You’re listening to “A Framework for Evaluating GenAI Adoption and Use in Software Engineering,” by Liang Yu and colleagues. Published in IEEE Transactions on Software Engineering in August 2026. Abstract. Generative Artificial Intelligence (GenAI) is increas-ingly integrated into software products to enable new features and user capabilities, from early exploration to operational deployment. GenAI adoption as a component within a soft-ware system introduces quality risks because GenAI outputs are probabilistic, prompt-sensitive, and may drift after release. Organizations, therefore, need to decide what to evaluate, when to evaluate, and who owns quality evaluation activities across software design, development, and operations. ISO/IEC 25059 standard distinguishes between software product quality (e.g., usability) and quality-in-use (e.g., satisfaction) for AI-enabled software, yet it provides limited operational guidance for these evaluation activities. We therefore investigate how industrial software teams adopt and use GenAI models in the software systems they build and operate, and how they evaluate system qualities when deciding to adopt GenAI during development and after deployment. We do not benchmark the underlying GenAI model itself. In this study, we conducted 19 semi-structured inter-views in two software development companies. We triangulated the interviews with archival data (15 internal documents and 184 internal wiki/web pages) to capture GenAI adoption steps, quality concerns, evaluation practices, and role responsibilities. Our findings describe a three-phase adoption process – Ideation, Development, and Operation – highlighting where quality eval-uations occur, which criteria are used, and how evaluation responsibilities are distributed. Based on observed practices and using ISO/IEC 25059 as an organizing lens, we synthesize a process-oriented quality evaluation framework. This framework maps metrics to explicit gatekeeping, validation, and monitoring checkpoints, bridging abstract ISO quality characteristics with engineering workflows. We applied the framework in a GenAI-enabled software product (SE4AI) use case and reported how it supported structured evaluation activities. We also observed that quality evaluations span legal, security, development, QA, and operations, but ownership is fragmented across phases. We therefore propose a GenAI Quality Lead responsibility (often assignable to an existing senior role) to coordinate criteria, evidence, and traceability across quality evaluation activities. The results contribute to Software Engineering for AI (SE4AI) by clarifying how teams can measure qualities when building software that adopts and uses GenAI. Liang Yu and Emil Alégroth are with the Department of Software Engi-neering, Blekinge Institute of Technology, 371 79 Karlskrona, Sweden (e-mail: the email address; the email address). Panagiota Chatzipetrou is with Örebro University, 701 82 Örebro, Sweden (e-mail: the email address). Tony Gorschek is with Blekinge Institute of Technology, 371 79 Karlskrona, Sweden, and also with Fortiss, 80805 Munich, Germany (e-mail: tony. the email address). I. INTRODUCTION. G ENERATIVE Artificial Intelligence (GenAI) with Large Language Models (LLMs) has been rapidly adopted in software engineering. This adoption covers two com-plementary perspectives: Software Engineering for AI (SE4AI), which provides practices for building reliable AI-enabled sys-tems, and AI for Software Engineering (AI4SE), where GenAI assists traditional development tasks. From requirements analysis to code generation, GenAI is transforming how software is developed and maintained. Industrial GenAI adoption typically starts with prompt-based experimentation and iterative refinement, gradually maturing into production-ready implementations. This adoption introduces risks for soft-ware quality. The main challenge is to define what to evaluate, when, and how — with adoption activities and monitoring — for product quality and quality-in-use. Unlike traditional determin-istic software, GenAI-based software produces probabilistic outputs that can vary across runs, even with identical inputs. The generated content can be plausible but factually incorrect, biased, or potentially harmful, raising concerns around accu-racy, security, and compliance. These characteristics – non-determinism, prompt and data sensitivity, and behavioral drift – limit the applicability of traditional software quality assurance techniques, which assume fixed logic and predictable outputs. ISO/IEC 25059 focuses on AI-enabled software, distin-guishing between product quality (e.g., functional suitability, reliability, security) and quality-in-use (e.g., effectiveness, ef-ficiency, freedom from risk). ISO standards define what quality means, but they do not specify when a team should verify each quality characteristic or what evidence is required to transition from a pilot to production. For instance, teams performed legal, security, and output evaluations, but these evaluations were scattered across roles and were rarely recorded in a way that supports later review. This creates a gap between standards and industrial engineering work: teams need process-oriented guidance that ties quality evaluations to adoption steps and role ownership. In this study, we examine the adoption of GenAI for software products, specifically how teams integrate and operate GenAI techniques within the software systems they develop. We focus © 2026 The Authors. This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see the linked source on engineering practices and quality evaluation activities across the development and operational lifecycle. A case study was conducted, which combined interviews with an analysis of internal web pages and documents, enabling us to capture both documented processes and practices. Based on the observed AI4SE usage practices, we derived SE4AI guidance in the form of a quality evaluation framework aligned with ISO/IEC 25059. Such guidance supports recurring decisions: whether a use case can proceed past ideation, whether a GenAI feature is ready for release, and whether drift or failures require rollback or revalidation after deployment. Our framework provides a lightweight structure for these decisions by linking evaluation targets, activities, and accountable roles. Organizations can use it to reduce ad hoc gatekeeping, enhance the traceability of quality decisions, and align internal policies with external standards. Our primary contribution is a Quality Evaluation Frame-work that synthesizes industrial practice with ISO standards. Specifically, we contribute: 1) Process. discovery (The “When”): Empirical evi-dence on the multi-phase adoption process and its phases (Ideation, Development, Operation) used in industry. 2) Quality characterization (The “What”): A mapping. of practitioner concerns (e.g., legal/security) to ISO/IEC 25059 quality characteristics. 3) Role. definition (The “Who”): The proposal of a GenAI Quality Lead role to coordinate cross-functional evaluation. 4) Framework. verification: A verification of the framework on a real-world use case, demonstrating its feasibility. The remainder of this study is structured as follows. Sec-tion II presents background information and related work on GenAI in software engineering and quality evaluation ap-proaches. Section III describes our research methodology, in-cluding research questions, case company details, and data collection and analysis methods. Section IV presents the pro-posed quality evaluation framework, and Section V illustrates the framework verification through a use case implementa-tion. Section VI addresses threats to validity. Section VII dis-cusses implications, including the potential emergence of new roles such as AI Quality Lead. Section VIII concludes the paper with a summary of contributions and future research directions. II. BACKGROUND AND RELATED WORK. A. Background GenAI models, such as GPT-4o, have been used in soft-ware engineering activities, including code generation, documentation, and requirements drafting. In industry, the adoption of GenAI is not always tied to a complete “system” from the beginning. It often begins with integrating a model into a development task or piloting an idea with prompts. This adoption view helps teams understand how quality considera-tions arise along GenAI adoption. ISO/IEC 25059 provides quality measurements on AI-enabled software that integrates AI models or services. It struc-tures quality into two complementary categories: • Product quality: characteristics include functional suitabil-ity, performance efficiency, reliability, security, maintain-ability, and portability. For example, ‘functional suit-ability’ concerns whether outputs satisfy specified