COBOL-experienced professionals retire.

Terry and Chandrasekar [17] developed a structured Agile lifecycle model to support the evolution of legacy systems into an enterprise by iteratively refining requirements and managing complexity during system transformation. AI-assisted abstraction tools can support this transition by automating system abstractions and generating target code in multiple languages/platforms while preserving the core functionality of legacy infrastructures [4].

In the case study by Sullivan [23], adopting an Agile approach with constant real-time feedback from end users in a high-pressure modernization scenario, such as the COVID-19 pandemic, is key to successful outcomes. However, the integration of AI tools into legacy infrastructures raises data privacy concerns, as AI-driven automation requires extensive access to historically sensitive information [31]. These challenges are particularly evident in the case of DOGE, which is presented in the next section.

2.3. DOGE Legacy Systems Characterization

After establishing the key concepts and discussing the main theoretical foundations in Section 2.1 and Section 2.2, this section provides the empirical context for the research by characterizing the federal modernization program launched under Executive Order 14158 and led by DOGE, which has spawned opposition and lawsuits [32]. This analysis is particularly relevant for practitioners and academics seeking to understand the technological characteristics of DOGE’s COBOL-based infrastructure as well as its modernization challenges and broader implications. Despite its origins in a politically charged context, DOGE is a legitimate empirical case because it was formally established through Executive Order 14158 [33], and it is nationally relevant in terms of shaping federal IT modernization policy.

COBOL-Based Legacy Systems and Modernization Challenges: The Case of DOGE

A notable example of a federal initiative addressing legacy modernization is the DOGE program, which was initiated by Elon Musk and Vivek Ramaswamy in 2024 and formalized in 2025 [34]. The initiative began in mid-2024, when Elon Musk proposed a “government efficiency commission” to Donald Trump, which later evolved into DOGE, and was officially established by Executive Order 14158 in January 2025 to consolidate agencies, cut costs, and modernize outdated IT systems with AI-driven alternatives [35]. Musk began highlighting the urgent need for this modernization on social media in 2024, saying, “The government runs on ancient computers & software. Needs an upgrade!” [36], and shared a Government Accountability Office (GAO) report from 2023 that revealed that several federal agencies still rely on outdated and obsolete IT systems using older languages such as COBOL, some of which exceed 50 years old in operation [34,37].

The urgency of this matter is reflected in previous GAO assessments. According to the GAO report in 2019 [37], ten critical federal agencies were identified as most in need of modernization in 2019, but only two agencies had documented modernization plans that included three key elements: (i) milestones to complete the modernization, (ii) description of the work necessary to modernize the legacy system, and (iii) details regarding the disposition of the legacy systems. By 2023, six of the remaining eight agencies had implemented modernization plans based on the GAO’s recommendations, leaving the Department of Transportation and the Office of Personnel Management as the only agencies without a fully developed plan.

Table 1. Progress in documenting modernization plans for critical federal legacy systems (adapted from GAO-23-106821 report).

Agency	System Description	Had Modernization Plan with Key Elements, as of June 2019?	Has Addressed Incomplete Elements of Modernization Plan, as of May 2023?
Department of Defense	A maintenance system that supports wartime readiness, among other things	Yes. Agency included all elements in its modernization plan	Not applicable
Department of Education	A system that contains student information	No. Agency did not have a documented modernization plan	Yes. Agency included all elements in its modernization plan
Department of Health and Human Services	An information system that supports clinical and patient administrative activities	No. Agency did not have a documented modernization plan	Yes. Agency included all elements in its modernization plan
Department of Homeland Security	A network that consists of routers, switches, and other network appliances	Partial. Agency had a modernization plan but it did not include milestones or the disposition of the legacy system	Yes. Agency included all elements in its modernization plan
Department of the Interior	A system that supports the operation of certain dams and power plants	Yes. Agency included all elements in its modernization plan	Not applicable
Department of the Treasury	A system that contains taxpayer information	Partial. Agency had a modernization plan but it did not fully include milestones and it did not include the disposition of the legacy system	Yes. Agency included all elements in its modernization plan
Department of Transportation	A system that contains information on aircraft	No. Agency did not have a documented modernization plan	No. In April 2022, agency officials informed us that they expected to go live with the modernized system in the fall of 2022; however, as of May 2023, we have not received documented plans for this modernization effort
Office of Personnel Management	Hardware, software, and service components that support information technology applications and services	Partial. Agency had a modernization plan but it did not fully include milestones or work necessary, and it did not include the disposition of the legacy system	No. As of May 2023, we have not received evidence that the agency has developed a comprehensive modernization plan for this system
Small Business Administration	A system that controls access to applications	Partial. Agency had a modernization plan but it did not include the work necessary	Yes. Agency included all elements in its modernization plan
Social Security Administration	A group of systems that contain information on Social Security beneficiaries	Partial. Agency had a modernization plan but it did not fully include milestones or work necessary, and it did not include the disposition of the legacy system	Yes. Agency included all elements in its modernization plan

