The State of Generative AI in Software Development: Insights from Literature and a Developer Survey

The St a te of Genera tive AI in Softw are Development: Insights fr om Litera ture and a Developer Sur vey A Preprint Vincen t Gurgul Chair of Information Systems Hum b oldt-Univ ersität zu Berlin Un ter den Linden 6, 10117 Berlin Robin Gub ela Chair of Information Systems Ho c hsc hule für T ec hnik und Wirtschaft Berlin T resko w allee 8, 10318 Berlin Stefan Lessmann Chair of Information Systems Hum b oldt-Univ ersität zu Berlin Un ter den Linden 6, 10117 Berlin Buc harest Universit y of Economic Studies 6 Piata Romana, 010374, Romania Marc h 19, 2026 Abstra ct Generativ e Artificial In telligence (GenAI) rapidly transforms soft ware engineering, yet existing researc h remains fragmented across individual tasks in the Softw are Dev elopment Lifecycle. This study integrates a systematic literature review with a survey of 65 softw are dev elop ers. The results sho w that GenAI exerts its highest impact in design, implemen tation, testing, and do cumen tation, where ov er 70 % of developers rep ort at least halving the time for boilerplate and documentation tasks. 79 % of survey resp onden ts use GenAI daily , preferring bro wser-based Large Language Mo dels o ver alternativ es integrated directly in their developmen t environmen t. Gov ernance is maturing, with tw o-thirds of organizations main taining formal or informal guidelines. In con trast, early SDLC phases such as planning and requiremen ts analysis show mark edly lo wer reported b enefits. In a nutshell, GenAI shifts v alue creation from routine co ding tow ard sp ecification quality , architectural reasoning, and o versigh t, while risks such as uncritical adoption, skill erosion, and technical debt require robust go vernance and human-in-the-loop mechanisms. Keywor ds Generativ e Artificial Intelligence · Soft ware Engineering · Large Language Mo dels · AI-assisted Co ding · IT Gov ernance · Agile Developmen t 1 In tro duction The rapid adoption of Generativ e Artificial Intelligence (GenAI), particularly Large Language Mo dels (LLMs), is substantially transforming the Softw are Developmen t Life Cycle (SDLC). Chui et al. [2023] identify soft ware engineering as one of the domains with the highest economic impact p oten tial from GenAI. AI-assisted to ols such as GitHub Copilot are increasingly embedded in professional softw are engineering environmen ts, supp orting activities that span the entire lifecycle—from requirements analysis to architectural design, co de generation, testing, debugging, do cumen tation, and main tenance. GenAI refers to a class of machine learning systems capable of producing nov el conten t (such as text, images, co de, or sound) b y learning patterns from large datasets and generating contextually coherent outputs [Go odfellow et al., 2016]. Prominent examples include LLMs such as ChatGPT, which pro cess and generate natural language. While these systems do not replace h uman judgment or accountabilit y , they might augment The State of Generativ e AI in Softw are Developmen t: Insights from Literature and a Developer Survey dev elop er capabilities, p oten tially enabling faster iteration cycles, improv ed efficiency , and higher softw are qualit y . A t the same time, its adoption raises questions related to maintainabilit y , security , intellectual prop ert y , transparency , regulatory compliance, and long-term developer skill developmen t. Despite growing interest, the literature remains fragmen ted: Studies typically fo cus on asp ects related to isolated SDLC phases, while comprehensive assessments across the full lifecycle with empirical data remain scarce. Research agendas suc h as F euerriegel et al. [2024] call for more empirical in vestigations into how GenAI transforms organizational work practices and, sp ecifically , ho w suc h systems affect softw are dev elopment pro ductivit y . T o address these gaps, we derive the following research questions. R Q1: How are GenAI to ols currently used by softw are developers? R Q2: How is the impact of GenAI p erceiv ed across the phases of the SDLC? R Q3: What go vernance mechanisms are implemen ted to manage the adoption of GenAI in softw are dev elopment departments? R Q4: What are the p erceiv ed risks asso ciated with GenAI in softw are dev elopment? This study addres ses these questions by integrating three complemen tary sources of evidence. First, a structured review of academic literature synthesizes current scholarly knowledge on the p oten tial and limitations of GenAI across the phases of the SDLC and in agile dev elopmen t. Second, industry rep orts are incorp orated to capture emerging practices, technological developmen ts, and practitioner-oriented assessments that often precede academic publication. Third, a cross-sectional survey of 65 softw are developers building on the T echnology Acceptance Mo del (T AM) [Da vis, 1989, V enk atesh and Da vis, 2000] provides primary empirical evidence on real-w orld to ol usage, p erceiv ed pro ductivit y effects, gov ernance maturity , and asso ciated risks. By integrating insights from academic research, industry discourse, and practitioner exp erience, the study enables a more comprehensive and empirically grounded assessment of how GenAI is reshaping developmen t practices, influencing role profiles, and altering go vernance requirements. 2 Metho dology for Literature Analysis This section outline s the metho dology used to identify and select relev an t studies for the literature analysis presen ted in Section 3. Our literature selection process follows established metho dological guidelines for conducting systematic literature reviews in Information Systems research [W ebster and W atson, 2002, Okoli, 2015]. F ollo wing the typology prop osed by Paré et al. [2015], who distinguish nine literature review types, ours can b e c haracterized as a scoping review: Its primary goal is to identify and structure the current b ody of researc h on GenAI across the SDLC, rather than to aggregate quantitativ e data or to develop new theoretical prop ositions. By organizing the literature along the SDLC phases, we follo w the concept-centric approach suggested by W ebster and W atson [2002], shifting the fo cus from what individual authors hav e found to what is collectiv ely known ab out each stage of the developmen t pro cess. The selection criteria included p eer-review ed journal articles, conference pro ceedings, preprin ts, and tec hnical rep orts published by established consulting and research organizations. Extending the evidence base b eyond academic publications is in line with Okoli [2015], who ackno wledges non-p eer-review ed sources as part of the searc hable evidence base in systematic reviews. Including such literature is relev an t for our research in the rapidly evolving research field of GenAI, where muc h no vel knowledge is not yet reflected in p eer-review ed publications. F urthermore, only publications released b et ween 2021 and 2026 were considered, reflecting the p eriod during which large-scale GenAI systems b ecame widely accessible. The primary databases searched were Scopus, IEEE Xplore, the ACM Digital Library , SpringerLink, arXiv, and Go ogle Scholar. These sources were chosen to ensure broad cov erage of softw are