func-tional requirements for the intended tasks. • Quality-in-use: characteristics cover effectiveness, effi-ciency, satisfaction, freedom from risk, and context cov-erage. For example, ‘freedom from risk’ covers legal, ethical, and security concerns. We distinguish two evaluation aspects in this context. Product quality refers to the quality of the delivered software system that integrates GenAI services, including prompts, orchestration logic, validators, fallback handling, and data handling. Quality-in-use refers to system outcomes and risks in operation, such as task success, user satisfaction, and freedom from risk in the deployment context. When solutions rely on third-party models, teams do not directly inspect the model’s internals (e.g., training data); instead, assess system (AI-integrated) behavior through black-box tests on representative inputs. Teams also complement the evaluation with runtime monitoring of failures and drifts. Transparency in this setting is achieved through traceable configuration and records, such as prompt templates, evaluation datasets, and monitoring logs, together with supplier documentation when available. B. Related Work 1) Quality evaluation methods for GenAI: Several methods. exist for evaluating the quality of AI-based software. Unlike this work, Zhang et al. introduced a structured quality frame-work addressing aspects such as data quality, model robustness, and integration maturity. Riccio et al. developed testing strategies targeting correctness and validation challenges in AI-driven systems. Similarly, Breck et al. proposed the ML Test Score, a checklist-based framework for assessing opera-tional readiness and reducing hidden technical debt. More recently, researchers have begun to explore evaluation techniques tailored to generative tasks such as code generation. Ribeiro et al. proposed CheckList, a behavioral testing model for NLP systems that identifies failures in specific lin-guistic capabilities, such as negation handling and coreference resolution. Wang et al. reviewed evaluation metrics for code generation, including functional correctness, code read-ability, and runtime performance. These approaches advance technical evaluation and can be applied in later development and product evaluation. Some of them offer guidance on system-level prototypes. However, their practical applicability, scala-bility, and suitability for industrial adoption remain uncertain due to limited validation in real-world contexts. Prior work has provided evaluation methods for AI-enabled systems, including readiness checklists and operational gates, testing guidance for AI-enabled software, and qual-ity frameworks that encompass data, model, and integration concerns. These contributions clarify what can be evalu-ated and what evidence can be collected, but they do not de-scribe how such evidence is positioned across adoption phases or how responsibility is assigned in industrial settings. Our study complements them by providing a process-oriented view that links evaluation targets to activities and roles across soft-ware development, using ISO/IEC 25059 as a shared quality vocabulary. 2) Empirical studies of GenAI in software engineering:. Empirical research has examined how GenAI is applied in real-world software development contexts. Barke et al. and Donvir et al. investigated developer interactions with GenAI-assisted tools, reporting both productivity gains and new categories of errors introduced by model-generated code. Other studies have demonstrated GenAI’s applicability in tasks such as test case generation, requirements analysis, and documentation. These studies also raise concerns regarding quality and trust. Donvir et al. observed that GenAI can produce subtle, hard-to-detect defects despite speeding up ideation and coding. Such defects are hard to detect because the generated code compiles and can pass limited tests while masking logic errors and missing checks. Non-determinism and prompt sensitiv-ity for open-ended tasks make failures difficult to reproduce and verify. Aleti et al. highlighted the difficulty of verify-ing correctness and security in GenAI-generated outputs. Park et al. emphasised the importance of prompt engineering and developer awareness for achieving high-quality outputs. Recent studies have further unpacked the drivers of this adoption. Khojah et al. and Russo et al. highlight that adoption is driven less by personal innovativeness and more by workflow compatibility and trust. Lambiase et al. and Choudhuri et al. discuss how cultural values and cognitive styles can hinder adoption. These studies primarily view adoption through the lens of individual developer productivity and trust (AI4SE). However, an organizational challenge remains: once adopted, how teams verify the quality of the resulting GenAI-enabled products (SE4AI). 3) Research gap: The existing literature on GenAI adop-. tion does not provide a usage-oriented view of how quality is evaluated in the industry. While prior works explain why and how individuals adopt GenAI, they neglect how organizations govern the quality of the resulting systems. Our study addresses this gap by documenting the process of GenAI adoption in software products and mapping observed evaluation practices to ISO/IEC 25059. III. RESEARCH METHODOLOGY. We conducted a case study following the guidelines of Rune-son et al.. Archival data and semi-structured interviews were used to collect in-depth insights into how quality is eval-uated throughout GenAI adoption in software products. An overview of the research process is shown in Fig. 1. A. Research Objectives Our study objective is, first, to investigate the technical workflows of individual engineers integrating GenAI into software products; second, to understand how the engineering organization governs and evaluates software quality throughout this adoption. We focus on capturing the exact processes fol-lowed by practitioners. In parallel, we examine how software quality is evaluated throughout the use of GenAI to identify re-curring practices or patterns. For example, the practices include which quality aspects are considered important, when and how evaluations are conducted, and who is responsible for carrying them out. B. Research Questions To reach the research objectives, we formulate the following research questions: • RQ1: What processes do individual engineers follow when adopting GenAI in software products? This question breaks down adoption into phases and prac-tical steps, covering model selection, system integration, and operational use. It provides the baseline for under-standing the surrounding activities where quality concerns may emerge. • RQ2: What quality aspects are evaluated across GenAI adoption and use in software products? This question focuses on the types of quality concerns that practitioners prioritize throughout the adoption and use, such as legal risk, security, or data handling. • RQ3: Who within the engineering organization is respon-sible for evaluating quality throughout GenAI adoption? Given the identified evaluation aspects, we examine which roles participate in quality-related work. It connects iden-tified quality concerns to organizational ownership. • RQ4: How does the organization conduct quality evalua-tion throughout GenAI adoption? We investigate existing approaches to carry out the evalu-ations, which reveal how quality is judged in practice. C. Case Selection We selected two software development companies, denoted as Company A and Company B, using purposive sampling from our industrial network. Both belong to the same corporate group but operate in different domains (FinTech and Telecom). We anonymized the companies under non-disclosure agreements. When this study began, few external partners had established processes for using GenAI in software development. The selected companies offered ongoing AI initiatives and ac-cess to internal artifacts, enabling a focused in-depth case study. We discuss implications for generalizability in Section VI. Table I presents the contexts of the selected companies. Company A produces a financial system (e.g., transactions, loans, and payments) that has successfully served customers for CONTEXT INFORMATION OF THE SELECTED COMPANIES approximately two decades. Company B is in the telecommu-nications domain, providing support services for network management. Both companies operate globally and expand rapidly. Our data collection focused on the Swedish and Indian devel-opment sites. Both have mature software engineering practices and have established organizational goals to integrate GenAI into their products. At the time of the study, a total of 272 AI initiatives were recorded. Most of the recorded initiatives were ongoing, and some had been completed. With