provides a comparison of agencies’ progress in implementing modernization plans from June 2019 to May 2023.

A preliminary analysis suggests that while federal agencies are increasingly prioritizing the modernization of legacy systems, challenges remain, particularly in ensuring accuracy and preventing critical operational failures. For instance, Burman [38] highlights that legacy COBOL infrastructures continue to pose significant operational challenges at the Social Security Administration (SSA), which is under scrutiny following critical errors in beneficiary records, including cases where people over the age of 150 were still listed as beneficiaries due to flaws in data processing systems. These issues led the author to set the goal of analyzing more closely the impact of outdated systems on government processes and regulated industries, particularly DOGE.

In addition to these technical setbacks, efforts to modernize also raise broader governance and security concerns. The MIT Technology Review article, published in February 2025 [14], warns that DOGE’s modernization strategy, while ambitious, has raised significant security risks by eliminating key oversight mechanisms. The so-called “Evil Housekeeper Problem”, which states that “once someone is in your hotel room with your laptop, all bets are off”, is used as an analogy to illustrate the risks linked to DOGE’s decision to remove the Authority to Operate (ATO) process, which allowed unauthorized individuals to make unauthorized changes to government systems.

In response to these risks, the implementation of DOGE is expected to comply with the GSA’s 2024 directive on artificial intelligence. This directive requires all federal AI systems to undergo pre-deployment registration and risk-based impact assessments. It also requires oversight by an AI Safety Team. This includes the introduction of mandatory human fallback protocols for public-facing systems and the labeling of AI-generated outputs. These safeguards are intended to ensure traceability, accountability, and secure deployment of high-impact government applications [39].

While these compliance measures are being institutionalized, DOGE is also actively expanding its use of AI in operational processes. According to Shrivastava [40], DOGE has begun to use AI models to optimize government operations, particularly in workforce management and the processing of large amounts of sensitive data. Kube et al. [41] explained that this initiative is controversial because DOGE plans to evaluate the productivity of government employees based on a Large Language Model (LLM) by asking employees to self-report what they got done in the last week via email, while AI determines which roles are redundant. Such initiatives could represent a first step in modernizing legacy systems by helping to improve reporting and data processing, but they also raise serious concerns about the reliability and fairness of AI-driven decision-making in government [42]. Although formal peer-reviewed evaluations are pending, internal assessments by the GSA report reductions in processing time and fewer administrative errors in early deployments [39].

This approach is in line with DOGE’s “AI-first strategy” perspective, as highlighted by Thomas Shedd’s recent comments to General Service Administration staff, and reinforces the administration’s push to restructure government agencies to operate like “startup software companies” [43]. The plan is to deploy AI widely across the federal government, using AI capabilities to analyze contracts for redundancies, root out fraud, and automate much of the work to facilitate a reduction in the federal workforce [44]. The initiatives have brought in skilled technologists to work with federal agencies, incorporating best practices from the private sector, such as Agile development, which emphasizes iterative development, continuous feedback loops, and user-centered design [45].

These advances in the public sector should be viewed with caution, considering that they require a high level of interaction, particularly in data-sensitive services [46]. Therefore, in line with other researchers, we also advocate the need for deeper empirical research to critically assess the effectiveness of integrating AI and Agile methodologies together in legacy system modernization.

