Submitted:
14 November 2026
Posted:
17 November 2026
You are already at the latest version
Abstract
The growing use of generative artificial intelligence (GenAI) in software development raises a knowledge-continuity question: how can organizations retain the human competences needed to evaluate, diagnose, and correct software as production is increasingly delegated? This theory-synthesis article connects research on production and evaluation expertise, practice-dependent competence maintenance, and software-process selection. Two literature-grounded conditional propositions identify residual organizational capabilities (P1) and the conditions supporting their continuity over time (P2). Their combined implications are examined through matched Waterfall and Scrum configurations. Prior studies explain when evaluation depends on production, how expertise may deteriorate, and how methods are selected; they do not by themselves determine when a competence-maintenance deficit is imposed by formal process commitments rather than adjustable work arrangements. The synthesis yields a conditional attribution criterion: a limitation can be assigned to the method only if feasible alternatives within that method cannot meet the same requirements under matched constraints. The constructed comparison establishes this distinction, not observed long-term competence retention. Supplementary File S1 provides a five-step managerial decision guide and two worksheets.
Keywords:
generative AI
; knowledge retention
; organizational learning
; competence maintenance
; software development
; human expertise
; Waterfall
; Scrum
; process configuration
; human control
1. Introduction
GenAI can reduce the work developers perform independently while changing the opportunities through which they exercise the competences needed to understand, evaluate, diagnose, and correct software. For information-systems directors, maintaining qualified human control as generative assistance becomes routine is both a work-design and an organizational knowledge-continuity challenge beyond immediate productivity. It concerns how technical decisions, corrective tasks, and feedback are allocated over time.
The management question is whether choosing Waterfall or Scrum changes the opportunities to maintain the competences needed for technical control.
Existing research provides several pieces of the answer, but not the comparison itself. He and Puranam [1] specify when evaluation depends on production experience; Siderius, Shumsky, and Taylor [2] examine the link between verification effort and retained expertise, while Avery and colleagues [3] synthesize risks of degradation of still-needed skills. Studies of software developers document changes in engagement and knowledge exchange [4,5,6,7]. These works establish relevant capability and maintenance questions without determining how Waterfall and Scrum differ in the feasible work arrangements that address them.
Method selection is itself an established information-systems management problem. Veloso, Varajão, and Moreira [8] consider team competence, organizational conditions, and method tailoring among the factors influencing the choice. Pileggi and Bharathy [9] discuss the rebalancing of human and AI roles in Agile development. Neither contribution resolves the different question considered here: when a process commitment, rather than an adjustable configuration, limits opportunities to maintain a capability that remains necessary for human control under GenAI.
The research question is: Does the choice between Waterfall and Scrum determine the conditions under which the human competences necessary for evaluation and corrective intervention are maintained in GenAI-assisted software development, or does the decisive difference lie in how either process is configured? The analysis also considers when constraints on resources, timing, or admissible adaptations make method choice relevant to this objective.
The article integrates three complementary research streams to derive P1 and P2, identify which process requirements are binding, and compare feasible Waterfall and Scrum configurations under common constraints. Its contribution is a conditional criterion for attributing a competence-maintenance limitation to the selected process or to its enactment. Supplementary File S1 translates the criterion into a managerial decision procedure.
2. Literature and the Research Gap
2.1. Expertise, Evaluation, and Maintenance
He and Puranam [1] establish when evaluation expertise can be decoupled from production expertise, depending on the information required for reliable judgment. Faraj and Sproull [10] examine the coordination of specialized knowledge across software teams, while Huang, Xiao, and Vishnoi [11] analyze delegation under different verification conditions. These sources justify asking which capabilities remain necessary for separate human control operations (P1). They do not establish that all productive abilities must be retained, nor do they compare software-process frameworks.
A separate research stream concerns continuity rather than present availability. Bainbridge [12] and Endsley and Kiris [13] identify difficulties associated with reduced human involvement under automation; Siderius, Shumsky, and Taylor [2] model how verification effort and incentives relate to future expertise. Avery, Dinger, and Maier [3] synthesize technology-driven degradation of essential capabilities and its implications for organizational resilience. These contributions support the conditional maintenance question captured by P2, but do not determine whether a Waterfall or Scrum commitment makes the necessary practice feasible.
Maintaining a specified capability can depend on practice or feedback. Coding, demanding review, independent fault diagnosis, simulation, and mentoring may contribute differently according to the capability and its task-specific requirements.
2.2. GenAI in Software Work and Organizational Responses
Shen and Tamkin examine how AI use influences the formation of programming knowledge [14]. Zacharias and colleagues describe how professional developers experience changes in technical engagement [4]; Martin-Lopez and colleagues distinguish completion of coding tasks from understanding the resulting software [5]. Together, these studies show why the organization of technical work matters in addition to the volume of software produced.