many parallel AI initiatives, ad hoc practices lead to non-comparable results, duplicated effort, and limited visibility into risk and drift across teams. Shared adoption patterns and processes, such as common checkpoints (e.g., verification of data privacy), a minimal metric start-set (e.g., metrics on out-put validity/cost), clear role ownership (e.g., product owners), and traceable records (e.g., model versions and prompt tem-plates), enable cross-team monitoring and incident review. This supports consistent decisions and reduces rework in legal and security reviews. D. Research Design Overview To address the research questions, we adopted a two-stage research design that combined an exploratory empirical study with a confirmatory case verification. • Stage 1: Framework Construction (RQ1–RQ4). We conducted an exploratory study at two companies to under-stand current GenAI adoption practices and quality eval-uation needs. We synthesized these findings into a quality evaluation framework. • Stage 2: Framework Verification. We then applied the framework in a specific project (Section V) to verify its feasibility in a real-world setting. This structure ensures that the framework is grounded in empirical observation (Stage 1) and validated through practical application (Stage 2). E. Participant Selection Participants for this study were selected using purposive sam-pling following a multi-step process. Firstly, we identified active GenAI initiatives within each company using internal documentation (see Table I for initiative counts). Secondly, the first author was embedded with development teams in both companies, which provided us with sufficient understanding of the organization. Thirdly, this embedded access gave visibility into AI initiatives (teams, internal documentation, and demo TABLE II PARTICIPANTS SELECTED FROM COMPANY A AND COMPANY B TABLE III LIST OF SELECTED ARCHIVAL DOCUMENTS sessions), which we used to identify participants across roles and responsibilities. We prioritized practitioners with hands-on GenAI responsibilities and direct involvement in quality evaluation activities. During interviews, we collected detailed information about participants’ specific roles, geographic loca-tions, and deeper insights into their GenAI experiences. As a result, shown in Table II, 19 participants were chosen. F. Archival Data In addition to interviews, we collected archival data (e.g., internal documents) to triangulate our findings and gain deeper insights into GenAI adoption and quality evaluation practices. We examined a total of 184 internal wiki/web pages, which are counted as unique URLs, not A4 pages. Pages ranged from short information (1–2 paragraphs) to multi-section guides, and the count indicates breadth rather than physical length. These pages span a wide range of topics related to GenAI adoption, in-cluding GenAI development portals, legal compliance and risk assessment pages, GenAI experimentation guidelines, system development processes, and tooling documentation (e.g., API consoles and quality dashboards). Furthermore, we analyzed 15 internal documents (e.g., PDF, Docx, Excel, slides), as shown in Table III, focused on how GenAI is adopted and evaluated in practice, such as “GenAI risk guardrails” (legal and technical risks), “AI adoption as-sessment” steps, “Responsible GenAI” checklists, model reg-istration and vendor assessment records, security and privacy reviews, metric definitions and dashboard specifications, and training slide decks. We selected archival artifacts using a two-step approach. Firstly, we collected pages directly linked to the GenAI initiative board and its referenced templates. Secondly, we applied a sampling by following links to referenced tech-nical pages, metric definitions, and review artifacts that were mentioned in interviews or embedded in the governance pages. We excluded artifacts that were not specific to GenAI adoption, lacked actionable procedures, or were duplicates. We treated archival artifacts as evidence to triangulate inter-view statements, by confirming the existence of practices (e.g., evaluation activities and dashboards), clarifying role ownership, and extracting concrete evaluation criteria and metrics. Firstly, they provided a high-level view of the GenAI adoption pro-cess, including handoffs and ownership that individual prac-titioners did not always see. Secondly, they enabled a cross-check through comparison with documented procedures and templates. Thirdly, they revealed organizational expectations for GenAI quality that were not yet consistently aligned with teams’ daily work. Although the archival data strengthened credibility, they also have limits. Access was restricted to internal platforms, so coverage reflects what was available during November 2024 to June 2025. As documents varied in detail and maintenance, we mitigated these risks by cross-checking with interviews. G. Interviews We conducted semi-structured interviews to investigate how industrial practitioners adopt GenAI in software products and how they evaluate quality during this process. The inter-views enabled an in-depth exploration of practices, decision-making rationales, and cross-role responsibilities not captured in internal documents, complementing the archival analysis. The interview guide was structured into four sections: Current GenAI adoption status and recent examples of use, Quality evaluation practices across design, development, and operations, Challenges and risks that triggered additional evaluation activities, and Roles and responsibilities, includ-ing who performed quality evaluations and who could approve progress. This design was aimed at aligning with RQ1–RQ4 by transitioning from practice descriptions to quality concerns and role ownership. The full guide is available in the replication package. We conducted four pilot interviews with senior developers and quality managers to calibrate the interview guide. Pilot data were excluded from the main study analysis. The pilot inter-views resulted in two changes. First, we separated questions about software ‘product quality’ from questions about ‘quality-in-use’, because pilot answers conflated static code properties with runtime user outcomes. Second, we adjusted the ordering so that participants first described a recent GenAI use case, and then discussed quality evaluation and role ownership, which improved recall and reduced abstract answers. A total of 19 interviews were held between November 2024 and June 2025. Twelve participants attended in person, and seven participated online, due to the geographical distribution of participants. The interviews were conducted by the first author and held in English, using an interview guide consisting of 21 questions. The interview guide is available on the IEEEDataPort. Each interview lasted between 46 and 60 minutes. We used the same interview guide across interviews. We asked core questions consistently and used follow-up probes to clarify spe-cific examples (e.g., what artifact was produced, who approved it, or what metric was checked). We took several actions to enhance interview validity. Firstly, we covered various roles during the participant selection. To encourage honest feedback during interviews, we informed par-ticipants that the study reports results only in anonymized form and that no raw transcripts were shared with the companies. We removed names of teams, tools, and internal systems during transcription and used participant identifiers (P1–P19). We also requested concrete, recent artifacts, such as checklist items, dashboards, and notes, to ground claims in observable practice rather than general opinions. Secondly, we refined the inter-view guide after pilot interviews based on emerging themes, then used the finalized guide consistently across all remaining interviews. Thirdly, we conducted member checking with par-ticipants through email and follow-up conversations to confirm our interpretations. We have obtained the consent information from all study participants. However, certain limitations exist. As participants came from two companies within a large corporate group with ongoing AI initiatives, the sample may not reflect settings with mini-mal tooling, restricted data access, or without in-house legal and security support. Moreover, self-reporting can introduce recall bias. We used internal documents and web pages men-tioned or shared by participants to confirm statements and elaborate on details from interview transcripts. However, we acknowledge that these sources may reflect similar organi-zational perspectives rather than providing fully independent triangulation. H. Data Synthesis and Analysis We employed thematic analysis to identify patterns and themes in the collected data. We selected thematic analysis because our goal is to derive a view that links technical work (e.g., prompts, metrics, monitoring) with organizational work (e.g., approvals, ownership, governance). This aligns with qual-itative analysis, where artifacts are treated as traces of work and coordination rather than isolated documents. We therefore coded interviews and archival artifacts together, recording both the technical content and its context, including role, phase, and decision point. This analysis followed a four-step approach, supported by iterative collaboration and joint interpretation sessions. 