3. Materials and Methods

This study follows a systematic literature review based on PRISMA guidelines to achieve the aim of providing the best possible evidence to inform academics and organizations, while employing a “replicable, scientific and transparent process” [47]. As shown in Figure 1, Scopus was chosen as our source due to its vast coverage of peer-reviewed articles [48] on Agile methodologies and AI applications when compared to other similar databases such as ScienceDirect and Web of Science. General-purpose search engines such as Google Scholar were excluded because they index non-peer-reviewed sources, which could compromise the consistency and quality standards of systematic reviews. The use of a single-source database was a deliberate choice to achieve a level of transparency that allows other researchers to easily reproduce the results. The literature searched covers the period from 2020 to the first three months of 2025. This period was chosen because the pandemic acted as a catalyst for digital transformation in the public sector, accelerating the need to modernize legacy systems and digital infrastructure [49].

As can be seen in Figure 1, there has been a significant increase in the number of articles published since 2020, due to the need to explore solutions for modernizing legacy systems, which will be the main topic explored in the case study presented in this article. Figure 1 further supports the relevance of the research topic and the time period chosen, as the number of publications related to legacy systems and COBOL has grown exponentially over the last few decades, with a particularly sharp increase in the last five years.

For the purpose of this study, the literature search was conducted on March 4th 2025, using the following combination of keywords in the Title, Abstract, and Keywords fields: (“COBOL” OR “legacy system*”) AND (“Artificial Intelligence” OR “AI” OR “LLM*” OR “Agile”).

The author chose to include “LLM” as a synonym for AI, but excluded other terms such as “genAI”, “copilot”, and “ChatGPT” after preliminary testing showed that they did not produce any unique results that were not already captured by broader terms such as “AI” and “LLM”. The results are illustrated in Figure 2.

To ensure a comprehensive coverage of international publications in both searches, only studies published in English and Portuguese were selected. Because of limited resources for translation, articles in other languages were excluded to avoid potential misinterpretation. There were a total of 188 studies listed. In the initial screening, the author excluded 29 documents that were not conference papers, articles, or reviews.

After an abstract screening was carried out, records were excluded. The remaining 57 articles underwent full-text eligibility assessment, where 40 records were further excluded for falling outside the scope defined by the inclusion and exclusion criteria (

Table 2. Inclusion and exclusion criteria.

Inclusion Criteria	Exclusion Criteria
Published in English or Portuguese	Published in languages other than English or Portuguese
Published between 2020 and 2025	Not published between 2020 and 2025
Classified as peer-reviewed journal articles, conference papers, or reviews	Not classified as a peer-reviewed article, conference paper, or review
Focused on legacy system modernization	Lacked relevant insights into legacy system modernization
Addressed AI and Agile applications in legacy system modernization	Did not focus on AI and Agile applications in legacy system modernization

). After this thorough evaluation, 17 articles were selected for the final inclusion stage of the PRISMA process flow (Figure 3). To support the process described above, the author used the open-source reference management software—Zotero—to manage bibliographic data. All citations were imported into Zotero.

In order to increase the comprehensiveness of the search, the gray literature was included in the researcher’s search strategy [50], through a ‘snowball’ technique where only sources with institutional credibility were considered, including official White House documents, executive orders, GSA directives, and government websites [51]. This gray literature focused on the DOGE case, particularly Elon Musk’s initiative to modernize government IT legacy systems [13].

To address the epistemological divide between Waterfall-based COBOL methodologies and Agile frameworks, this study applied Epistemological Paradigm Analysis [29].

In addition to the Epistemological Paradigm Analysis, this study is theoretically grounded in well-established knowledge management frameworks. These include Nonaka and Takeuchi’s SECI model [52], Wiig’s model of structured knowledge building [53], and Choo’s sense-making model [54]. These models offer complementary yet distinct insights into how tacit and explicit knowledge circulates within complex organisations. Together, they support a deeper interpretation of the DOGE initiative’s knowledge challenges, particularly with regard to legacy expertise, AI-driven abstraction, and interdepartmental coordination.

4. Results and Discussion

In this section, the author explores the pros and cons of legacy system modernization by comparing the existing frameworks identified in the systematic literature review. Next, Section 4.2 applies the findings of Section 4.1 to the case of DOGE, analyzing how AI tools and Agile methodologies contribute to tacit knowledge management, address the COBOL talent gap, and accelerate digital transformation in legacy IT environments in the case of DOGE.