engineering, information systems, and applied AI researc h. Searc h queries combined general and domain-sp ecific k eywords. Core searc h terms included “AI-assisted co ding”, “GitHub Copilot”, “Generative AI”, “large language mo dels”, “agile framew ork” and “softw are developmen t to ols”. These w ere complemented b y phase-sp ecific terms corresp onding to the SDLC, such as “requirements engineering”, “soft ware design”, “co de generation”, “testing”, “debugging”, “maintenance” and “pro ject managemen t”. In addition, selected industry rep orts from consulting and tec hnology firms such as McKinsey , Boston Consulting Group (BCG), Deloitte, and A ccenture were included, as these publications often pro vide early insights into technological adoption patterns and practical c hallenges that precede academic research. The screening and selection pro cess follow ed the three-stage framework of Levy and J. Ellis [2006] following prior literature screening approaches. In the first stage (input), the initial search across all databases yielded 2 The State of Generativ e AI in Softw are Developmen t: Insights from Literature and a Developer Survey 959,044 publications. In the second stage (pro cessing), duplicate records across databases, non-English man uscripts, and publications prior to 2019 were remo ved, and results were filtered b y publication type follo wing established literature screening pro cedures [for example F röhnel et al., 2024]. This step reduced the corpus to 1,712 candidate publications. The remaining records w ere then screened in tw o stages: first, titles and abstracts were reviewed for topical relev ance, resulting in 160 candidate publications. Second, the remaining articles underwen t full-text screening. Selection criteria included relev ance to the impact of GenAI on one or more SDLC phases, metho dological transparency , and substantiv e con tribution to understanding tec hnical, organizational, or managerial implications. Studies fo cusing solely on unrelated AI applications or lac king sufficient metho dological clarit y were excluded. W e identified additional publications through bac kward search [W ebster and W atson, 2002] by reviewing the reference lists of the remaining studies and applying the same selection criteria as b efore. The screening by full-text resulted in 63 publications included in the review. In the third stage, the final corpus of publications was organized according to global observ ations, individual SDLC phases, agile metho dologies, and risk analysis, as presented subsequently . 3 Literature Review The rapid adv ancemen t of GenAI is increasingly transforming soft ware developmen t practices across the SDLC. Among AI-assisted developmen t to ols, con versational LLMs suc h as ChatGPT and assis tan ts in tegrated directly in the dev elopment en vironment lik e GitHub Copilot, Amazon Co deWhisperer, JetBrains AI Assistan t, and T abnine hav e b ecome widely adopted, supp orting tasks suc h as co de generation, explanation, and review [Sergeyuk et al., 2025]. Empirical studies and field res earc h consistently report pro ductivit y gains across dev elopment activities, particularly in co de optimization (approximately 60 %), bug fixing (26 %), and do cumen tation supp ort (12 %) [Collan te et al., 2025]. Evidence further suggests that less exp erienced developers b enefit disprop ortionately from AI assistance, achieving higher relative pro ductivity improv ements [Brynjolfsson et al., 2023, Dell’Acqua et al., 2023, Ng et al., 2024]. At the same time, studies indicate that co de quality generally remains stable or impro ves despite reduced developmen t time [Shihab et al., 2025, Y adav and Mondal, 2025]. Ho wev er, correctness cannot b e guaranteed and long-term risks remain, making human v alidation and ov ersight essen tial [Y etiştiren et al., 2023, Atif et al., 2025]. Beyond individual productivity , industry rep orts highlight broader organizational implications, including accelerated prototyping, shorter time-to-market, and shifting accoun tability tow ard pro duct and pro ject leadership [Deniz et al., 2023, Gnanasambandam et al., 2025]. The following subsections synthesize the literature across the phases of the SDLC, b efore addressing agile soft ware developmen t as well as the risks and roadblo c ks asso ciated with GenAI adoption. 3.1 GenAI in the Softw are Dev elopment Lifecycle 3.1.1 Planning. In the planning phase, GenAI supp orts the creation of foundational pro ject artifacts such as scop e definitions, timelines, milestone structures, role descriptions, communication plans, and risk assessments [Barcaui and Monat, 2023, Hughes et al., 2025, Maggo o et al., 2025]. AI systems are also used to conduct market analyses, syn thesize comp etitiv e information, and formulate pro duct goals and visions [Lechner et al., 2025, Dell’Acqua et al., 2023]. Cost estimation and story p oin t appro ximation hav e likewise b een explored as application areas [W agner and W agner, 2024]. Empirical evidence suggests considerable productivity gains during pro ject initiation and planning, inclu ding faster setup of new pro jects and improv ed alignment b et ween technical teams and management [Aramali et al., 2025, Brynjolfsson et al., 2023, Hughes et al., 2025]. GenAI is typically deplo yed via structured prompts and increasingly integrated in to pro ject management to ols suc h as Jira, Confluence, or Microsoft Pro ject [Barcaui and Monat, 2023, Ng et al., 2024]. These capabilities contribute to early identification of risks and structured pro ject preparation. 3.1.2 Requiremen ts Analysis. During requirements analysis, GenAI assists in drafting, refining, and v alidating functional and technical requiremen ts. LLMs are used to generate and elab orate user stories and use cases based on high-level pro ject documentation [Cico et al., 2023, Pinto et al., 2023, W ei et al., 2025] and to create preliminary arc hitectural drafts and structured feature sp ecifications [Glev er et al., 2024, Gong et al., 2025, Nitin, 2024]. AI-assisted approaches reduce conceptualization time and accelerate the creation of prototypes, including 3 The State of Generativ e AI in Softw are Developmen t: Insights from Literature and a Developer Survey visually oriented UI/UX artifacts [W ei, 2024, W ei et al., 2025, Gnanasambandam et al., 2025]. T ools such as Figma AI and related plugins seem promising implementations for rapid prototyping [Petridis et al., 2024]. F urthermore, AI systems can analyze requirement do cumen ts to identify inconsistencies, ambiguities, and missing elements, enabling earlier error detection and reducing downstream rework [Gong et al., 2025, P orter et al., 2025]. Improv ed consistency b et ween business and IT stakeholders has also b een rep orted [Pinto et al., 2023, Brynjolfsson et al., 2023]. 