Studies of software knowledge processes provide the organizational setting for these conditions. Kirchner and colleagues [6] document changes in knowledge creation and exchange under GenAI; Salomon and colleagues [7] examine displaced and continuing conversations among developers. Yu and Moon [15] discuss experience acquisition by junior engineers, while Krishnan and Thokala [16] consider apprenticeship and substantive review. These studies identify plausible learning and feedback pathways, not a matched method-level attribution test for retaining established evaluation and corrective capabilities.
Code review supports defect detection and knowledge transfer [17], task allocation can diffuse code knowledge [18], and layered supervision distinguishes automated checks from human technical judgment [19]. Avery, Dinger, and Maier [3] synthesize the risk that essential skills still needed by an organization deteriorate under sustained technology reliance. Moe and Stray [20] identify cognitive debt as an emerging concern in AI-augmented Agile teams, referring to a possible weakening of shared understanding. These results concern related but distinct levels: availability of task-specific human capabilities and knowledge shared across a team. Both make substantive practice, feedback, and coordination relevant to competence continuity; neither establishes an unavoidable Waterfall–Scrum difference.
2.3. Process Selection and the Missing Comparison
Veloso, Varajão, and Moreira [8] develop a multidimensional account of information-systems method selection, explicitly including team competence and method tailoring. Thus, neither contextual selection nor the importance of adaptation is new here. The unanswered question in the sources examined is narrower: after GenAI alters work, which process commitments, if any, prevent the organization from supplying practice and qualified access required to sustain a still-needed capability? This requires assessing alternatives available within a method, not comparing two arbitrary implementations.
Process-specific studies address related but different outcomes. Zhang and colleagues [21] develop Agile-oriented human–AI generation to improve requirements alignment. Ha and colleagues [22] compare classical process structures as coordination mechanisms for LLM-based agents, evaluating code quality, cost, and productivity rather than human skill continuity. Pileggi and Bharathy [9] synthesize changes to Agile principles, including human–AI role rebalancing; Moe and Stray [20] identify emerging skill, workflow, and shared-understanding concerns in AI-augmented Agile teams. Ahmad and colleagues [23] examine learning difficulties in student Scrum projects. These contributions make process organization and human competence salient without supplying the present conditional Waterfall–Scrum test of residual capability requirements, maintenance practice, and feasible within-method alternatives.
The Scrum Guide specifies Sprints, usable Increments, and a Definition of Done [24]. General software-lifecycle references recognize alternative arrangements of engineering activities [25,26]. These sources describe process commitments and engineering options, without establishing whether either framework necessarily provides or excludes the work required to maintain a specified human capability.
The gap is therefore an attribution problem, not a claim that expertise, skill degradation, or contextual method selection have been overlooked. Existing work identifies residual control requirements [1,10,11], the importance of continuing practice and feedback [2,3,12,13], knowledge-sharing changes in software teams [6,7,20], and the relevance of tailoring process choices [8,9,21,22]. The synthesis connects these findings to ask whether a specified competence-maintenance deficit persists across feasible configurations of the same method under matched task, quality, time, and resource constraints. Its added result is a conditional criterion separating a limitation imposed by Waterfall or Scrum commitments from one removable through organizational reconfiguration. The claim is restricted to the relationships established by the examined sources and the stated assumptions.
3. Conceptual Method: Theory Synthesis and Comparison Rule
3.1. Unit of Analysis and Fixed Conditions
The unit of analysis is a recurring class of changes to an existing business application across a stated period T. A change to an authorization rule provides a common illustration: the software must meet the same defect and acceptance requirements whether the work is organized through Waterfall or Scrum. The comparison holds the technical task, GenAI capability, acceptance threshold, initial qualified personnel, and external control obligations constant. Delivery commitments, workload, staff availability, and an allowed resource envelope Ω must be specified before a real manager could make a choice. Any additional reviewers or training resources must be included in the comparison.
A method is the selected process framework. A configuration is the concrete arrangement of responsibilities, review and decision points, independent diagnostic work, feedback, and substitute practice through which the framework is enacted. For method m in {W, S}, let C_m(Ω) denote the configurations compatible with its commitments and feasible under Ω. The analysis does not assume that these sets are identical: phase-gated and Sprint-based commitments may impose different timing or coordination constraints. It asks whether each set contains an arrangement that satisfies the same maintenance requirements.
For an evaluation capability e and planning horizon T, M(c,e,T) records whether configuration c meets the specified organizational maintenance conditions under the stated premises; it is not an empirical measure of whether the capability actually persists. P1 determines the human evaluation and intervention capabilities that must be accessible when required; P2 identifies necessary practice and feedback, or adequate substitutes, supporting the relevant competence through T. A distinct corrective capability has to be assessed in its own right. M is evaluated from these conditions, not from the Waterfall or Scrum label. Where a necessary dependency or substitute is unknown, the configuration remains unclassified on that dimension.