1) Initial coding: A starting set of codes was generated. from interview transcripts and archival artifacts. A mixed coding strategy was used. We applied inductive coding to derive adoption phases, steps, and evaluations from practice descriptions. We applied deductive coding when organizing quality concerns, using ISO/IEC 25059 as a naming and grouping index. For interviews, a meaning unit is a participant statement that describes an action, decision, artifact, metric, or role of responsibility. For archival data, a noted unit is a document fragment that defines a rule, checklist item, gate criterion, metric defi-nition, template field, or role ownership statement. Codes were generated based on their relevance to themes that connect to our research questions. For example, codes reflect GenAI adoption (e.g., “prompt piloting”), evalu-ation (e.g., “risks” and “metrics”), and role responsibil-ities (e.g., “product owner” and “security”). We coded excerpts when they reported concrete work (e.g., actions, EXAMPLES OF CODING PROGRESSION (RAW DATA → CODE → THEME) artifacts, or metrics), rather than general opinions without a specific example. 2) Refinement and disambiguation: The initial codes were. iteratively reviewed to improve granularity and avoid se-mantic overlap. Disagreements were resolved by revisit-ing the raw excerpt, aligning on the intended meaning, and either refining code definitions or splitting merged codes. For example, “correctness” was split into “output accuracy” and “code validity” to capture domain-specific distinctions. 3) Theme construction: We organized codes along two cate-. gories: process area (adoption phase and step) and quality dimension (ISO/IEC 25059 characteristics). We formed themes when repeated evidence appeared in both cate-gories. Edge cases were resolved through team discussion to avoid overlap. 4) Mapping to research questions: Themes were organized. according to their relevance to each research question. For example, codes concerning risk evaluations were con-nected to RQ2, while those related to developer respon-sibilities were connected to RQ3. Table IV illustrates our coding progression with four exam-ples from raw excerpts to initial codes and themes. Each row starts from a quote, either an interview excerpt (tagged as [P#]) or an archival fragment (tagged as Doc-∗). The “Initial code” column shows the first label assigned to that quote. The “Sub-theme” column groups related initial codes that reflect a shared evaluation practice or concern. The “Theme (ISO/Process)” column links the sub-theme to an adoption phase and, when relevant, the ISO/IEC 25059 characteristic used to organize quality concerns. Together, the rows illustrate both inductive coding for process elements and deductive grouping for quality characteristics. The first author performed the initial coding and maintained the codebook. The second and third authors reviewed the evolv-ing codes and themes in repeated discussion rounds using shared excerpts and artifact fragments. A shared coding pro-tocol was used to align definitions, coding rationale, and theme development. Preliminary interpretations were verified through feedback sessions with selected participants to check alignment with their experiences. I. Steps to Derive a Quality Evaluation Framework Through integrating findings from RQ1 (Process), RQ2 (Quality Aspects), and RQ3 (Roles) into a unified view, we synthesized a quality evaluation framework (Section IV.D). The derivation followed three steps: First, we mapped the identified quality aspects (from RQ2) to the corresponding adoption steps (from RQ1) where they were most frequently observed. For example, “legal risk” was mapped to the “Ideation” phase as a gatekeeping evaluation. Second, we assigned the identified responsible roles (from RQ3) to quality evaluation activities based on interview consensus. Third, we categorized the evalu-ations into three types: Gatekeeping (pre-development), Valida-tion (development), and Monitoring (operations), forming the framework’s core logic. This synthesis integrates the “when” (process), “what” (quality), and “who” (role) into a single ac-tionable structure. IV. RESULTS. We present study findings to answer our research questions in this section. Evidence attribution: Interview evidence is tagged with participant identifiers [P1]–[P19]. Archival evi-dence is tagged with document identifiers [Doc-∗ ]. If a claim is supported by both, we include both tags. Table V provides a summary mapping of these main findings to their empirical origins, indicating whether the insight stems primarily from interviews, archival data, triangulation of both, or higher-level interpretive synthesis. A. Results of RQ1—What Processes do Individual Engineers Follow When Adopting GenAI in Software Products? Fig. 2 illustrates the identified phases of GenAI adoption and use in software products. The process is organized into three main phases: Ideation, Development, and Operation. These phases reflect a combination of exploratory use, iterative design, and deployment-oriented practices. 1) Phase 1: Ideation: This phase is initiated by identifying. high-level business opportunities (Step 1). The ideas are typi-cally sourced by product teams, often without strong input on technical feasibility at this stage [P3, P8, P17]. This reflects a trend in which business-driven interests precede technical alignment [P5, P7, Doc-GenAI-Portals]. Once candidate ideas are selected, engineers explore poten-tial GenAI solutions by surveying available models (Step 2). At this stage, engineers screen for early fit and risk rather than formal accuracy. They verify task fit (outputs align with the intended task and domain), data and policy fit (inputs comply with privacy and security rules), the legal acceptability of the TRACEABILITY OF MAIN FINDINGS TO DATA SOURCES license, basic output quality on a few representative prompts, and integration feasibility (including API access, hosting, and deployment path) [P5, P13, Doc-ModelSelection]. Piloting sample prompts (Step 3) serves as a lightweight fea-sibility check. It enables teams to evaluate whether GenAI out-put aligns with expectations before making further investments. In the studied companies, this step was often used to discard ideas quickly, reflecting a low-cost experimentation practice [P12, Doc-ExpGuide]. If the output appears promising, teams proceed to select a specific AI model (Step 4). Selection is commonly based on empirical prompt testing, vendor trust, and ease of integration, rather than formal quality guarantees [P5, P19, Doc-ExpGuide]. This also implies that model selection is frequently revisited as the project progresses. 2) Phase 2: Development: In the development phase, the. focus shifts from ideation to system integration. Teams re-fine business requirements (Step 5) to accommodate GenAI capabilities and constraints, often reducing the granularity of traditional specifications [P15, P18]. In contrast to conventional software, requirements are shaped around acceptable variation in model output, making this step inherently iterative. The development step (Step 6) involves incorporating the se-lected GenAI model into an application or service. This may in-clude prompt engineering, workflow orchestration, or API-level integration. Participants noted that development often advances without complete clarity on evaluation criteria, reinforcing the need for later adjustments [P4, P9]. Testing and dataset preparation (Step 7) vary considerably in formality. Some teams curate specific test cases to val-idate expected behaviors, while others rely on production-like data to simulate real-world input [Doc-TestPlan]. The level of rigor here depends on the perceived risk and target user. Output evaluation (Step 8) is where the prototype is re-viewed for its usefulness, correctness, and consistency. Unlike deterministic software, evaluation involves human-in-the-loop judgment and often lacks predefined acceptance thresholds [P1, P5, P10]. This introduces challenges in comparing iterations or tracking progress over time. 3) Phase 3: Operation: After deployment (Step 9), sys-. tems enter the operations phase, where GenAI becomes part of production environments. This transition often marks a shift in responsibility from development teams to maintenance and operations teams [P6, P8, P14, Doc-DevOps]. Runtime monitoring (Step 10) is implemented to track anomalies, drift, or failure cases. However, monitoring practices remain immature; few teams employ systematic mechanisms for runtime evaluation beyond basic usage statistics [P2, P16, P18]. Evaluation of the runtime system (Step 11) focuses on sta-bility, user satisfaction, and reliability in production contexts. This step is typically reactive—issues are addressed after sur-facing, rather than through predictive assessments [P15, P17, Doc-DevOps]. Improvement cycles (Step 12) are driven by operational feed-back and usage data. These iterations may involve switching models, re-engineering prompts, or redesigning workflows [P1, P7, P16, Doc-ModelSelection]. However, the improvement pro-cess is rarely governed by formal change management, reflect-ing the still-exploratory nature of GenAI operations. 