3.1.3 Design and Implementation. The design and implementation phase represen ts the most extensively studied application domain for GenAI. AI-assisted co ding to ols such as GitHub Copilot provide real-time co de completion, refactoring suggestions, b oilerplate generation, and context-a w are optimization [Peng et al., 2023, Y etiştiren et al., 2023, Solohubov et al., 2023, Glever et al., 2024, Gerdemann et al., 2024]. They also supp ort automatic do cumen tation generation, including co de comments and API do cumen tation [Atif et al., 2025, Pinto et al., 2023, Ndiay e et al., 2025], as well as recommendations to improv e co de structure and maintainabilit y [Gong et al., 2025, Zhong et al., 2025]. Exp erimen tal and field studies consistently rep ort increased dev elopmen t sp eed [Peng et al., 2023, Cui et al., 2024, Shihab et al., 2025, Struever et al., 2025, Deniz et al., 2023]. Routine tasks are particularly affected, with automation reducing manual effort for standardized co ding activities [Solohubov et al., 2023]. Sev eral analyses also indicate improv ements in review efficiency and defect reduction, although h uman o versigh t remains essential [Collante et al., 2025, Y etiştiren et al., 2023, Ndiay e et al., 2025]. Impl emen tation t ypically o ccurs via integration into In tegrated Dev elopment En vironmen ts (IDEs) such as Visual Studio Code or In telliJ, often com bined with secure APIs or isolated mo dels to address data protection and compliance requiremen ts [Ng et al., 2024, Eb ert and Louridas, 2023]. 3.1.4 T esting and In tegration. In the testing and integration phase, GenAI is used to automatically generate unit, integration, and end-to-end test cases [Atif et al., 2025, Glever et al., 2024, Maggo o et al., 2025, Zhong et al., 2025]. Additional applications include the creation of mock ob jects and syn thetic test data [Y etiştiren et al., 2023, Gong et al., 2025], suggestions for missing test paths and cov erage gaps [Cico et al., 2023, Nitin, 2024], and AI-assisted security analysis [Ding et al., 2024, Pearce et al., 2025, Ndiay e et al., 2025]. The literature do cumen ts reductions in man ual testing effort for routine scenarios and broader, faster test cov erage [Kathiresan, 2024, Solohubov et al., 2023, Zhong et al., 2025] as well as earlier detection of edge cases and integration issues [Y etiştiren et al., 2023, Kathiresan, 2024]. In practice, GenAI is often integrated into CI/CD pipelines and connected to test framew orks or embedded directly into developmen t en vironments for automated co de analysis and test case generation [Ng et al., 2024, Pin to et al., 2023]. 3.1.5 Op eration and Main tenance. In the op eration and maintenance phase, GenAI supports the analysis of logs, error messages, and stac k traces, providing diagnostic insights and suggesting p oten tial remediation steps based on learned patterns [Ng et al., 2024, Gong et al., 2025, Zhong et al., 2025]. It is also used to explain, summarize, and restructure p oorly do cumen ted or complex legacy co de, improving main tainability and knowledge transfer [Nitin, 2024, A tif et al., 2025, Maggo o et al., 2025]. Rep orted effects include faster issue detection and resolution, improv ed traceability , and enhanced resilience to staff turnov er due to automated documentation and knowledge preserv ation [Ng et al., 2024, Pinto et al., 2023, Nitin, 2024]. Integration commonly o ccurs within inciden t and log managemen t platforms suc h as Jira, ServiceNow, or Grafana, and in sensitive environmen ts through locally deploy ed or access-controlled AI systems to mitigate compliance risks [Ng et al., 2024, Eb ert and Louridas, 2023]. 3.2 GenAI in Agile Softw are Dev elopment GenAI influences the dynamics of agile dev elopment practices b ey ond productivity gains. The literature rep orts high p erceiv ed usefulness and satisfaction in agile teams when using AI-assisted to ols [Geyer et al., 2025] and accelerated task completion, leading to shorter iteration cycles and more frequent releases [Zhang et al., 2024, Ulfsnes et al., 2024, Bahi et al., 2024]. By automating routine co ding, do cumen tation, and testing tasks, GenAI enables teams to fo cus on v alue-generating increments, reinforcing core agile principles suc h as rapid feedback and contin uous delivery . A t the same time, role exp ectations shift: Pro duct owners, 4 The State of Generativ e AI in Softw are Developmen t: Insights from Literature and a Developer Survey dev elop ers, and scrum masters increasingly engage in strategic, creative, and ev aluativ e activities, while routine tasks are partially delegated to AI systems [Dieb old, 2025, Bahi et al., 2024, Nasir and Hussain, 2024]. The literature recommends adapting role profiles and strengthening AI-related comp etencies, particularly critical ev aluation skills in AI-supp orted environmen ts. 3.3 Risks and Roadblo c ks in the A doption of GenAI Despite substantial pro ductivit y gains, the literature identifies a range of short- and long-term risks asso ciated with the integration of GenAI into softw are developmen t. In the short term, concerns relate to reliability , qualit y con trol, and security . AI-generated co de suggestions are frequently erroneous, with manual correction efforts av eraging appro ximately ten minutes p er iden tified defect [Y etiştiren et al., 2023]. A significant prop ortion of outputs requires p ost-editing before pro ductiv e use [Ziegler et al., 2022]. Developers often accept AI-generated snipp ets with limited scrutin y , suggesting sup erficial quality control practices [Ziegler et al., 2022, Chen et al., 2021]. P erceiv ed pro ductivit y impro vemen ts do not alwa ys align with ob jective activit y metrics, raising the risk of o v erestimating efficiency gains [Stra y et al., 2025]. Securit y vulnerabilities represen t a further concern: coding assistan ts can be susceptible to prompt injection attac ks, potentially leading to data contamination, unintended co de execution, or data leak age [Liu et al., 2025, Cotroneo et al., 2024, Norton et al., 2025]. Long-term risks are more structural. Contin uous reliance on AI-generated suggestions may ero de deep tec hnical knowledge and increase dep endency on AI systems [Ng et al., 2024, Brynjolfsson et al., 2023, Ding et al., 2025]. Automatically generated co de can introduce opaque dep endencies and arc hitectural inconsistencies, contributing to technical debt [Eb ert and Louridas, 2023, Atif et al., 2025, Anderson et al., 2025, Moresc hini et al., 2026]. Even syntactically correct co de may contain subtle semantic or securit y-relev ant w eaknesses [Y etiştiren et al., 2023, A tif et al., 2025]. Organizational dep endence on sp ecific vendors p oses additional strategic risks, including tec hnology lo c k-in and compliance exp osure [Ng et al., 2024]. A t the team level, some studies indicate reduced direct h uman interaction when AI to ols mediate commu- nication, co de explanation or problem-solving [Ulfsnes et al., 2024, Cabrero-Daniel et al., 2024], en tailing risks of kno wledge silos and diminished interpersonal trust. T o mitigate these effects, research recommends main taining structured exc hange formats such as co de reviews and sprint retrosp ectiv es. More broadly , gov er- nance demands increase as GenAI b ecomes embedded in pro duction workflo ws. Issues of quality assurance, accoun tability , compliance, and traceability b ecome more salient when AI-generated artifacts are incorp orated in to pro duction systems [Geyer et al., 2025, Zhang et al., 2024, Ulfsnes et al., 2024, Dieb old, 2025, Bahi et al., 2024]. The literature emphasizes the need to formalize review obligations, clarify resp onsibilit y for AI-assisted outputs, and establish dedicated gov ernance frameworks to prev ent quality degradation and unmanaged risk exp osure. Sev eral roadblo cks