3.2. What Counts as a Method Effect
A method-level distinction requires comparing configurations under matched technical tasks, control requirements, GenAI capability, initial expertise, and resource constraints. Differences imposed by process commitments or binding feasibility constraints count as method effects; differences arising from how a team organizes practice, review, and feedback are configuration effects.
A method-specific preference requires a feasible adequate configuration under one framework and evidence that adequate alternatives under the other are infeasible. Adequate configurations in both frameworks establish that either model can meet the stated maintenance conditions, although costs and workload may differ.
A method-level limitation cannot be established by observing an inadequate configuration alone. The comparison must also examine whether an admissible alternative within the same method could satisfy the required conditions under the same resource and control constraints. Only a justified exclusion of such alternatives can support attributing the limitation to the method rather than to its configuration.
3.3. Theory-Synthesis Design and Inferential Status
Following Jaakkola’s theory-synthesis approach [27], this article integrates complementary literatures to clarify a single organizational decision problem. The selection is purposive and problem-driven, not a systematic or exhaustive review. Each stream has a distinct role: evaluation and coordination research [1,10,11] specifies residual human-control capabilities (P1); automation and expertise-maintenance research [2,3,12,13] specifies conditional practice and feedback dependencies over time (P2); method-selection and software-process research [8,9,21,22,24,25,26] identifies formal commitments and configurable activities. Studies of knowledge exchange in software teams [6,7,20] help explain why access, exercise, and feedback have organizational consequences. The integration connects what must remain possible, what sustains that capability, and which configurations can provide the necessary contribution under matched constraints.
The inferential step specific to this synthesis is the within-method alternative test: a deficit observed in one configuration cannot be attributed to Waterfall or Scrum unless an adequate, feasible alternative under the same method and constraints has been ruled out. The constructed W+/S+/W−/S− cases expose this logic, rather than measure actual skill retention. The criterion is conditional on task-specific capability needs, the adequacy of maintenance activities, and feasible resource arrangements.
3.4. Five-Step Inferential Sequence
The argument proceeds as follows: (1) derive P1 from research on task requirements, evaluation, and coordination to identify residual human capabilities; (2) derive P2 from research on practice, feedback, and continuity to specify when current evaluative competence may not suffice over T; (3) identify which features belong to the bounded Waterfall and Scrum process commitments and which remain configurable; (4) compare method-compatible configurations holding the relevant obligations, initial expertise, and resources constant; and (5) derive the method-choice result from the existence or absence of adequate feasible configurations. Steps 1 and 2 yield intermediate conditional propositions; step 5 yields the paper’s distinctive process-comparison finding.
4. Derivation of P1 and P2: Conditions for Competence Maintenance
4.1. Constructs, Source Roles, and Bounds
A human competence is treated here as a task-specific capability: an ability defined for a particular task, quality threshold, and set of operating conditions. Independent production capability G_g is the ability to produce the relevant output without delegating its generative component to GenAI, while ordinary non-generative tools remain available. Evaluation capability is the ability to judge the result against specified criteria. Corrective intervention includes rejecting, repairing, or recovering the work, informed when necessary by diagnosis. Verification gathers or applies evidence about criteria; evaluation makes the judgment; control additionally requires authority and timely means to act. These operations may call for different capabilities in different roles.
He and Puranam delimit conditions for separating production and evaluation expertise [1]; Faraj and Sproull establish the importance of coordinating specialist knowledge [10]; Huang, Xiao, and Vishnoi analyze how verification affects delegation [11]; and Kwarsick proposes capacity-accounting tools for retained human oversight [28,29]. Earlier automation studies [12,13] and the expertise model of Siderius and colleagues [2] inform the temporal relationship between practice and maintenance.
A task performer’s skill requirement, a capability possessed by the organization, its timely availability for a residual operation, and the activity through which it is maintained are distinct objects. An adequate alternative must satisfy the same control standard and decision deadline; nominal staff availability and untested training are insufficient evidence of adequacy.
4.2. Derivation of P1: From Individual Requirements to Organizational Availability
Let Q_g(A) denote the capabilities required from the person assigned to task g in a GenAI-assisted mode. Let Q_O(A) denote the capabilities required by separate human evaluation and intervention operations in the activity. The two sets describe different functions and must not be assumed identical. If independent production capability G_g is no longer necessary for the assisted performer, G_g may be absent from Q_g(A); it does not follow that G_g is absent from Q_O(A). Whether a remaining operation requires G_g is an independent, task-specific premise. This separation follows the conditional account of production and evaluation [1] and the need to coordinate expertise across roles [10].