4.1. Perspectives in Legacy System Modernization: Results from the Systematic Literature Review

The modernization of legacy systems, particularly COBOL-based infrastructures, presents challenges that go beyond technical limitations, extending into epistemological, organizational, and operational concerns [29]. While AI and Agile methodologies offer solutions, integrating them into rigid legacy systems requires overcoming barriers related to knowledge retention, system interoperability, cybersecurity, and organizational adaptation [1,15]. In this regard, studies such as DOGE, which is trying to have a more modern footprint while addressing the talent deficit of COBOL experts, have attracted attention in newspapers and industry blogs, but there is still a notable lack of academic research on these topics, reflecting the novelty of the issue.

It is evident that legacy system modernization involves many different approaches, integrating AI tools and Agile methodologies, leading to widespread discussion within the academic community. However, while many studies examine these domains in parallel, few explore how Agile practices can effectively support the implementation of AI in legacy environments. In this regard, the author has identified three generic points of view:

• Optimistic view: Elahi et al. [15] conducted research to investigate how AI enhances the migration of legacy systems towards modernization through AI-based decision-making. Their results suggest that the strategic implementation of AI enables organizations to achieve higher productivity and cost-effectiveness. AI-driven knowledge automation helps mitigate skill shortages, ensuring knowledge extraction from legacy systems can be automated and leveraged for modern development.
• Cautious view: Other scholars adopt a more cautious approach, acknowledging the benefits of AI while highlighting potential cybersecurity vulnerabilities and interoperability issues in legacy systems. Ntafalias et al. [26] developed an AI-enabled architecture to connect legacy-building automation systems, structured across multiple interoperability layers. Their work demonstrates that, although AI can enhance control and responsiveness, achieving full integration requires resolving semantic conflicts and ensuring compatibility across the entire system. Although Agile methodologies are not explicitly discussed, the modular design and layered orchestration of their framework align with Agile principles, such as iterative development and incremental integration. This suggests that Agile-inspired strategies could help to address integration issues, particularly in heterogeneous legacy environments.
• Skeptical view: The most skeptical positions point out that while AI and Agile methodologies hold promise, their full potential in large-scale legacy system modernization remains uncertain due to unresolved issues such as scalability, standardization, data sharing, and interoperability, as many proposed frameworks are still at the pilot stage [27].

Nerur, Mahapatra, and Mangalaraj (2005) developed a comparative framework contrasting the positivist epistemology underpinning Waterfall with the constructivist paradigm of Agile [29]. Waterfall methodologies depend on structured workflows, predetermined phases, and extensive documentation to ensure system predictability, yet they lack adaptability. In contrast, Agile fosters adaptability, continuous feedback, and collaborative decision-making through dynamic knowledge creation. However, transitioning from Waterfall to Agile poses risks to institutional knowledge retention. Nerur, Mahapatra, and Mangalaraj (2005) argue that hybrid models combining structured documentation with Agile’s iterative learning cycles can maintain system integrity while fostering adaptability [29]. Lano et al. [4] highlight that Agile fosters collaboration between legacy system experts and development teams, ensuring knowledge retention and mitigating the risks of workforce attrition in COBOL-based infrastructures.

Several researchers have proposed solutions to mitigate these challenges. Terry and Chandrasekar [17] introduced a structured Agile lifecycle model designed to facilitate legacy system evolution while maintaining knowledge retention. Their framework incorporates iterative requirement refinement, allowing legacy systems to transition gradually without compromising documentation integrity. Lano et al. [4] extended this approach by proposing AI-assisted abstraction techniques to automate code migration while maintaining system integrity. Their research highlights that while AI-driven abstraction tools accelerate modernization, they must be complemented by structured oversight mechanisms to prevent technical debt and system instability. Meanwhile, Brataas et al. [28] proposed a knowledge-sharing model that leverages Agile practices such as pair programming and cross-functional collaboration to bridge generational gaps between COBOL experts and modern developers, reducing workforce transition risks. These examples show how Agile frameworks can be adapted to support the integration of AI by encouraging collaboration, enabling iterative validation and ensuring system continuity.

As Wang et al. [1] note, the financial sector provides a critical example of how AI and Agile methodologies are shaping legacy system modernization. Their research outlines how cloud computing, big data analytics, and blockchain have revolutionized financial services, enabling greater scalability and automation. However, they emphasize that rapid digitalization introduces cybersecurity risks, regulatory compliance challenges, and interoperability constraints.