constrain the realization of GenAI’s full p otential. A large-scale industry survey [Ahlaw at et al., 2024] shows that co ding accounts for only 10–15 % of total developmen t time and less than 50 % of dev elop ers use GenAI. They conclude, that improv ements in co de generation alone ha ve limited ov erall impact. T o prev en t automation risks from offsetting pro ductivit y gains, a human-in-the-loop approac h (including qualit y review, domain judgment, and requirements o versigh t) is emphasized by industry rep orts [Deniz et al., 2023, Ahla wat et al., 2024, Struever et al., 2025]. Ov erall, the literature indicates that GenAI exerts measurable influence across all phases of the SDLC, with the strongest effects observ ed in design, implementation, and testing. How ever, these gains are neither uniform nor automatically realized—short-term limitations in reliabilit y and security , long-term risks such as skill erosion and technical debt as well as organizational roadblo c ks ma y constrain or offset pro ductivit y impro vemen ts. Effective gov ernance frameworks, structured human-in-the-loop v alidation, and delib erate organizational adaptation remain critical to translating GenAI’s tec hnical capabilities into lasting b enefits. 4 Surv ey Design T o complement the literature analysis with practitioner p erspectives, we conduct a standardized online survey informed by the T AM [Da vis, 1989, V enk atesh and Davis, 2000]. T AM provides a widely used framework for examining how users perceive the usefulness and adoption of new tec hnologies in organizational contexts. The surv ey in vestigates how softw are practitioners use GenAI tools, how they assess productivity effects across differen t SDLC phases, how gov ernance mec hanisms for GenAI adoption are implemented in practice, and whic h risks practitioners asso ciate with the use of GenAI. 5 The State of Generativ e AI in Softw are Developmen t: Insights from Literature and a Developer Survey Data were collected using a self-developed online questionnaire. The instrument comprised five sections: (1) demographic and professional background v ariables (age, education, role, team size); (2) primary developmen t metho dology (agile, traditional, h ybrid) and degree of adherence; (3) GenAI to ol usage b eha vior and frequency; (4) assessments of pro ductivit y , quality , and security implications across SDLC phases; and (5) time-based efficiency estimations and op en-ended questions on p erceiv ed opp ortunities and risks. A pretest ( n = 3 ) was conducted to refine w ording and improv e the distinction b et ween pro cess mo del categories. The target p opulation comprised professionals inv olved in softw are dev elopment and related disciplines, including softw are engineers, architects, team leads, DevOps engineers, pro duct owners, QA engineers, as w ell as students and in terns. Data collection took place betw een 31 December 2025 and 19 January 2026 using a multi-stage sno wball sampling approach [Döring, 2023] through professional and academic netw orks, including direct outreach on Link edIn. While the primary fo cus w as the DA CH region, participation from other regions w as not restricted. 5 Surv ey Results In the following we present the findings of our quantitativ e online surv ey (n = 65). Although the sample exhibits a comparatively young age structure, it demonstrates a strong professional profile. The ma jorit y of resp onden ts are industry practitioners, including 30 professional softw are developers, one team lead, and four pro duct owners. Notably , ten participants rep ort more than ten years of professional exp erience, indicating that the sample includes senior exp ertise despite its youthful demographic comp osition. Figure 5.1: Developmen t metho dology and age distribution With regard to developmen t metho dologies, the agile approaches Scrum and Kanban constitute the largest share at about 42 %. Appro ximately 34 % report op erating without a clearly defined pro cess mo del, follo wed by 17 % of resp onden ts who use a hybrid approach with b oth agile and traditional (waterfall-based) metho dologies. The comparably smallest group with ab out 8 % of resp onden ts work in formalized or plan-driv en (“T raditional”) environmen ts. Figure 5.2: GenAI to ol usage 6 The State of Generativ e AI in Softw are Developmen t: Insights from Literature and a Developer Survey As Figure 5.2 presents, the survey reveals widespread in tegration of GenAI into daily developmen t activities. Appro ximately 79 % of resp onden ts rep ort using AI to ols at least once p er day , indicating that GenAI has b ecome a routine component of professional practice. In terms of tools, browser-based LLMs dominate. ChatGPT is used by 77 % of participants, making it the most frequently adopted to ol, follow ed by Go ogle Gemini/Bard at 52 %. GitHub Copilot, which combines IDE-in tegrated co de completion with a conv ersational c hat in terface, is used b y about 48 % of resp ondents. IDE-in tegrated coding assistan ts such as JetBrains AI Assistant (11 %), T abnine (3 %), and Amazon Co deWhisp erer (2 %) remain marginal b y comparison. 17 % of respondents reported using other to ols while 3 % did not respond to the to ol-related question. This distribution suggests that developers fav or dialog-based interaction with general-purp ose LLMs ov er sp ecialized IDE-integrated alternatives. Figure 5.3: Impact of GenAI by SDLC phase Figure 5.3 display s the distribution of responses regarding the SDLC phases in which participants p erceiv ed the greatest utility of GenAI (m ultiple choice). 85 % of resp onden ts identify “Design and Implementation” as the phase with the strongest p erceiv ed b enefit, follow ed by “Op eration and Maintenance” at 75 %. Earlier lifecycle phases rece iv e notably low er ratings, with “Requirements Analysis” at 43 %, “T esting and Integration” at roughly 37 %, and “Planning” at 18 %. Although the low v alue for “Planning” might b e attributed to the high prop ortion of junior softw are developers in the sample, none of the Pro duct Owners and only 5 % of resp onden ts with more than ten years of exp erience rep ort a p erceiv ed impact in this phase, suggesting the opp osite. Figure 5.4: Estimated time saving with GenAI by co ding task (100 % = no time sa ving, 0 % = full automation) 7 The State of Generativ e AI in Softw are Developmen t: Insights from Literature and a Developer Survey T o quan tify these b enefits at the task lev el, respondents were ask ed to estimate the time required for six common dev elopment activities when using GenAI in relation to working without AI supp ort (see Figure 5.4). The results show that the strongest effects are reported for writing b oilerplate co de and do cumen tation, where 72 % and 69 % of resp onden ts, resp ectiv ely , estimate at least halving the required time. These tw o tasks fall under design and implemen tation as well as op eration and main tenance; the t wo SDLC phases also rated highest in Figure 5.3. W riting new co de shows comparably more mo derate p ositiv e effects, with the largest group (37 %) rep orting a notable reduction to approximately 75 %. Creating unit tests also sho ws substantial gains among the resp onden ts, although the high “cannot assess” rate (28 %) suggests that not all resp onden ts regularly p erform