The derivation holds three premises apart. First, the assisted performer can meet the specified task requirement without G_g. Second, a separately identified evaluation or corrective operation remains necessary at the unchanged quality standard and deadline. Third, that operation requires G_g and no available substitute can meet the same requirement. The third premise must be substantiated through task analysis or other relevant evidence; it cannot be established merely because the organization already possesses G_g. Under these conditions, removing organizational access to G_g would make the residual operation impossible at the required standard, even though assisted production remains assignable to its performer.
Proposition 1 (P1) — When GenAI makes a production capability no longer necessary for the person performing a task, that capability remains necessary to the organization if another evaluation or intervention operation within the activity continues to require it and no alternative capable of meeting the same requirements is available.
P1 therefore derives a necessary organizational-availability condition, not a universal rule that every reviewer must be able to produce independently. The capability may be held by a specialist, distributed among coordinated roles, or accessed through qualified external support, provided the required judgment or intervention can actually occur in time. If the separate operation does not require G_g, or an adequate alternative exists, this particular necessity conclusion does not follow.
4.3. Derivation of P2: From Current Competence to Its Maintenance over Time
P1 concerns a capability required for a residual operation at a given time; it does not establish that people will remain able to perform it in the future. Let e identify a specified human evaluation capability, p a practice supplying diagnostic experience or feedback relevant to e, and T a finite planning horizon. Assume initially that e meets the required threshold. Earlier automation studies connect reduced routine involvement with difficulties in later intervention [12,13]; He and Puranam [1] identify circumstances in which production experience matters for evaluation, and Siderius, Shumsky, and Taylor [2] explicitly model the connection between verification effort and retained expertise. These contributions motivate a dependence that must still be established for the specific capability, practice, and period.
Here the premises are: (i) e remains required for control during T; (ii) p supplies a necessary contribution to maintaining e at the required threshold; (iii) GenAI delegation reduces that contribution; and (iv) neither the practice remaining nor any alternative supplies an adequate replacement. From the initially adequate level of e one therefore cannot infer its continuing adequacy. Practice may be partly retained or replaced by substantive review, independent diagnosis, simulation, mentoring, or other activities only when the substitute is shown adequate for e.
Proposition 2 (P2) — If delegation to GenAI reduces a practice that makes a necessary contribution to maintaining an evaluation capability required for control, and that contribution is not replaced, the initial availability of the capability is not sufficient to sustain continuity of control over the period considered.
P2 makes the temporal requirement explicit. The relevant contribution to evaluation competence must be identified for the task and time horizon. Evidence on developer experience, code understanding, and programming-skill formation [4,5,14] helps identify practices worth examining. If a substitute supplies adequate experience and feedback, preserving the previous production practice is not required by P2.
4.4. Joint Implication for the Waterfall–Scrum Comparison
P1 and P2 establish two requirements for comparing configurations: (a) timely access to the human capabilities needed for evaluation and intervention, and (b) adequate practice and feedback, or substitutes, to maintain the specified evaluation capability over T. Distinct corrective capabilities require separate assessment when relevant. The comparison then asks which conditions follow from formal process commitments and which depend on configuration.
5. Results: Waterfall–Scrum Comparison
5.1. Process Commitments and Configurable Practices
Waterfall W is defined here as a bounded sequential, phase-gated process model: requirements and design precede implementation, with agreed review and handoff gates. Independent fault diagnosis, correction, and feedback can be scheduled during implementation and verification before final handoff; a gate does not prescribe their depth. Scrum S requires Sprints, usable Increments, and a Definition of Done [24]. Technical quality required for Done must be established before an Increment is complete, while the Scrum Guide leaves technical review techniques and diagnostic practice to the team. Scrum’s Sprint Review inspects the Increment and considers future adaptations; technical code review is a separate engineering activity. In both processes, the organization determines the specific diagnostic practice, review depth, and feedback arrangements. Table 1 distinguishes the process commitments from the activities requiring configuration.
The comparison therefore examines substantive diagnostic work and qualified access rather than the number of gates or Sprints.
5.2. Matched Adequate Configurations in Waterfall and Scrum
Consider recurring changes to an authorization rule over T. The defect classes, acceptance criteria, GenAI capability, human-control obligations, and initial qualified specialists are fixed across both methods. For P1, suppose the assisted performer no longer needs independent production capability G_g, while qualified diagnosis of a consequential authorization fault still requires G_g and has no adequate substitute. Organizational access to G_g is then necessary by the decision time. For P2, let e be the capability to evaluate and diagnose those faults, whose exercise in this example requires G_g, and suppose independent fault reconstruction and corrective feedback make a necessary contribution to maintaining e. For this construction, assume these activities are adequate over T for e and for any distinct corrective capability required. Allocate the same qualified effort to this work in both methods and stipulate decision windows compatible with the common available resources.