In summary, as the urgency to modernize legacy infrastructures grows, AI-driven tools and frameworks are being developed and, in this regard, researchers have increasingly focused on the complexities of technological adoption and regulatory compliance. Despite the rise in investigations (Figure 1), these studies have primarily examined specific technological approaches tailored to individual systems rather than providing a holistic assessment of the advantages and limitations of existing frameworks for organizations [15].

Table 3. Summary of perspectives on legacy system modernization.

Resume of Relevant Categories	Main Authors
Advantages of AI in modernization: AI improves productivity, reduces costs, and automates knowledge extraction, helping mitigate skill shortages in legacy system migration	[15]
Challenges in AI and Agile integration for legacy systems: Concerns over cybersecurity risks, interoperability issues, and organizational resistance create barriers to seamless AI and Agile adoption	[1,26]
Scalability and feasibility of AI and Agile frameworks: The large-scale adoption of AI and Agile is limited by unresolved issues in standardization, data sharing, and system interoperability. Many frameworks remain in pilot stages	[27]
Knowledge retention and transition from Waterfall to Agile: Waterfall ensures structured documentation but limits adaptability. Agile supports collaboration but risks knowledge loss without proper transition strategies. Hybrid models are needed	[4,29]

summarizes Section 4.1, highlighting that AI-driven decision-making optimizes modernization and reduces costs. However, cybersecurity, interoperability, and organizational adaptation challenges persist. Most authors advocate a structured, cautious approach, integrating AI and Agile while ensuring compliance with security and governance standards. These findings lay the analytical groundwork for interpreting the DOGE initiative, which will be examined in the next section.

4.2. The DOGE Case: Incorporating Insights from Gray Literature Analysis

This section presents a critical synthesis of the DOGE case through the lens of the theoretical findings in Section 4.1. As noted in Section 2.3, DOGE launched an ambitious initiative to modernize legacy IT systems that are “outdated or obsolete” [34]. The initiative aimed to eliminate bureaucracy and adopt a “startup-like” digital operating model [55]. However, the execution raised concerns among experts, particularly due to the removal of traditional security mechanisms such as ATOs [14]. Despite these concerns, DOGE has continued with its “AI-first” strategy, in line with the principles of the Agile Manifesto [43]. For example, according to anonymous sources, the use of ML to analyze administrative text data is becoming increasingly common within DOGE [41].

Although DOGE has faced criticism, the push for AI-driven modernization should not be completely dismissed, but rather re-evaluated. Compared to the traditional approaches, the integration of AI tools, such as the General Services Administration (GSA) chatbot, can streamline administrative tasks (e.g., email drafting and document analysis), improve fraud detection in contract reviews, and increase the accessibility and responsiveness of government services [46].

In this section, the aim is to discuss DOGE’s modernization efforts in the light of AI-driven automation and Agile methodologies at two levels (technical and organizational), as well as the associated pros and cons, as shown in

Table 4. Summary of Section 4.1 and Section 4.2.

Impact Levels	Key Aspects of Modernization	Findings from the Gray Literature (DOGE Case)	Pros of AI and Agile Modernization	Cons of AI and Agile Modernization	Main Authors
Technical Level	Security and Compliance	Removal of ATO security mechanisms in government IT increased cybersecurity risks [14]. DOGE’s Agile-first approach accelerated processes but weakened regulatory compliance, exposing systems to vulnerabilities [46].	AI-driven frameworks enhance fraud detection and automate risk assessments [26]. Machine learning models improve vulnerability detection [43]. Agile enables continuous security updatesm [26].	Over-reliance on AI for security decisions increases false positives/negatives [46].	[14,26,43,46]
	System Interoperability	DOGE introduced AI automation in legacy systems, but compatibility issues with COBOL-based architectures led to operational failures [46].	AI enhances interoperability through intelligent middleware and APIs, allowing modern platforms to interact with legacy systems. Digital Twin technology enables modernization without full replacements [15].	Lack of transition planning led to system failures. Heavy reliance on third-party AI-driven solutions increased costs and maintenance Complexity [15,46].	[15,46]
Organizational Level	Knowledge Retention and Workforce	AI evaluated workforce productivity and redundancies, improving knowledge transfer but raising ethical concerns over job losses [41].	AI facilitates knowledge capture and sharing, improving workforce training and reducing institutional knowledge loss. Agile methodologies enable continuous adaptation [28].	AI bias in workforce evaluations led to unfair layoffs. Over-reliance on AI automation reduced human oversight in personnel decisions [15].	[15,28,41]
	AI-Driven Decision-Making	DOGE automated hiring, budgeting, and contract oversight to improve efficiency but faced transparency issues [46]. The AI-first strategy reinforced automation in governance [43].	AI enhances bureaucratic efficiency, streamlining processes and enabling predictive analytics for better decision-making [17].	AI’s bias and lack of interpretability raise ethical concerns, limiting accountability in automated decisions [15,17].	[15,17,43,46]
	Cost Efficiency and Bureaucratic Optimization	The GSA chatbot automated administrative tasks, reducing costs but causing unintended workforce reductions [46].	AI optimizes resource allocation and improves public service efficiency [15].		[15,46]