this activit y . Understanding unfamiliar or legacy co de pro duces the most p olarized resp onses, with 29 % rep orting extreme sp eedup to less than 25 % but also the highest “no c hange” rate (11 %) across all tasks. Debugging yields the weak est p erceived sa vings and the highest share of resp onden ts rep orting increased effort (6 %), consistent with the cognitively complex nature of this task. T ogether, those findings confirm the primary role of GenAI as a co ding assistant that accelerates rep etitiv e and standardized developmen t tasks, while also suggesting substan tial p erceiv ed v alue in op eration and main tenance activities such as debugging, log analysis, and legacy co de comprehension. Figure 5.5: GenAI p olicies and guidelines in the resp onden t’s companies As Figure 5.5 shows, the survey results indicate a comparatively mature gov ernance landscap e for GenAI usage across organizations. The largest share of resp ondents (40 %) report the existence of official tools com bined with a formal AI p olicy . An additional 26 % indicate the presence of informal recommendations, suggesting that in roughly t wo thirds of organizations AI usage is at least partially embedded within recognized guidance structures. How ever, gov ernance is not uniformly institutionalized. Around 20 % of participants state that they are unaw are of an y existing p olicies, and approximately 8 % rep ort that no regulation exists at all. Smaller proportions indicate no resp onse, other miscellaneous categories or that AI tools are explicitly forbidden in the company . This distribution rev eals a dual dynamic: While structured gov ernance mec hanisms are already widespread, a non-negligible segment of organizations still exhibits regulatory ambiguit y or limited transparency regarding AI usage p olicies. The qualitative resp onses largely reinforce the quan titativ e findings and provide deep er insight in to p erceived opp ortunities, risks, and anticipated role changes asso ciated with GenAI in softw are developmen t. Across roles and exp erience levels, the dominan t theme is efficiency . Participan ts rep eatedly emphasize time savings through automation of rep etitiv e and low-complexit y tasks such as b oilerplate co de generation, do cumen tation, test creation, debugging supp ort, and co de explanation. Many resp onden ts describ e GenAI as a “sparring partner” that accelerates learning, facilitates onboarding in to new tec hnologies, and supp orts rapid prototyping. Exp erienced developers highlight its v alue in reviewing nov el co de, interfacing with unfamiliar systems, and exploring alternativ e solution paths. Several comments state that AI reduces “toil” and allows developers to fo cus more on architecture, system design, and complex problem solving. A t the same time, concerns are substan tial and remark ably consistent. The most frequently mentioned risks are uncritical adoption of AI-generated co de, loss of deep technical understanding, and data protection and securit y vulnerabilities. Many resp onden ts warn against “blind copying” and emphasize that AI outputs often app ear syntactically correct but may contain logical, arc hitectural, or securit y flaws. Senior practitioners stress that exp erience remains indisp ensable for ev aluating AI-generated artifacts. A recurring theme is the 8 The State of Generativ e AI in Softw are Developmen t: Insights from Literature and a Developer Survey danger of cognitive offloading: i f AI is used as a shortcut rather than as a supp ort to ol, individual learning curv es and long-term problem-solving capabilities ma y deteriorate. Data security—particularly the risk of exp osing sensitive or company-specific information to external mo dels—is also rep eatedly cited as a ma jor concern. Regarding the future of the profession, most participants do not anticipate the disappearance of softw are dev elopment but rather a structural transformation of the role. A common exp ectation is a shift from pure co de writing tow ard architectural thinking, requirement formulation, quality assurance, and orchestration of AI systems. Several resp onden ts predict a p olarization effect: junior roles may decline in num b er, while senior and architect-lev el comp etencies b ecome more v aluable. Others foresee new hybrid roles, such as AI sup ervisors. While a minority expresses skepticism or uncertain ty ab out long-term effects, the prev ailing view is that GenAI will b ecome a standard to ol, and that developers who fail to in tegrate it in to their workflo ws ma y fall b ehind. 6 Conclusion W e examined the impact of GenAI on the SDLC by integrating a systematic literature review with empirical evidence from a cross-sectional survey of 65 softw are developers. The findings consis ten tly demonstrate that AI-assisted co ding to ols exert substantial influence across multiple SDLC phases, with the most pronounced and empirically v alidated effects observed in design, implemen tation, testing, and do cumen tation. W e determine that prior research rep orts measurable pro ductivit y gains, reductions in routine workload, and generally stable or impro ved co de quality when human v alidation mechanisms remain in place. The surv ey results largely corroborate these findings. Resp onden ts iden tify implementation, boiler plate co de, and do cumen tation as the domains with the highest p erceiv ed v alue, confirming the role of GenAI as a pro ductivit y multiplier in syntactically structured and rep etitiv e tasks. In contrast, less structured and ev aluativ e tasks such as planning, requiremen ts analysis, debugging, and testing show considerably lo wer p erceiv ed pro ductivit y gains. Beyond phase-sp ecific effects, the study highlights structural and organizational dimensions of AI integration. GenAI has b ecome deeply embedded in daily workflo ws, yet security concerns remain mo derately elev ated, indicating that developers are aw are of p oten tial risks despite widespread adoption. The gov ernance analysis rev eals that a ma jority of organizations hav e already established either formal AI p olicies or at least informal usage guidelines, yet, a meaningful minority rep ort regulatory ambiguit y or limited a wareness of existing p olicies, p ointing to communication and implemen tation gaps. Imp ortan tly , the findings suggest that GenAI do es not eliminate the need for exp ertise but instead shifts v alue creation within the SDLC. As routine co ding activities b ecome increasingly automated, higher-order compe- tencies, such as requirements precision, architectural reasoning, critical v alidation, and gov ernance ov ersight, gain imp ortance. Exp erience level mo derates p erception: early-career professionals rep ort particularly strong p erceiv ed learning b enefits, whereas exp erienced practitioners tend to frame AI primarily as an efficiency- enhancing tool. This dynamic underscores the imp ortance of structured training and h uman-in-the-lo op mec hanisms to preven t ov erreliance and ensure sustainable skill developmen t. Sev eral limitations warran t consideration. The survey sample is skew ed tow ard younger, developer-centric resp onden ts primarily from the DA CH region. Moreov er, the reliance on self-reported pro ductivit y estimates in tro duces p oten tial resp onse biases, as p erceiv ed and ob jective pro ductivit y gains may diverge. In summary , GenAI is not merely an incremental productivity to ol but a transformativ e force reshaping w orkflows, role profiles, and gov ernance structures within softw are developmen t. How ever, the magnitude and durability of its b enefits dep end on delib erate organizational integration, transparent p olicy framew orks, and contin ued human accoun tability . Promising future research av enues lie in examining longitudinal effects on skill dev elopment, architectural quality , and organizational p erformance. 