In W+, the team performs the stipulated independent authorization-fault reconstruction during the implementation/verification work, documents the diagnosis, obtains qualified correction and feedback, and completes any required technical approval by the phase handoff. Repeating that activity over T is compatible with the bounded phase-gated sequence. In S+, the same qualified people perform an equivalent reconstruction and corrective debrief during Increment development before the relevant work meets the Definition of Done, again repeated over T. Neither schedule adds experts or training time available only to one method; the common resource envelope is assumed to permit the respective control windows. The positive constructions differ in when the work is embedded, not in the required defect, standard, practice contribution, or qualified effort.
Under those premises, W+ provides qualified judgment and intervention by the designated handoff, while S+ provides them before the corresponding Increment meets the Definition of Done. Each configuration includes the stipulated diagnostic work and feedback. Waterfall and Scrum thus both admit, under the matched assumptions, configurations meeting the same competence-maintenance conditions. This establishes conditional feasibility, not observed retention of those competences.
5.3. Within-Method Variation in Maintenance Practice
Now retain the same initial experts, task, quality requirement, GenAI capability, and resource envelope, but remove the recurrent independent fault-reconstruction exercise and its useful feedback without an adequate substitute. W− still has the prescribed phases and gates, with routine automated checks and sign-offs; S− still builds Increments to a Definition of Done, with routine technical checks. In both, the experts remain available and authorized to handle consequential faults when necessary at the beginning of T. Their current availability is not the same as the sustained practice stipulated to maintain their evaluative competence. Current acceptance controls remain in place even though the specified maintenance practice is omitted.
In both negative configurations, qualified specialists can initially carry out the required control, so P1 can be satisfied at that point. The stipulated diagnostic exercise and feedback are absent, and no adequate substitute is provided. Given the specified maintenance dependence, W− and S− fail to meet the full P2-based conditions for continuity through T. These cases demonstrate the absence of a stipulated necessary maintenance contribution, not observed skill deterioration. Table 2 summarizes the four matched configurations and their conditional implications.
5.4. Feasibility Under Binding Constraints
Process commitments can affect competence-maintenance opportunities through the timing and coordination of qualified technical work. Phase-gated handoffs and Sprint-based completion requirements may create different demands on specialist availability and diagnostic practice. These differences matter when, under the same resource constraints, they limit the feasible arrangements through which the necessary competences can be exercised and maintained. If admissible changes within the same method remove the limitation, the difference is attributable to configuration rather than to an unavoidable process constraint.
5.5. Fifth Step: Comparative Result and Method-Choice Rule
Because W+ and S+ are admissible under the same stipulated resource constraints as W− and S−, the inadequacy of the negative configurations does not establish a method-level limitation. Under these assumptions, the deficits can be removed through within-method reconfiguration.
Comparative result — The comparison establishes that the formal identity of Waterfall or Scrum is neither sufficient to guarantee the stipulated maintenance conditions nor, under the matched assumptions, necessary for their realization. Whether a method-specific difference exists depends on the feasible configurations permitted by its commitments and the applicable constraints. Such a difference can be attributed to the method only when adequate alternatives within the apparently disadvantaged framework have been ruled out. Table 3 sets out the resulting conditional method-choice cases.
6. Discussion and Implications for Information-Systems Management
6.1. What the Theory Synthesis Adds Beyond Prior Work
The article does not claim to discover the conditions for separating production from evaluation: He and Puranam [1] examine them directly. Nor is skill deterioration under AI new: Siderius and colleagues [2] model its dynamics, while Avery and colleagues [3] synthesize the implications of degradation of still-required abilities. Kirchner and colleagues [6], Salomon and colleagues [7], and Moe and Stray [20] document or synthesize how GenAI changes knowledge exchange, learning, and shared understanding in software teams. These contributions establish why capability continuity matters, but do not determine the process-choice attribution made here.Veloso and colleagues [8] already incorporate skills and method tailoring into method selection; Pileggi and Bharathy [9] revisit Agile principles and human–AI roles; Ha and colleagues [22] compare process scaffolds for LLM-generated software on quality and cost. The present synthesis addresses a different outcome and reverses the analytical direction: from competences as inputs to choosing a method to the feasible maintenance of those competences as a criterion for evaluating that choice. By combining P1 and P2 with the admissible configurations of Waterfall and Scrum, it distinguishes a shortcoming of one enacted arrangement from a limitation attributable to a formal process commitment under binding constraints. This conditional attribution criterion—not the general insight that organizational context matters—is the specific theoretical addition. Supplementary File S1 translates the criterion into a five-step decision procedure for safeguarding organizational access to necessary expertise.
6.2. Practical Implications
In phase-gated Waterfall, managers should examine the diagnostic work, technical review, corrective activity, and feedback performed before handoffs. In Scrum, they should examine the corresponding activities within Increments and before the Definition of Done. These choices also shape opportunities for sharing and exercising technical knowledge across roles; neither the number of gates nor the number of Sprints demonstrates long-term competence retention.