4) Observations: We observed that GenAI adoption in indus-. try follows a typical pattern of iterative experimentation, model probing, and incremental discovery, similar to how organizations adopt other new technologies [P2, P12, P19]. This reflects a shift from deterministic system design toward empirical con-vergence, where systems evolve through usage feedback and feasibility trials. Moreover, we learned that quality evaluation occurs in most phases, but practices are often informal and subjective [P4, P8, P18] (e.g., accepting outputs without predefined thresholds for accuracy, latency, or safety). Practices such as prompt piloting, runtime observation, and prototype judgment serve as implicit gates for decision-making. This raises concerns about traceabil-ity and accountability in GenAI-based systems. For practitioners, the identified three phases can serve as a reference to position their own efforts. It also offers researchers an example of how GenAI adoption unfolded in two industrial settings. B. Results of RQ2—What Quality Aspects are Evaluated Across GenAI Adoption and Use in Software Products? We report the quality concerns that emerged from our em-pirical data. Fig. 3 summarizes the quality aspects identified by participants across the three phases of GenAI adoption and use. At the end of each phase, we present “Proposed Actions” that serve as the foundation for the evaluation framework proposed in this paper. Quality aspects in Phases 1 and 2 relate to Product Quality, and Phase 3 focuses on operation, which falls under the Quality-in-Use characteristics as defined in ISO/IEC 25059. These as-pects reflect where practitioners currently focus their attention, shaped by their roles and organizational constraints [P3, P7, P13, P16, Doc-GenAI-Portals]. 1) Model Quality Aspects (Phase 1) — ISO 25059 Product. Quality: a) Security and legal compliance: Participants empha-sized the importance of understanding legal implications early, particularly regarding intellectual property, data privacy, and third-party model licensing [P9, P10, P11, P12, P13, P14]. This concern is not only about selecting ‘safe’ models but also about whether the use case is legally viable [P10, Doc-RiskGuardrails]. The motivation stems from the uncertainty around AI model behavior and ownership of training data, which introduces legal risks. Practitioners noted that without early legal validation, entire business ideas may later be blocked or delayed [P9, P11, P12, P14, Doc-ExpGuide]. b) Business risks: Feasibility was often framed in busi-ness rather than technical terms, for example, unclear value propositions, unpredictable behaviour in user-facing contexts, or prohibitive costs [Doc-ExpGuide, Doc-GenAI-Portals, Doc-DevOps]. Participants noted that even technically functional GenAI models might conflict with user trust or supportability goals [P2, P3, P5, P7, P16]. As a result, early-phase quality eval-uation often prioritizes business viability alongside compliance. c) Proposed Actions: At this phase, quality evaluation fo-cuses on compliance and purpose. Organizations are advised to involve legal, compliance, and product strategy teams early to assess feasibility before prototyping. 2) Development Quality Aspects (Phase 2) — ISO 25059. Product Quality: a) Data quality: Data quality emerged as a central concern. Participants were uncertain whether test datasets and prompts accurately reflected real-world adoption, especially when out-puts were highly sensitive to subtle input changes. In some teams, overfitting prompts to hand-crafted cases reduced gen-eralization after deployment [P1, P2, P4, P15, Doc-ExpGuide, Doc-TestPlan]. b) Efficiency and Effectiveness: Effectiveness was inter-preted contextually: some defined it as “acceptable output for task completion”, others included time-to-answer or resource efficiency (e.g., latency, token cost). Formal benchmarks were rarely used, according to participants. Instead, teams used quick iterations with subjective judgment [P2, P3, P5, P6, P7, P18, Doc-ExpGuide, Doc-GenAI-Portals]. While agile, this infor-mality makes scaling and repeatability difficult. c) Proposed Actions: Informal assessments should be complemented with measurable criteria, for example, output consistency, processing time, and task success rates. Data quality should be revisited during development, not assumed from earlier checks. 3) Operation quality aspects (Phase 3) — ISO 25059. Quality-in-Use: a) System install and upgrade: Participants reported de-ployment friction for GenAI components due to version mis-matches, hosting constraints, and unclear dependencies [P1, P3, P4, P5, P7, P18, P19]. Such issues are more disruptive for GenAI than traditional systems, given their dynamic depen-dencies and licensing restrictions [Doc-RiskGuardrails, Doc-TestPlan, Doc-DevOps]. b) System quality trends: Monitoring quality trends in pro-duction was valued but underdeveloped [P3, P5, P15, P17, Doc-DevOps]. Most participants lacked formal systems in place to track model drift, unexpected failures, or changing behavior across environments. Instead, most relied on user complaints or ticket reports from support teams. This limits feedback loops and hinders root-cause analysis [P6, P7, P8, P17, Doc-DevOps]. c) Proposed Actions: Practitioners should treat GenAI-based components as evolving artifacts. Continuous monitor-ing, logging, and sampling should be used to detect unexpected changes, supported by feedback channels from users and oper-ations teams. 4) Observations: While the identified aspects align with ISO/IEC 25059, several characteristics remain under-addressed: • Functional adaptability — system responsiveness to dy-namic environments or evolving data. • Robustness — stable performance under bias, adversarial inputs, or novel conditions. • Transparency — clear documentation of AI models’ logic, data provenance, and decision rationale. • Societal and ethical risk mitigation — adherence to fairness, privacy, and accountability principles. During the interviews, these quality characteristics were rarely mentioned unless specifically asked about. For instance, functional adaptability – an important reason for choosing GenAI in the first place – was not systematically evaluated. Ro-bustness and transparency were typically implicit or informal (e.g., “seems to work well”) rather than captured through struc-tured criteria or logging. Without explicit evaluation, systems may degrade in unobserved ways, particularly after deployment [P15, Doc-DevOps, Doc-TestPlan]. Engineering teams should integrate selected ISO/IEC 25059 attributes into lightweight quality evaluation checklists or design review templates. The following aspects are actionable: • Transparency: Document prompt structures, data origins, and model versioning. Use structured logging to maintain prompt output behavior over time. • Robustness: Evaluate and guard against structured-output (e.g., JSON/XML) failures. Use schema validators (e.g., JSON Schema, XSD), constrained decoding or typed func-tion calling, and strict parsers. Add automatic repair, en-force timeouts, and use retries with backoff and circuit breakers. Define fallbacks for known failure cases (e.g., cached answers, rule-based generation, or human-in-the-loop). • Societal and ethical risk mitigation: Align GenAI adoption with Responsible AI frameworks. Include non-engineering stakeholders early, such as legal, security, compliance officers, domain experts, and representative users. These steps do not require full ISO compliance but can raise quality awareness and connect daily GenAI practice with standardized guidance. C. Results of RQ3—Who Within the Engineering Organization is Responsible for Evaluating Quality Throughout GenAI Adoption? GenAI integration