9 The State of Generativ e AI in Softw are Developmen t: Insights from Literature and a Developer Survey References Mic hael Chui, Roger Rob erts, Lareina Y ee, Eric Hazan, Alex Singla, Kate Sma je, Alex Sukharevsky , and Ro dney Zemmel. The economic p oten tial of generativ e AI: The next pro ductivit y fron tier. T ec hnical rep ort, McKinsey , June 2023. URL https://www.mckinsey.com/capabilities/tech- and- ai/our- insights/ the- economic- potential- of- generative- ai- the- next- productivity- frontier . Ian Go odfellow, Y oshua Bengio, and Aaron Courville. De ep L e arning . Adaptiv e Computation and Machine Learning series. MIT Press, Cambridge, MA, USA, Nov ember 2016. ISBN 978-0-262-03561-3. URL https://mitpress.mit.edu/9780262035613/deep- learning/ . Stefan F euerriegel, Jo c hen Hartmann, Christian Janiesch, and Patric k Zschec h. Generativ e AI. Business & Information Systems Engine ering , 66:111–126, 2024. ISSN 2363-7005. doi:10.1007/s12599-023-00834-7. F red Davis. P erceived Usefulness, Perceiv ed Ease of Use, and User Acceptance of Information T ec hnology. Management Information Systems Quarterly , 13(3), December 1989. ISSN ISSN 0276-7783/ISSN 2162-9730. URL https://aisel.aisnet.org/misq/vol13/iss3/6 . Visw anath V enk atesh and F red D. Da vis. A Theoretical Extension of the T echnology Acceptance Mo del: F our Longitudinal Field Studies. Management Scienc e , 46(2):186–204, F ebruary 2000. ISSN 0025-1909, 1526-5501. doi:10.1287/mnsc.46.2.186.11926. URL https://pubsonline.informs.org/doi/10.1287/ mnsc.46.2.186.11926 . Jane W ebster and Richard T. W atson. Analyzing the Past to Prepare for the F uture: W riting a Literature Review. MIS Quarterly , 26(2):xiii–xxiii, 2002. URL http://www.jstor.org/stable/4132319 . Chitu Ok oli. A guide to conducting a standalone systematic literature review. Communic ations of the Asso ciation for Information Systems , 37, 2015. Guy Paré, Marie-Claude T rudel, Mirou Jaana, and Spyros Kitsiou. Syn thesizing information systems kno wledge: A typology of literature reviews. Information & Management , 52(2):183–199, Marc h 2015. ISSN 03787206. doi:10.1016/j.im.2014.08.008. URL https://linkinghub.elsevier.com/retrieve/pii/ S0378720614001116 . Y air Levy and Timothy J. Ellis. A Systems Approach to Conduct an Effective Literature Review in Supp ort of Information Systems Research. Informing Scienc e: The International Journal of an Emer ging T r ansdiscipline , 9:181–212, 2006. ISSN 1547-9684, 1521-4672. doi:10.28945/479. URL https://www. informingscience.org/Publications/479 . Kim F röhnel, Gabriel Benedikt Thomas Adams, and Rüdiger Zarneko w. F acing the Multifaceted Impact of F ak e Reviews: A Comprehensive Literature Review. Wirtschaftsinformatik 2024 Pr o c e e dings , Jan uary 2024. URL https://aisel.aisnet.org/wi2024/39 . Agnia Sergeyuk, Y aroslav Golubev, Timofey Bryksin, and Iftekhar Ahmed. Using AI-based co ding assistan ts in practice: State of affairs, p erceptions, and wa ys forw ard. Information and Softwar e T e chnolo gy , 178:107610, F ebruary 2025. ISSN 0950-5849. doi:10.1016/j.infsof.2024.107610. URL https://www.sciencedirect. com/science/article/pii/S0950584924002155 . An tonio Collante, Samuel Ab edu, Say edHassan Khato onabadi, Ahmad Ab dellatif, Ebub e Alor, and Emad Shihab. The Impact of Large Language Models (LLMs) on Code Review Pro cess, August 2025. URL http://arxiv.org/abs/2508.11034 . arXiv:2508.11034 [cs]. Erik Brynjolfsson, Danielle Li, and Lindsey R Raymond. Generative AI at work. 2023. URL https: //www.nber.org/system/files/working_papers/w31161/w31161.pdf . F abrizio Dell’Acqua, Edw ard McF owland, Ethan R. Mollick, Hila Lifshitz-Assaf, Katherine Kellogg, Saran Ra jendran, Lisa Kray er, F rançois Candelon, and Karim R. Lakhani. Navigating the Jagged T ec hnological F rontier: Field Exp erimental Evidence of the Effects of AI on Knowledge W ork er Pro ductivit y and Qualit y. SSRN Ele ctr onic Journal , 2023. ISSN 1556-5068. doi:10.2139/ssrn.4573321. URL https: //www.ssrn.com/abstract=4573321 . Kevin KB Ng, Liyana F auzi, Leon Leow, and Jaren Ng. Harnessing the Poten tial of Gen-AI Co ding Assistants in Public Sector Softw are Dev elopment, September 2024. URL . arXiv:2409.17434 [cs]. Md Istiak Hossain Shihab, Christopher Hundhausen, Ahsun T ariq, Summit Haque, Y unhan Qiao, and Brian Wise Mulanda. The Effects of GitHub Copilot on Computing Students’ Programming Effectiveness, Efficiency, and Processes in Brownfield Co ding T asks. In Pr o c e e dings of the 2025 ACM Confer enc e on International Computing Educ ation R ese ar ch V.1 , pages 407–420, Charlottesville USA, August 2025. 10 The State of Generativ e AI in Softw are Developmen t: Insights from Literature and a Developer Survey A CM. ISBN 979-8-4007-1340-8. doi:10.1145/3702652.3744219. URL https://dl.acm.org/doi/10.1145/ 3702652.3744219 . Dev ansh Y adav and Shouvick Mondal. Ev aluating Pre-trained Large Language Mo dels on Zero Shot Prompts for P arallelization of Source Co de, Jan uary 2025. URL https://osf.io/jmevx_v1 . Burak Y etiştiren, Işık Özsoy , Miray A yerdem, and Eray Tüzün. Ev aluating the Code Quality of AI-Assisted Co de Generation T o ols: An Empirical Study on GitHub Copilot, Amazon Co deWhisp erer, and ChatGPT, Octob er 2023. URL . arXiv:2304.10778 [cs]. Mohammad Atif, Kriti Chopra, Ozgur Kilic, Tianle W ang, Zhih ua Dong, Charles Leggett, Meifeng Lin, Paolo Calafiura, and Salman Habib. CelloAI: Leveraging Large Language M odels for HPC Softw are Developmen t in High Energy Ph ysics, August 2025. URL . arXiv:2508.16713 [cs]. Begum Karaci Deniz, Chandra Gnanasam bandam, Martin Harrysson, Alharith Hussin, and Shiv am Sriv asta v a. Unleash developer pro ductivit y with generative AI. T echnical rep ort, McK- insey , June 2023. URL https://www.mckinsey.com/capabilities/tech- and- ai/our- insights/ unleashing- developer- productivity- with- generative- ai . Chandra Gnanasam bandam, Martin Harrysson, and Rikki Singh. Ho w an AI-enabled soft ware pro duct dev elopment life cycle will fuel innov ation. T echnical rep ort, McKinsey , Octob er 2025. URL https: //www.mckinsey.com/industries/technology- media- and- telecommunications/our- insights/ how- an- ai- enabled- software- product- development- life- cycle- will- fuel- innovation . André Barcaui and André Monat. Who is b etter in pro ject planning?Generative artificial intelligence or pro ject managers? Pr oje ct L e adership and So ciety , 4:100101, Decem b er 2023. ISSN 2666- 7215. doi:10.1016/j.plas.2023.100101. URL https://www.sciencedirect.com/science/article/pii/ S2666721523000224 . Laurie Hughes, Reza Kiani Ma vi, Masoud Agha jani, Keith Fitzpatrick, Senali Madugoda Gunaratnege, Sey ed Ashk an Hosseini Shek arabi, Richard Hughes, Ahmad Khanfar, Ahdieh Khatav akhotan, Neda Kiani Ma