Where the required practice is inadequate, managers can examine protected diagnosis, substantive review, feedback, specialist reassignment, and tested substitutes within the current process, recording their recurring resource requirements. Supplementary File S1 provides a five-step procedure for this assessment.
6.3. Limitations and Testable Implications
The comparison is conceptual, not an empirical test of long-term competence retention or loss. Its conclusions depend on the specified control requirements, initial expertise, sufficient effectiveness of the stipulated diagnostic practice and feedback (or adequate substitutes), and comparable timing and resource constraints. These premises must be assessed in a real organization. The article does not estimate rates of skill deterioration, the quantity of practice needed, or systematic differences in effectiveness and cost between the methods. Those questions require longitudinal evidence from comparable settings, including independent assessments of diagnostic and corrective performance and records of the work and feedback actually provided. A claimed method-level limitation would be overturned in a given case by identifying an admissible, feasible configuration within that same method that meets the specified requirements under the same constraints.
The scope is recurring changes to existing software with specified human evaluation and intervention obligations. Waterfall is represented by a particular sequential, phase-gated enactment; other plan-driven, hybrid, or locally constrained processes may admit different configurations. Evaluation and corrective intervention may require distinct capabilities and maintenance practices. Broader objectives—such as innovation, supplier independence, or developing new generations of specialists—may warrant preserving additional skills beyond those examined here.
The comparison also suggests a practical research agenda: identify the diagnostic activities and feedback that are indispensable to each competence, examine the feasible adaptations under each process, and distinguish constraints built into the framework from restrictions introduced by organizational implementation. This is the information needed to determine when a method change could offer an actual competence-maintenance advantage.
7. Conclusions
The central result of this theory synthesis is a conditional attribution criterion for competence maintenance in GenAI-assisted software development. Under matched task and resource conditions, neither the Waterfall nor Scrum label alone determines whether the specified requirements can be satisfied; feasible process configurations and binding commitments must be examined together.
This result follows from two literature-grounded propositions. P1 identifies when a production competence made unnecessary for the assisted performer remains necessary for organizational evaluation or intervention. P2 identifies the practice and feedback required to maintain evaluation capability over time. In the matched comparison, W+ and S+ satisfy the resulting conditions, while W− and S− omit a necessary maintenance contribution. Both models therefore permit different arrangements for sustaining the required human capabilities.
For information-systems managers, the priority is to organize meaningful diagnosis, corrective responsibilities, feedback, and timely specialist access within the chosen process. Changing methodologies becomes justified for competence-maintenance reasons when binding constraints permit an adequate arrangement in one framework but prevent it in the other. Supplementary File S1 provides the corresponding decision procedure.
Funding.: This research received no specific external funding.
Supplementary Materials
The following supporting information can be downloaded at the website of this paper posted on Preprints.org. Supplementary File S1, Configuration-Based Managerial Guide for Competence Maintenance under GenAI, contains a five-step application procedure and two worksheets.
Author Contributions
Conceptualization, methodology, analysis, writing—original draft, review and editing: A.S.
Institutional Review Board Statement
Not applicable. This conceptual study did not involve human participants.
Informed Consent Statement
Not applicable.
Data Availability Statement
No new empirical data were collected or analyzed in this conceptual study. The analytical premises, constructed comparisons, and decision conditions are presented in the article; the accompanying decision guide is provided in Supplementary File S1. Data sharing is not applicable to this article.
Conflicts of Interest
The author is employed by Energee3 S.r.L., Reggio Emilia, Italy. The author declares no additional financial conflicts of interest relevant to this conceptual study.
Use of GenAI: OpenAI ChatGPT (including GPT-6, used in 2026) assisted with organization of existing materials, drafting, language editing, and consistency checks. No language-model output was treated as empirical evidence. The author reviewed and edited the outputs and retains responsibility for source interpretation, inference assessment, editorial decisions, and the final content.