into software products creates new quality challenges, with responsibility for addressing these challenges remaining fragmented across roles. Fig. 4 summarizes the roles identified in each phase of GenAI adoption—ranging from busi-ness ideation to runtime operations—and the specific quality concerns associated with them. 1) Responsible roles: In the Ideation phase, responsibil-. ity is distributed among business owners, legal and licensing teams, and security units. Business owners assess the viabil-ity of use cases and business risks, while security and legal experts validate compliance with regulations and intellectual property policies when selecting models [P5, P18, P19, Doc-ModelSelection]. As a participant stated: “We had to get the legal team involved early because we weren’t sure if the use case was even allowed with the data we had [P9].” In the Development phase, developers and architects lead integration, while quality engineers and analysts support testing and validation. Yet evaluation practices often remain infor-mal or rely on domain intuition [P2, P6, P7, P8, Doc-GenAI-Portals]. Several participants reported: “It is mostly trial and test prompts, tweak outputs, and we go with what feels right [P16].” Some teams assign evaluation duties to developers by default, and in the absence of structured criteria, they rely on domain expertise and human judgment. Several participants [P4, P5, P18] noted: “There is no fixed checklist. We just look at whether the response makes sense or is dangerous.” In the Operation phase, roles such as infrastructure engi-neers and support analysts become involved. However, runtime feedback mechanisms are often weakly connected to earlier evaluation decisions [P15, P16, P17, Doc-DevOps]. Partici-pants [P7, P15, P18] stated that “We rely on support tickets to learn when something’s off. There’s no automated alerting tied to model output yet.” 2) The need for a dedicated role in ownership: Across all. phases, we found limited unified ownership for GenAI quality evaluation. Although responsibilities are scattered among do-main experts, the lack of coordination and traceability reduces the accountability and repeatability of evaluation activities. For example, a participant stated: “Everyone is CERTAIN that SOMEONE is checking the quality, but no one knows who actually does it [P1].” This ambiguity is not unique to GenAI. Conventional projects may also suffer from diffuse quality own-ership in siloed organizations [P7, P8, Doc-TestPlan]. GenAI magnifies the problem because quality decisions span multi-ple areas, including legal, security, compliance, and software development. A recurring theme in interviews was the absence of a desig-nated role that integrates technical, legal, and ethical concerns across the GenAI adoption and use [P2, P9, P11, P14, Doc-RiskGuardrails]. As participants [P4, P5, P6, P10, P19] repeated that “It would help to have someone who knows what to eval-uate, not just technically, but in terms of risk or impact too.” Based on this finding, we outline a recommended GenAI Qual-ity Lead role in the Discussions section, alongside alternative organizational solutions. Responsibilities of the quality lead include: • Facilitate early-phase risk identification (e.g., legal, secu-rity, bias). • Establish and adapt quality criteria with legal, security, and domain stakeholders. • Defining evaluation criteria with product and engineering teams. • Drive evaluation practices across development and opera-tions (e.g., output testing, runtime monitoring). • Translating quality standards (e.g., ISO/IEC 25059) into project-specific practices. • Serve as a contact point for review boards, audits, and compliance alignment. This role builds upon principles of SE4AI, which emphasizes quality assurance throughout the AI engineering lifecycle. By formalizing this role, companies can better integrate emerging standards (e.g., ISO/IEC 25059) and internal Responsible AI guidelines into their everyday practices. Importantly, the role is not a replacement for domain-specific experts, but rather a coordinating role that enhances quality, accountability, and traceability. It helps move beyond ad hoc evaluations by em-bedding structured thinking into the day-to-day engineering of GenAI-based solutions [P2, P5, P18, P19, Doc-TestPlan, Doc-ExpGuide]. D. Results of RQ4—How Does the Organization Conduct Quality Evaluation Throughout GenAI Adoption? Based on our findings from RQ1 to RQ3, we synthesized a process-oriented framework that integrates GenAI adoption phases with quality evaluation practices and role responsi-bilities. This framework illustrates HOW quality is evalu-ated during GenAI adoption in phases and indicates WHO is accountable. Fig. 5 consolidates our findings into a process-oriented view of GenAI adoption, with quality evaluations integrated into each phase and corresponding roles. Each phase in the figure contains steps (white boxes), among which specific actions related to quality evaluation are high-lighted. The ⋆ marks indicate points where evaluation is di-rectly enabled, for example, assessing legal risks, verifying outputs, or analyzing system qualities. The ♦ marks identify supporting evaluation activities, for example, model prompt-ing or installability that shape or constrain quality assurance decisions. Together, they form a network of quality signals – some proactive, others reactive. Below each phase, the gray panels describe evaluation prac-tices, which range from compliance vetting to metric-based evaluations. At the bottom, the roles responsible are connected, revealing a distributed accountability structure where legal, en-gineering, and quality teams are all involved at different stages. 1) Quality evaluation practices in GenAI adoption: Based on. collected data, three interconnected evaluations emerge: Gate-keeping – evaluation before development, Validation – test-ing during development, and Monitoring – evaluation during operation. a) Gatekeeping (before development): Evaluations in the Ideation phase are often conducted to determine whether a GenAI model or idea can be pursued [P6, P7, P18, Doc-ExpGuide, Doc-RiskGuardrails]. These gatekeeping controls are typically compliance-driven, involving legal and security teams. The checklist in Table VI was referenced by multiple partici-pants as a tool to assess risk exposure and vendor accountability before prototyping. Participants from legal teams [P9, P10, P11] mentioned: “We block use cases if the supplier can’t commit to data privacy or copyright terms.” These evaluations serve as safeguards, especially in large organizations where regulatory pressure is high. b) Validation (during development): In the development phase, evaluation activities shift from permission to execution. Here, teams examine if an AI model behaves appropriately in their specific settings. This includes assessing data quality, response consistency, efficiency, and effectiveness in relation to internal policies. We found that evaluation practices often lack mature tooling and rely on custom metrics [P1, P6, P7, P8, Doc-TestPlan, Doc-GenAI-Portals]. For example, a senior developer shared that “We often create our specific key performance indicators in our implementation, aiming to give a signal on hallucinations [P3].” Table VII outlines a cross-functional practice that was used in multiple settings to structure quality assessments during im-plementation. These assessments are rarely isolated. They are tightly coupled to development iterations. To better understand the relationship between the observed practices and software quality, we mapped the evaluation activ-ities to ISO/IEC 25059 characteristics in Fig. 6. As shown in Fig. 6, observed practices (e.g., prompt testing, the use of tools, and metrics) map implicitly to ISO quality characteristics. Specifically, they reflect: • Functional Suitability. It includes the correctness and appropriateness of output. • Efficiency. It covers measures for inference time duration and resource usage. • Effectiveness. It covers task success and utility in the target context, assessed through targeted tests and usage evaluation. • Reliability. It refers to output repeatability and failure handling. These mappings indicate what to evaluate; they do not pre-scribe how. In our cases, metrics were selected per context, and some remained unsettled. Future tooling and policy definitions may benefit from translating these ISO sub-characteristics into practical evaluations and observable metrics. c) Monitoring (during operation): Evaluation does not end at deployment. Once a system is deployed successfully, teams perform runtime assessments based on logs, live quality trends, and user feedback [P15, P17, Doc-DevOps]. This mode fo-cuses on identifying performance