vi, Keyao Li, Moataz Mahmoud, T egw en Malik, Sashah Mutasa, F arzaneh Nafar, Ross Y ates, Rasha Alahmad, Il Jeon, and Y ogesh K. Dwivedi. Impact of artificial in telligence on pro ject management (PM): Multi-exp ert p ersp ectiv es on adv ancing kno wledge and driving inno v ation tow ard PM2030. Journal of Innovation & Know le dge , 10(5):100772, September 2025. ISSN 2444-569X. doi:10.1016/j.jik.2025.100772. URL https://www.sciencedirect.com/science/article/pii/S2444569X25001179 . Laksh Maggoo, Charles De Pommerol, Ashik Ardeshna, Emmanuel Amiot, Laurent Bensoussan, Jad Haddad, Lindy Kouy oumji, Iv an P alencia, and Benjamin T ubiana. Next F rontier in T ransform- ing T elecom: Practical Benefits of Agentic AI and Digital Agents. T echnical rep ort, Oliver W yman, 2025. URL https://www.oliverwyman.com/content/dam/oliver- wyman/v2/publications/ 2025/march/next- frontier- in- transforming- telecom.pdf . Christoph Lechner, Nikolaus Lang, Siegfried Handsch uh, Olivier Bouffault, and Julian Co oper. Can GenAI do your next strategy task? Not yet. California Management R eview , January 2025. URL https: //cmr.berkeley.edu/2024/09/can- genai- do- your- next- strategy- task- not- yet/ . Phillip W agner and Reinhard W agner. The Evolution of T echnology in Artificial Intelligence and Its Impact on Pro ject Managemen t. In D. Jude Heman th, Utku K ose, Bogdan Patrut, and Mevlut Erso y , editors, Innovative Metho ds in Computer Scienc e and Computational Applic ations in the Er a of Industry 5.0 , pages 268–293, Cham, 2024. Springer Nature Switzerland. ISBN 978-3-031-56322-5. doi:10.1007/978-3-031-56322- 5_19. V artenie Aramali, Namho Cho, F alguni Pande, M. K. S. Al-Mhdawi, Udech ukwu Ojiako, and Abro on Qazi. Generative AI in pro ject managemen t: Impacts on corp orate v alues, employ ee perceptions, and organizational practices. Pr oje ct L e adership and So ciety , 6:100191, Decem b er 2025. ISSN 2666- 7215. doi:10.1016/j.plas.2025.100191. URL https://www.sciencedirect.com/science/article/pii/ S266672152500016X . Orges Cico, Betim Cico, and Andja Cico. AI-assisted Softw are Engineering: a tertiary study . In 2023 12th Me diterr ane an Confer enc e on Emb e dde d Computing (MECO) , pages 1–6, June 2023. doi:10.1109/MECO58584.2023.10154972. URL https://ieeexplore.ieee.org/document/10154972 . ISSN: 2637-9511. Gusta vo Pinto, Cleidson de Souza, Thayssa Ro c ha, Igor Steinmacher, Alb erto de Souza, and Edw ard Monteiro. Dev elop er Exp eriences with a Contextualized AI Co ding Assistant: Usability, Exp ectations, and Outcomes, No vem b er 2023. URL . arXiv:2311.18452 [cs]. 11 The State of Generativ e AI in Softw are Developmen t: Insights from Literature and a Developer Survey Jialiang W ei, Anne-Lise Courbis, Thomas Lambolais, Gérard Dray , and W alid Maalej. On AI-Inspired User In terface Design. IEEE Softwar e , 42(3):50–58, Ma y 2025. ISSN 1937-4194. doi:10.1109/MS.2025.3536838. URL https://ieeexplore.ieee.org/document/10886968 . Pierre-Y ves Glev er, Arnaud Balssa, Stephane Girard, Thilo Hermann, Serge Baccou, and Lorna Neville. Redefining soft ware engineering: The impact of generative AI. T ec hni- cal report, Capgemini, 2024. URL https://www.capgemini.com/insights/research- library/ redefining- software- engineering- the- impact- of- generative- ai/ . Jingzhi Gong, V ardan V osk any an, Paul Bro ok es, F an W u, W ei Jie, Jie Xu, Rafail Giavrimis, Mike Basios, Leslie Kanthan, and Zheng W ang. Language Mo dels for Co de Optimization: Survey, Challenges and F uture Directions, Jan uary 2025. URL . arXiv:2501.01277 [cs]. Vikram Nitin. Using AI to Automate the Mo dernization of Legacy Softw are Applications. In Pr o c e e dings of the 39th IEEE/ACM International Confer enc e on A utomate d Softwar e Engine ering , ASE ’24, pages 2514–2517, New Y ork, NY, USA, Octob er 2024. Asso ciation for Computing Machinery . ISBN 979-8-4007-1248-7. doi:10.1145/3691620.3695610. URL https://doi.org/10.1145/3691620.3695610 . Bingy ang W ei. Requiremen ts are All Y ou Need: F rom Requiremen ts to Code with LLMs. In 2024 IEEE 32nd International R e quir ements Engine ering Confer enc e (RE) , pages 416–422, June 2024. doi:10.1109/RE59067.2024.00049. URL https://ieeexplore.ieee.org/abstract/document/10628487 . ISSN: 2332-6441. Sa vv as Petridis, Michael T erry , and Carrie J Cai. PromptInfuser: How Tightly Coupling AI and UI Design Impacts Designers’ W orkflows. In Pr o c e e dings of the 2024 ACM Designing Inter active Systems Confer enc e , DIS ’24, pages 743–756, New Y ork, NY, USA, July 2024. Asso ciation for Computing Machinery . ISBN 979-8- 4007-0583-0. doi:10.1145/3643834.3661613. URL https://dl.acm.org/doi/10.1145/3643834.3661613 . Dylan Porter, Joanna F. DeF ranco, and Phil Laplante. Requirements Sp ecification Automated Qualit y Analysis: Past, Present, and F uture. Computer , 58(1):101–104, January 2025. ISSN 0018-9162, 1558-0814. doi:10.1109/MC.2024.3480629. URL https://ieeexplore.ieee.org/document/10834143/ . Sida P eng, Eirini Kalliamv akou, Peter Cihon, and Mert Demirer. The Impact of AI on Dev elop er Pro- ductivit y: Evidence from GitHub Copilot, F ebruary 2023. URL . arXiv:2302.06590 [cs]. Illia Solohubov, Artur Moroz, Mariia Y u Tiahuno v a, Halyna H Kyrychek, and Stepan Skrupsky . Accelerating soft ware developmen t with AI: exploring the impact of ChatGPT and GitHub Copilot. 2023. URL https://ceur- ws.org/Vol- 3679/paper17.pdf . Dieter Gerdemann, Sv en Garrels, Sophie Glusac, and Konstan tin Bykhanov. Wh y soft ware + generativ e AI could take v alue creation to a whole new level | Kear- ney , April 2024. URL https://www.kearney.com/service/digital- analytics/article/ why- software- generative- ai- could- take- value- creation- to- a- whole- new- level . Alassane Ndiay e, Xhorxhina T ara j, Bostjan Rob eznik, Gerd Herzog, Christian Gerb er, and Hendrik Purwins. Generativ e AI in Softw are Engineering: T ransforming the Softw are Dev elopment Pro cess. T ec hnical rep ort, A ccenture & German Research Center for Artificial Intelligence, Octob er 2025. Ming Zhong, Xin Sun, F ang Lv, Lulin W ang, Hongna Geng, Lei Qiu, Huimin Cui, and Xiaobing F eng. BePilot: An AI Programming Assistant for Compiler Back end Developmen t. ACM T r ans. Softw. Eng. Metho dol. , August 2025. ISSN 1049-331X. doi:10.1145/3764585. URL https://dl.acm.org/doi/10.1145/3764585 . Just A ccepted. Kevin Zheyuan Cui, Mert Demirer, Sonia Jaffe, Leon Musolff, Sida P eng, and T obias Salz. The Pro ductivit y Effects of Generative AI: Evidence from a Field Exp erimen t with GitHub Copilot. An MIT Explor ation of Gener ative AI , March 2024. doi:10.21428/e4baedd9.3ad85f1c. URL https://mit- genai.pubpub.org/ pub/v5iixksv/release/2 . Max Struever, Matthew Kropp, and Julie Bedard. F rom Dev Sp eed to Business Impact: The Case for AI- Assisted Co ding and Generative Engineering, July 2025. URL https://www.bcg.com/x/the- multiplier/ ai- assisted- coding- generative- engineering . Christof Eb ert and Panos Louridas. Generative AI for Softw are Practitioners. IEEE Softwar e , 40(4):30–38, July 2023. ISSN 0740-7459, 1937-4194. doi:10.1109/MS.2023.3265877. URL https://ieeexplore.ieee. org/document/10176168/ . A o Ding, Gaolei Li, Xiaoyu Yi, Xi Lin, Jianh ua Li, and Chaofeng Zhang. Generative AI for Softw are Security Analysis: F undamentals, Applications, and Challenges. IEEE Softwar e , 41(6):46–54, Nov