References
- He, T.; Puranam, P. Can You Judge without Doing? The Conditions for Decoupling Expertise in Evaluation from Production. In INSEAD Working Paper 2026/36/STR; 2026. [Google Scholar] [CrossRef]
- Siderius, J.; Shumsky, R.; Taylor, A. Use it or Slowly Lose it: Expertise Atrophy with Organizational AI Usage. In Working paper; 2026. [Google Scholar] [CrossRef]
- Avery, A.; Dinger, M.; Maier, C. Technology-Driven Skill Degradation and Organizational Resilience in the Era of Automation and AI: An Integrated Systematic Review and Research Agenda. ACM SIGMIS Database 2026, 57(3), 42–80. [Google Scholar] [CrossRef]
- Zacharias, J.; Popova, A.; von Zahn, M.; Chen, J.; Hinz, O. Developers’ Dilemma: Opportunities and Pitfalls of Generative AI for Software Development. In Business & Information Systems Engineering; 2026. [Google Scholar] [CrossRef]
- Martin-Lopez, A.; Tufano, R.; Guglielmi, E.; Sánchez, A.B.; Sanz, A.L.; Scalabrino, S.; Oliveto, R.; Segura, S.; Bavota, G. More Code, Less Understanding? On the Impact of AI Assistants on Developers’ Productivity and Code Ownership. IEEE Trans. Softw. Eng. 2026, 52(7), 2173–2187. [Google Scholar] [CrossRef]
- Kirchner, K.; Bolisani, E.; Kassaneh, T.C.; Scarso, E.; Taraghi, N. Generative AI Meets Knowledge Management: Insights From Software Development Practices. Knowl. Process Manag. 2025, 32(4), 223–235. [Google Scholar] [CrossRef]
- Salomon, M.; Koshchenko, E.; Sergeyuk, A.; Holmes, R.; Murphy, G.C.; Fritz, T. From Disruptions to Discussions: How GenAI Impacts Human Interactions in Software Development. IEEE Trans. Softw. Eng. 2026, 52(7), 2095–2110. [Google Scholar] [CrossRef]
- Veloso, M.; Varajão, J.; Moreira, F. A Comprehensive Model for the Selection of Methods in Information Systems. Manag. Rev. Q. published online. 2026. [Google Scholar] [CrossRef]
- Pileggi, S.F.; Bharathy, G. AI-gile: Revisiting Agile Principles in the Era of AI. Inf. Softw. Technol. 2026, 194, 108073. [Google Scholar] [CrossRef]
- Faraj, S.; Sproull, L. Coordinating Expertise in Software Development Teams. Manag. Sci. 2000, 46, 1554–1568. [Google Scholar] [CrossRef]
- Huang, L.; Xiao, W.; Vishnoi, N. K. Delegation and Verification Under AI. In Proceedings of the 43rd International Conference on Machine Learning, PMLR, 2026; 306, pp. 48039–48071. Available online: https://proceedings.mlr.press/v306/huang26bp.html.
- Bainbridge, L. Ironies of Automation. Automatica 1983, 19(6), 775–779. [Google Scholar] [CrossRef]
- Endsley, M. R.; Kiris, E. O. The Out-of-the-Loop Performance Problem and Level of Control in Automation. Hum. Factors 1995, 37(2), 381–394. [Google Scholar] [CrossRef]
- Shen, J. H.; Tamkin, A. How AI Impacts Skill Formation. arXiv 2026, arXiv:2601.20245v1. [Google Scholar] [CrossRef]
- Yu, S.; Moon, T. Who Will Become the Next Senior? How Generative AI Erodes the Development Pathway in Software Engineering. arXiv 2026, arXiv:2607.17067. [Google Scholar] [CrossRef]
- Krishnan, S.; Thokala, G.B. Who Trains the Next Senior? Managing Software Engineering When AI Makes Code Cheap. SSRN Work. Pap. 2026, 7083078. [Google Scholar] [CrossRef]
- Bacchelli, A.; Bird, C. Expectations, Outcomes, and Challenges of Modern Code Review. In Proceedings of the 35th International Conference on Software Engineering (ICSE), 2013; pp. 712–721. [Google Scholar] [CrossRef]
- Etemadi, V.; Bushehrian, O.; Robles, G. Task Assignment to Counter the Effect of Developer Turnover in Software Maintenance: A Knowledge Diffusion Model. Inf. Softw. Technol. 2022, 143, 106786. [Google Scholar] [CrossRef]
- Stolze, M.; Strässle, M. When Review Alone No Longer Scales: Layered Supervision in AI-Assisted Software Engineering. Proc. 20th Int. Symp. Empir. Softw. Eng. Meas. (ESEM 2026) LIPIcs 394, 90:1–90:12. [CrossRef]
- Moe, N.B.; Stray, V. AI-Augmented Agile Teams: A Research Agenda and Summary of the Human-AI Collaboration Workshop. In Agile Processes in Software Engineering and Extreme Programming—Workshops (XP 2026); Lecture Notes in Business Information Processing; first online; Springer: Cham, 2027; vol. 596, pp. 51–60. [Google Scholar] [CrossRef]
- Zhang, S.; Xing, Z.; Guo, R.; Xu, F.; Chen, L.; Zhang, Z.; Zhang, X.; Feng, Z.; Zhuang, Z. Empowering Agile-Based Generative Software Development through Human-AI Teamwork. ACM Trans. Softw. Eng. Methodol. 2025, 34(6), 156. [Google Scholar] [CrossRef]
- Ha, D. M.; Kien, P. T.; Quan, T.; Nguyen-Duc, A. Evaluating Classical Software Process Models as Coordination Mechanisms for LLM-Based Software Generation. arXiv. 2025. Available online: https://arxiv.org/abs/2509.13942.