degradation, systemic errors, or violations of usage policies. As shared by an operations engineer, “It is required that we detect issues not when they happen, but when someone flags them downstream [P16].” These practices are increasingly important in GenAI, where be-havior can drift over time. Importantly, monitoring efforts feed back into development and legal reassessments, completing the lifecycle. TABLE VI LEGAL RISK ASSESSMENT CHECKLIST AND OWNERS TABLE VII QUALITY EVALUATIONS FOR GENAI APPLICATION IMPLEMENTATION The integrated view in Fig. 5 reveals a decentralized ar-chitecture of quality assurance in GenAI adoption. Evaluation is not owned by a single role or function. It is distributed, contextual, and embedded in day-to-day decisions. Tables VI and VII reflect how companies attempt to formalize what would otherwise remain tacit. However, several challenges remain: metrics are inconsistent, practices vary across teams, and accountability is shared but unclear. Some participants described this as a lack of a coherent QA strategy tailored to GenAI, “We have a security master in each dev team. But who is the quality champion for GenAI? [P7]” This observation motivates the discussion of a new role in Section VII-B, and underscores the practical value of treating quality evaluation not as an afterthought, but as an ongoing design activity. 2) Model reuse: The quality evaluation framework presented. in Fig. 5 offers a reusable structure for organizations seeking to evaluate GenAI-enabled software. By clarifying when and how evaluations are conducted, and by whom, the overview can serve as a reference for orga-nizations lacking established quality practices for GenAI. It also helps bridge informal routines and ISO-inspired evaluation dimensions, as illustrated in Fig. 6. Therefore, the framework is a synthesized process reflecting actual engineering practice. To further examine its practical utility, we conducted a case implementation in one of the studied companies. The aim was to apply the evaluation structure to a real GenAI adoption and assess its feasibility. The next section presents this framework verification case. V. VERIFICATION OF THE FRAMEWORK. The goal of this verification is to demonstrate how the framework can be instantiated in a real-world project by trans-lating the abstract quality characteristics into project-specific evaluations, metrics, and role assignments. We focus on fea-sibility and traceability by verifying whether teams can per-form the necessary evaluations using available artifacts (test suites, validators, dashboards, and logs), and whether the re-sulting evidence supports the adoption of GenAI and post-deployment monitoring. The verification case was conducted in Company B. A. Case Description The case was a GenAI-assisted application for generating XML code used in Unstructured Supplementary Service Data (USSD) menus. USSD is a widely used communication pro-tocol in mobile services that enables interaction through short codes, eliminating the need for internet access. The target users of this tool were merchants who needed to create dynamic menus for product browsing and purchasing without manually writing XML code. The tool enabled these users to provide nat-ural language descriptions of their menu logic, which the system converted into valid, executable XML files for integration into mobile financial platforms. OPERATIONALIZATION OF QUALITY CRITERIA FOR USSD CASE This project was selected as a verification case for two rea-sons. First, it represents a realistic and commercially relevant adoption of GenAI, involving code generation in a critical domain. Second, it required cross-role collaboration between developers, quality assurance engineers, legal reviewers, and product owners, reflecting the multi-stakeholder structure of GenAI adoption. B. Operationalization of the Framework The development process was structured around the phases and evaluation steps outlined in the quality evaluation frame-work. To apply the framework, the team first translated the abstract ISO/IEC 25059 quality characteristics into concrete project metrics (Table VIII), such as semantic alignment score, XML validation pass rate, average generation latency, and com-pliance sign-off. This step was critical to move from “quality” as a vague concept to measurable acceptance criteria. With the project-specific quality criteria definitions in Ta-ble VIII, the framework verification process was executed through three phases, as shown in Fig. 7. 1) Phase 1: Ideation (Before Implementation): The Ideation. Checkpoint was conducted by the Product Owner and Legal Team. They evaluated Security Risk by reviewing the data flow. Since the tool would process merchant inputs but no end-user personal data, the risk was classified as low, provided no training occurred on customer data. Model selection (GPT-4o vs Llama 3) was evaluated based on the Functional Suitability of initial prototypes. 2) Phase 2: Development (During Implementation): During. development, the Development Checkpoint focused on Relia-bility and Performance Efficiency. Developers integrated a val-idator that automatically checked every generated XML against the schema. The QA team ran a regression suite of 50 complex menu descriptions. Metrics were tracked daily: initial pass rates were low (60%) but improved to 92% after prompt engineering refined the XML structure instructions. 3) Phase 3: Operation (After Implementation): Post-deployment, the Operation Checkpoint involved continuous monitoring via Grafana. The Operations team tracked Performance Efficiency (latency) and user feedback (implied Func-tional Suitability). An alert threshold was set for latency spikes (> 45s), which would trigger a fallback to a smaller, faster model. C. Verification Outcomes The proposed framework was applied to all three phases of the USSD XML Generator case. It enabled the team to address quality concerns with defined responsibilities and metrics. In the early phase, the model assessment helped filter out unsuitable options and anticipate risks. Legal and security engi-neers used the checklist to validate compliance concerns before deployment. One engineer noted, “The checklist turned vague worries into actionable points, such as licensing, datasets, in-puts, and outputs.” During development, quality metrics were embedded into the implementation process. Accuracy and la-tency were measured continuously using in-house scripts and Prometheus monitoring. A quality assurance engineer re-marked, “The quality characteristics gave us criteria early, so we knew what to test and why.” Operational monitoring added a feedback loop post-deployment. Monitoring dash-boards showed runtime trends that directly informed model prompt adjustments. “Seeing the correctness rate of generated outputs improve week by week gave us real confidence,” said an operations engineer. Despite its practical value, the framework had limitations during the case implementation. First, the success of its appli-cation relied on the willingness and capacity of team members to take ownership of additional evaluation tasks. Second, al-though the framework included evaluation practices with ISO alignment, the evaluation results still depended on the quality of test data and prompt engineering, where guidance was limited. Third, the monitoring approach, while useful, required specific tooling and metrics setup that may not generalize to other organizations. We position the framework as a meta-level structure with a stable core (checkpoints, role duties, and a minimal metric set) and configurable parts (attributes, datasets, and safeguards). Teams can instantiate it per feature or domain using shared tem-plates and steps, rather than building a new method or bringing new tools each time. In the USSD case, the core metrics were reused; prompts and test data were tailored. VI. VALIDITY THREATS. We follow Runeson et al.’s guideline to address validity threats: construct, internal, external, and conclusion validity. a) Construct validity: To identify quality evaluation practices in GenAI adoption, we conducted 19 interviews with practition-ers who had hands-on experience with GenAI-enabled software products. Participants were selected from multiple roles, includ-ing development, testing, architecture, legal, and management, to ensure diverse perspectives. Although each participant had experience with parts of the entire process, their combined accounts provided a holistic view across its constituent steps. However, participants may have used inconsistent terminol-ogy or lacked shared definitions of “quality,” “evaluation,” or “GenAI adoption.” To mitigate this, we clarified questions dur-ing interviews (e.g., “What did your team do to ensure the model was suitable?”), followed