ember 2024. ISSN 12 The State of Generativ e AI in Softw are Developmen t: Insights from Literature and a Developer Survey 1937-4194. doi:10.1109/MS.2024.3416036. URL https://ieeexplore.ieee.org/abstract/document/ 10634314 . Hammond Pearce, Baleegh Ahmad, Benjamin T an, Brendan Dolan-Gavitt, and Ramesh Karri. Asleep at the Keyb oard? Assessing the Security of GitHub Copilot’s Co de Contributions. Commun. A CM , 68(2):96–105, Jan uary 2025. ISSN 0001-0782. doi:10.1145/3610721. URL https://dl.acm.org/doi/10.1145/3610721 . Gopinath Kathiresan. Automated T est Case Generation with AI: A No vel F ramework for Impro ving Soft ware Quality and Cov erage. W orld Journal of A dvanc e d R ese ar ch and R eviews , 23(2):2880– 2889, 2024. ISSN 2581-9615. doi:10.30574/wjarr.2024.23.2.2463. URL https://wjarr.com/content/ automated- test- case- generation- ai- novel- framework- improving- software- quality- and- coverage . Last Mo dified: 2025-03-24T17:45+05:30. W erner Gey er, Jessica He, Daita Sark ar, Michelle Brachman, Chris Hammond, Jennifer Heins, Zahra Ashktorab, Carlos Rosemberg, and Charlie Hill. A Case Study Inv estigating the Role of Generative AI in Qualit y Ev aluations of Epics in Agile Softw are Developmen t, May 2025. URL 2505.07664 . arXiv:2505.07664 [cs]. Sai Zhang, Zhenc hang Xing, Rongh ui Guo, F angzhou Xu, Lei Chen, Zhao yuan Zhang, Xiaow ang Zhang, Zhiy ong F eng, and Zhiqiang Zhuang. Emp o w ering Agile-Based Generativ e Soft ware Developmen t through Human-AI T eamw ork, Nov em b er 2024. URL . arXiv:2407.15568 [cs]. Rasm us Ulfsnes, Nils Brede Mo e, Viktoria Stra y , and Marianne Sk arp en. T ransforming Softw are Developmen t with Generative AI: Empirical Insights on Collab oration and W orkflo w, F ebruary 2024. URL http: //arxiv.org/abs/2405.01543 . arXiv:2405.01543 [cs]. Anas Bahi, Jihane Gharib, and Y oussef Gahi. In tegrating Generative AI for Adv ancing Agile Soft ware Dev elopment and Mitigating Pro ject Management Challenges. International Journal of A dvanc e d Computer Scienc e and Applic ations , 15(3), 2024. ISSN 21565570, 2158107X. doi:10.14569/IJA CSA.2024.0150306. URL http://thesai.org/Publications/ViewPaper?Volume=15&Issue=3&Code=IJACSA&SerialNo=6 . Philipp Dieb old. F rom Backlogs to Bots: Generative AI’s Impact on Agile Role Evolution. J. Softw. Evol. Pr o c ess , 37(1), January 2025. ISSN 2047-7473. doi:10.1002/smr.2740. URL https://doi.org/10.1002/ smr.2740 . Bilal Nasir and Ijaz Hussain. Automating Agile W orkflo ws: The Role of Multi-Agent LLMs in Mo dern Soft ware Engineering. T ec hnical rep ort, December 2024. Alb ert Ziegler, Eirini Kalliamv ak ou, Shawn Simister, Ganesh Sittampalam, Alice Li, Andrew Rice, Devon Rifkin, and Edward Aftandilian. Pro ductivit y Assessmen t of Neural Co de Completion, May 2022. URL http://arxiv.org/abs/2205.06537 . arXiv:2205.06537 [cs]. Mark Chen, Jerry T worek, Heewoo Jun, Qiming Y uan, Henrique Ponde de Oliv eira Pinto, Jared Kaplan, Harri Edwards, Y uri Burda, Nicholas Joseph, Greg Bro c kman, Alex Ray , Raul Puri, Gretchen Krueger, Mic hael Petro v, Heidy Khlaaf, Girish Sastry , Pamela Mishkin, Bro ok e Chan, Scott Gra y , Nick Ryder, Mikhail Pa vlov, Alethea Po wer, Luk asz Kaiser, Mohammad Bav arian, Clemens Winter, Philipp e Tillet, F elip e Petroski Such, Dav e Cummings, Matthias Plapp ert, F otios Chantzis, Elizab eth Barnes, Ariel Herb ert- V oss, William Hebgen Guss, Alex Nichol, Alex Paino, Nikolas T ezak, Jie T ang, Igor Babusc hkin, Suchir Bala ji, Shantan u Jain, William Saunders, Christopher Hesse, Andrew N. Carr, Jan Leike, Josh A chiam, V edant Misra, Ev an Morik aw a, Alec Radford, Matthew Knight, Miles Brundage, Mira Murati, Katie May er, P eter W elinder, Bob McGrew, Dario Amo dei, Sam McCandlish, Ilya Sutskev er, and W o jciech Zaremba. Ev aluating Large Language Models T rained on Co de, July 2021. URL . arXiv:2107.03374 [cs]. Viktoria Stray , Elias Goldmann Brandtzæg, Viggo T ellefsen Wivestad, Astri Barbala, and Nils Brede Mo e. Dev elop er Pro ductivity With and Without GitHub Copilot: A Longitudinal Mixed-Metho ds Case Study, Septem b er 2025. URL . arXiv:2509.20353 [cs]. Y ue Liu, Y anjie Zhao, Y unbo Lyu, Ting Zhang, Haoyu W ang, and David Lo. "Y our AI, My Shell": Dem ystifying Prompt Injection Attac ks on Agentic AI Co ding Editors, Septem b er 2025. URL http: //arxiv.org/abs/2509.22040 . arXiv:2509.22040 [cs]. Domenico Cotroneo, Cristina Improta, Pietro Liguori, and Rob erto Natella. V ulnerabilities in AI Co de Generators: Exploring T argeted Data Poisoning Attac ks. In Pr o c e e dings of the 32nd IEEE/A CM Inter- national Confer enc e on Pr o gr am Compr ehension , ICPC ’24, pages 280–292, New Y ork, NY, USA, June 2024. Asso ciation for Computing Mac hinery . ISBN 979-8-4007-0586-1. doi:10.1145/3643916.3644416. URL https://dl.acm.org/doi/10.1145/3643916.3644416 . 13 The State of Generativ e AI in Softw are Developmen t: Insights from Literature and a Developer Survey Kieran Norton, Tim Li, Tim Davis, Emily Mossburg, Diana Kearns-Manolatos, Saurabh Ban- so de, Chris Thomas, and Alfons Buxó F errer. How can tec h leaders manage emerging gener- ativ e AI risks to da y while keeping the future in mind? T echnical rep ort, Deloitte, F ebru- ary 2025. URL https://www.deloitte.com/us/en/insights/topics/digital- transformation/ four- emerging- categories- of- gen- ai- risks.html . Zhiguo Ding, Songyang W u, Yilin Zhang, T ao Y ang, Yingna Li, Zh uhua Li, Hang Shao, and Yicong Shi. Co dingCare: AI Co de Generation Security F ramework for Common V ulnerability Mitigation. In Pr o c e e dings of the 2025 4th International Confer enc e on Intel ligent Systems, Communic ations and Computer Networks , ISCCN ’25, pages 307–312, New Y ork, NY, USA, June 2025. Association for Computing Mac hinery . ISBN 979-8-4007-1520-4. doi:10.1145/3732945.3732990. URL https://dl.acm.org/doi/10.1145/3732945. 3732990 . Edw ard Anderson, Geoffrey Park er, and Burcu T an. The Hidden Costs of Co ding With Generative AI. MIT Slo an Management R eview , 67(1), August 2025. URL https://sloanreview.mit.edu/article/ the- hidden- costs- of- coding- with- generative- ai/ . Sergio Moreschini, Elvira-Maria Arv anitou, Elisav et-Persefoni Kanidou, Nikolaos Nikolaidis, Ruoyu Su, Ap ostolos Ampatzoglou, Alexander Chatzigeorgiou, and V alen tina Lenarduzzi. The Evolution of T ec hnical Debt from DevOps to Generative AI: A multiv o cal literature review. Journal of Systems and Softwar e , 231: 112599, January 2026. ISSN 0164-1212. doi:10.1016/j.jss.2025.112599. URL https://www.sciencedirect. com/science/article/pii/S0164121225002687 . Beatriz Cabrero-Daniel, T omas Herda, Victoria Pichler, and Martin Eder. Exploring Human-AI Collab oration in Agile: Customised LLM Meeting Assistan ts, April 2024. URL . arXiv:2404.14871 [cs]. Prana y Ahlaw at, Julie Bedard, Sank alp Damani, Ben F eldman, Elizab eth Lucero, Abhishek Samdaria, and Geoffrey Sipp erly . The Art of Scaling GenAI in Softw are. T echnical rep ort, Boston Consulting Group, May 2024. URL https://www.bcg.com/publications/2024/the- art- of- scaling- genai- in- software . Nicola Döring. F orschungsmetho den und Evaluation in den Sozial- und Humanwissenschaften . Springer, Berlin, Heidelb erg, 2023. ISBN 978-3-662-64761-5 978-3-662-64762-2. doi:10.1007/978-3-662-64762-2. URL https://link.springer.com/10.1007/978- 3- 662- 64762- 2 . 14