.

The author will consider the following dimensions:

• Security and compliance;
• System interoperability;
• Knowledge retention and workforce;
• AI-driven decision-making;
• Cost efficiency and bureaucratic optimization.

Technical level: AI-driven modernization in DOGE has significantly improved system interoperability, enabling seamless data exchange between legacy COBOL-based infrastructures and modern platforms. AI-driven automation has optimized fraud detection and administrative processes, reducing inefficiencies and increasing accuracy. However, the removal of oversight mechanisms, such as the ATO process, has weakened cybersecurity and created vulnerabilities that could be exploited. Furthermore, while AI-driven workforce assessments offer efficiencies, they lack contextual understanding, raising concerns about biased assessments and misclassification of employee performance. This suggests a misalignment with best practices in Agile-based legacy transformation, which emphasize iterative validation and collaborative evaluation. For this reason, a structured AI framework is critical to ensure system reliability and mitigate the risks associated with automation.

Organizational level: The integration of AI and Agile methodologies has introduced greater adaptability, enabling refinement of government operations. However, the rapid and unstructured implementation of AI has created inconsistencies across departments, disrupted knowledge transfer, and led to resistance from employees who perceive automation as a threat to job stability. Relying on AI to assess employees has also raised ethical concerns, as automated decisions risk overlooking tacit knowledge and human expertise. A hybrid approach, in which AI complements rather than replaces human decision-making, would ensure that DOGE modernizes effectively while preserving institutional knowledge, upholding ethical standards, and fostering balanced human–AI collaboration within the organization.

In summary, the DOGE case illustrates both the potential and limitations of AI-driven modernization of legacy systems. Although the adoption of automation and Agile methodologies has improved efficiency by automating processes and enhancing system interoperability, particularly in COBOL-based legacy systems where expertise is already scarce, DOGE’s strategy aims to reduce dependency on a shrinking pool of COBOL experts through AI-based process automation. However, the rapid and unstructured adoption of AI has also raised concerns about knowledge sharing and human–AI collaboration, as automation alone cannot fully replace the tacit knowledge embedded in legacy systems [15,28]. Unlike the traditional Waterfall approach to system development, which follows a rigid, sequential structure that limits adaptability, the Agile approach facilitates iterative improvements and continuous adaptations to new requirements, making it more suitable for modernizing legacy infrastructures [3,16]. A hybrid approach, in which AI supports rather than replaces human decision-making, is essential for fostering sustainable digital government operations. Aspects related to such concerns may be a target for future research.

5. Conclusions

Organizations are increasingly trying to modernize their legacy systems, many of which still rely on outdated programming languages such as COBOL. However, only a few have succeeded in doing so. The literature reveals three dominant perspectives: optimistic, cautious, and skeptical. Some researchers see AI as a game changer, capable of automating processes and addressing the shortage of COBOL experts [9,15]; others argue that while AI and Agile methodologies are promising, their large-scale applications in legacy system modernization are still uncertain due to issues such as scalability, standardization, data interoperability, and governance [27]. Despite the growing interest in AI-driven modernization, academic research tends to focus on specific case studies rather than broader, structured frameworks that could guide organizations in different contexts. This suggests a gap in the literature—there is still no clear consensus on how AI and Agile can be effectively integrated to modernize legacy systems while ensuring knowledge retention.