- Ahmad, M.O.; Al-Baik, O.; Abu Al-haija, M.; Husein, A.; Morales-Trujillo, M. Artificial Intelligence Learning Paradox: How Over Reliance on It Undermines Mastery in Agile Software Project Management. Procedia Comput. Sci. 2026, 278, 1877–1884. [Google Scholar] [CrossRef]
- Schwaber, K.; Sutherland, J. The Scrum Guide: The Definitive Guide to Scrum: The Rules of the Game. November 2020. Available online: https://scrumguides.org/scrum-guide.html.
- IEEE/ISO/IEC. ISO/IEC/IEEE 12207:2026; Systems and Software Engineering—Software Life Cycle Processes. IEEE, 2026. Available online: https://standards.ieee.org/ieee/12207/11416/.
- IEEE Computer Society. Guide to the Software Engineering Body of Knowledge (SWEBOK Guide), Version 4.0a; IEEE Computer Society, 2025; Available online: https://www.computer.org/education/bodies-of-knowledge/software-engineering.
- Jaakkola, E. Designing conceptual articles: four approaches. AMS Rev. 2020, 10, 18–26. [Google Scholar] [CrossRef]
- Kwarsick, J. The Diffusion Constraint. Working paper, SSRN 7464739. 2026. Available online: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=7464739.
- Kwarsick, J. The Capacity Ledger: A Measurable Specification for Capacity-Degradation Risk in Human-AI Collaboration. Work. Pap. 2026, SSRN 7011518. [Google Scholar] [CrossRef]
- Cui, K. Z.; Demirer, M.; Jaffe, S.; Musolff, L.; Peng, S.; Salz, T. The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers. Manag. Sci. published online. 2026. [Google Scholar] [CrossRef]
- Becker, J.; Rush, N.; Barnes, E.; Rein, D. Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. arXiv 2025, arXiv:2507.09089. [Google Scholar] [CrossRef]
Table 1.
Method commitments and competence-maintenance variables.
| Dimension | Sequential, phase-gated Waterfall W | Scrum S | What requires configuration |
| Delivery commitment | Planned phases and specified handoff or acceptance gates | Sprints, usable Increments, Definition of Done | Actual timing of qualified technical decisions |
| Qualified judgment | Specialists may review during phases or at gates | Team and/or accessible specialists may evaluate before Done | Authority, availability, depth of independent checking |
| Diagnostic practice | Can occur during construction, review and correction | Can occur during Increment development, review and correction | Necessary practice, intensity, task relevance, feedback |
| Maintenance substitutes | Training, simulation or protected diagnosis can be scheduled | The same types of substitutes can be scheduled | Demonstrated sufficiency over T and recurring resource cost |
Table 2.
Paired configurations for the same software obligation.
| Enactment | Maintenance-relevant work actually organized | Conditional conclusion | Implication for the method-level claim |
| W+ | At each change, independent authorization-fault reconstruction, qualified correction and feedback during implementation and verification; completed before the planned handoff. | Conditionally meets P1/P2-based maintenance requirements under assumed practice efficacy and compatible resources. | Waterfall permits the required diagnostic practice within a phase-gated configuration. |
| S+ | The same fault reconstruction, qualified correction and feedback during Increment work, completed before the relevant Definition of Done decision. | Conditionally meets the same requirements under assumed practice efficacy and compatible resources. | Scrum permits the same practice within Increment and Done commitments. |
| W− | Phase gates and current qualified cover remain, but recurrent independent reconstruction and useful feedback are omitted; no adequate substitute. | Specified necessary maintenance contribution absent under P2; initial capability loss not established. | Phase gates by themselves do not ensure the maintenance contribution. |
| S− | Done and current qualified cover remain, but recurrent independent reconstruction and feedback are omitted; no adequate substitute. | Specified necessary maintenance contribution absent under P2; initial capability loss not established. | Sprint and Done commitments by themselves do not ensure the maintenance contribution. |
Table 3.
What the manager may conclude after examining feasible configurations under the same constraints.
Table 3.
What the manager may conclude after examining feasible configurations under the same constraints.
| Feasible adequate W configuration? | Feasible adequate S configuration? | Competence-maintenance inference | Consequence for method choice |
| Yes | Yes | Both methods can realize the specified maintenance conditions | No need to switch methods solely for this objective; compare other costs and project needs |
| Yes | Demonstrably no | Bounded advantage of W under the declared constraints | W may be preferred for this objective; document why S adaptations fail |
| Demonstrably no | Yes | Bounded advantage of S under the declared constraints | S may be preferred for this objective; document why W adaptations fail |
| Demonstrably no | Demonstrably no | Neither feasible approach meets the stated need | Change resources, controls, or the proposed work organization |
| Unassessed or uncertain | Any / unassessed | No supported comparison of maintenance feasibility | Obtain missing evidence; do not infer superiority or equivalence |
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. |
© 2026 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.