5.1. Theoretical Implications

This research contributes to the theoretical understanding by clarifying that AI is most effective when augmenting—not replacing—human expertise. The DOGE case highlights both the potential and the risks of AI adoption, reinforcing the idea that AI should streamline tasks rather than eliminate roles. This is in line with the growing academic consensus that AI integration is most beneficial when it supports decision-making, increases productivity, and facilitates knowledge transfer. This key finding helps to demystify concerns about AI-driven job displacement and strengthens the argument that AI’s role in modernizing legacy systems should focus on collaboration rather than replacement [4].

5.2. Practical Implications

For managers looking to modernize legacy systems, the key is balance—AI can automate processes, improve system interoperability, and address the shortage of COBOL experts, but its implementation requires clear governance mechanisms to avoid knowledge gaps, employee resistance, and security risks. The DOGE case highlights the risks of rushing to adopt AI without structured oversight. Rather than replacing human expertise, organizations should use AI to support and enhance existing capabilities. Agile strategies are more adaptable for legacy system modernization. These methodologies allow for the gradual integration of AI tools, enabling real-world testing and adjustment without compromising system stability or institutional knowledge. A practical approach is to prioritize small-scale, well-defined AI deployments, gather user feedback, and scale adoption incrementally. This not only reduces disruption but also ensures that employees are adequately trained to collaborate with AI systems, rather than being displaced by them. Ultimately, AI should complement human expertise, not replace it, and organizations that take a phased, well-structured approach will modernize more effectively while minimizing disruption. Based on our analysis, we propose three practical steps for implementing a hybrid AI–Agile model in government legacy modernization projects:

• Use Agile teams to structure iterative AI integration. Begin with cross-functional teams that apply Agile sprints to define, test, and refine AI-supported functionalities (e.g., code refactoring and anomaly detection).
• Implement oversight and feedback loops. Combine AI-driven automation with continuous human validation to ensure that outputs remain aligned with operational goals and legal standards.
• Facilitate knowledge continuity. Use AI tools, such as NLP-based documentation, to extract and preserve legacy knowledge. Agile rituals, such as retrospectives and demos, ensure that knowledge is continuously shared across teams.

5.3. Limitations and Future Research

This study, which employs a systematic literature review and gray literature analysis, has some limitations, mainly due to its scope and the choices made during the research process. A key limitation is access to internal data, as the analysis is based solely on publicly available documents and industry reports rather than direct access to confidential sources or stakeholder interviews. In addition, time constraints posed challenges, as DOGE’s modernization initiative is still in its early stages. Long-term impacts, such as the sustainability of AI-driven transformations, workforce adaptation, and knowledge retention, cannot be fully assessed, only predicted. Another limitation is the scope of the literature review. While this study follows a systematic literature review methodology prior to gray literature analysis, it is limited by the scope of the Scopus database, which is continuously updated, meaning that recent studies beyond early 2025 were not included. In particular, Executive Order 14158, which officially established the DOGE initiative on 20 January 2025, is not yet indexed in major academic databases, and no peer-reviewed publications analyzing its implementation were available at the time of writing. Given the novelty of the topic, the gray literature and industry reports were included to capture the most recent developments; however, these sources are not peer-reviewed and may reflect editorial or political biases that affect the objectivity of the information presented. This strategy carries inherent trade-offs in terms of academic rigor, objectivity, and the generalizability of conclusions drawn from these sources. Furthermore, as the analysis focuses on one politically sensitive case, the generalizability of the findings is inherently limited. Conclusions should therefore be interpreted in the context of the DOGE initiative specifically, with careful distinction between peer-reviewed evidence and insights drawn from speculative or journalistic content.

Future research should aim to address these gaps by conducting empirical case studies in multiple federal agencies, not limited to DOGE, to compare different approaches to AI-enabled legacy system modernization. Qualitative research, including interviews with policy makers, IT specialists, and frontline workers, would provide richer insights into the decision-making processes and real-world challenges of modernization efforts. In addition, longitudinal studies should track KPIs such as cost savings, error rates, and processing speed across time. The development of structured frameworks for AI and Agile integration in legacy system modernization—taking into account governance, security, and ethical implications—should be a priority for future research. In particular, it would be valuable to test the influence of Agile practices (e.g., iterative feedback loops and sprint cycles) on deployment success in high-risk environments in the public sector. Pilot studies simulating AI–human collaboration in critical systems such as benefits processing, tax systems, and contract review could provide valuable insights for large-scale implementation.