by clarifying questions and validation of summaries during transcription checks. b) Internal validity: Our thematic analysis followed a staged coding process, including initial open coding, coding for pattern identification, and iterative refinement. Multiple researchers re-viewed the coding structure and figure derivation to minimize personal bias. However, participant statements may still include misinterpretations. We mitigated this through three group ses-sions to review collected artifacts (e.g., internal documents and metrics). For the verification case, the same research team col-laborated with practitioners during implementation, which may introduce bias toward confirming our proposed framework. To mitigate this, study participants interpreted evaluation metrics and improvement steps. c) External validity: Because participants from two se-lected companies work within the same corporate group, their accounts may reflect shared governance norms and vocab-ulary, so complete independence of perspectives cannot be guaranteed. We mitigated this by sampling across roles, sites, and functions and by triangulating interview accounts with archival artifacts, but we still interpret the findings as re-flecting this organizational setting and discuss transfer limits accordingly. The framework was derived in a setting with development teams, in-house legal and security support, and cross-site col-laboration. Organizations that lack these prerequisites, such as smaller teams without dedicated risk or operations functions, may need to adapt roles, activities, and metrics before adoption. Therefore, we present the framework as process-oriented guid-ance rather than a universal prescription, and we distinguish elements from configurable ones to aid tailoring. The verifi-cation case demonstrates feasibility in the studied context, but applying the framework elsewhere will require adaptations to local constraints. d) Conclusion validity: We documented research procedures, coding steps, and analytical decisions to support transparency. The first three researchers analyzed the data and resolved dif-ferences to ensure reliability. Participants’ feedback was used to validate our interpretations of key themes. We acknowledge potential threats to conclusion validity. For example, if the data concerning legal evaluation practices were incomplete or misinterpreted, an alternative conclusion might be that legal responsibilities are decentralized rather than co-ordinated through a central unit. However, our collected data showed that legal evaluation was explicitly addressed by multi-ple participants. Furthermore, our multi-step verification, which includes peer checking, participant feedback, and triangulation, reduces the risk of such misinterpretations. Additionally, we reported both positive outcomes and chal-lenges during the verification of the proposed framework, which aims to mitigate confirmation bias. Future studies are needed to understand the framework’s long-term effects in other industrial settings. VII. DISCUSSIONS. A. Legal Risk and Security Compliance Come First? In our cases, preliminary legal and security checkpoints blocked or delayed pilots. The reasons included supplier terms that allowed training on user inputs, unclear owner-ship of generated code, and data-residency constraints that conflicted with company policy (Table VI). While many en-gineering activities are iterative and exploratory, legal and compliance violations can trigger irreversible consequences, such as breach of contract, regulatory sanctions, or reputa-tional damage. This risk profile justifies the placement of le-gal and security evaluation as the first quality gates in our framework. The necessity of prioritizing these concerns became evident in our interviews. As one participant noted: “We did not even get to try the model, legal shut it down because of the terms of service.” Unlike performance issues or usability flaws, legal violations typically cannot be patched silently once a product reaches customers. This means that the cost of deferred evalua-tion is not technical debt, but rather organizational risk. Treating these issues as upstream concerns helps ensure that software built on GenAI is legally viable and security-resilient before other quality aspects are even meaningful. Additionally, the placement of these assessments early in the workflow promotes shared accountability across teams. Rather than delegating compliance evaluation to external gatekeepers, integrating them into quality evaluation encourages engineers to consider these aspects as part of their technical responsibility. This shift from reactive compliance to proactive governance requires additional resources to address risk and compliance evaluation. B. Introducing a GenAI Quality Lead Role Our findings show that quality evaluation responsibilities in GenAI-enabled software products are distributed across several functional roles, including legal teams, developers, security engineers, analysts, and quality managers. While each role addresses a specific concern—such as compliance, technical performance, or operational monitoring—no desig-nated function oversees quality holistically across the entire lifecycle. This absence of centralized responsibility leads to fragmen-tation in evaluation practices and limited traceability of deci-sions. Practices such as legal risk assessment or runtime evalu-ation are conducted in isolation, without a unifying perspective that ensures GenAI-specific quality concerns are systematically addressed. To mitigate this gap, we propose the establishment of a dedicated role: the GenAI Quality Lead (GQL). The GQL is envisioned as a cross-functional role that ensures quality considerations are proactively embedded throughout the GenAI adoption lifecycle—from model selection and data preparation to deployment and continuous operation. Specifically, the GenAI Quality Lead would coordinate the following responsibilities: • Collaborate with legal and security teams during model sourcing to address compliance and licensing risks. • Work with developers and analysts to define evaluation criteria aligned with business and technical requirements. • Ensure structured assessment of output quality and run-time behavior, drawing on relevant metrics and feedback loops. • Facilitate alignment with external standards, such as ISO/IEC 25059, by translating quality characteristics into actionable practices. This role draws conceptual support from the SE4AI (Soft-ware Engineering for AI) perspective, which emphasizes the need for dedicated quality assurance practices tailored to AI-based systems. The GenAI Quality Lead does not replace ex-isting domain experts, but rather functions as a coordinating actor that integrates diverse quality concerns into a consistent and traceable process. Establishing such a role may enhance an organization’s capa-bility to evaluate GenAI applications using measurable metrics. By providing transparent quality metrics, this role addresses the “trust” and “risk perception” barriers identified in recent adoption studies. This role also helps bridge the gap between “Product Qual-ity” (e.g., accuracy) and “Quality-in-Use” (e.g., user satisfac-tion). Organizations often focus on static product metrics, yet adoption failure often stems from poor quality-in-use. The GQL ensures that operational feedback loops explicitly mea-sure these dynamic aspects, thereby preventing the trust gap that arises when technically suitable models fail in practice. While proposing a GenAI Quality Lead explicitly addresses the coordination gaps we observed, organizations may explore alternative solutions depending on their structure. For instance, rather than creating a dedicated new role, companies could establish cross-functional AI Review Boards to periodically evaluate GenAI projects at critical phase gates. Alternatively, organizations might integrate GenAI-specific evaluation criteria into the existing duties of traditional Quality Managers, aligning GenAI quality assurance with established Quality Management Systems (QMS) and ISO standards. The essential requirement is that coordination of GenAI quality becomes a deliberate, cen-tralized responsibility, regardless of the specific organizational form. C. Practical Implications Organizations can use the framework to transition from ad-hoc prompting to a structured quality assurance process. We recommend a three-step instantiation approach: 1) Map existing gates: Identify where current quality eval-. uations (e.g., legal/security) occur and overlay the frame-work’s processes. For GenAI adoption, the ideation gate is critical to filter high-risk use cases early. 2) Define. the GQL responsibilities: Even without a dedicated new hire, assign the “GenAI Quality Lead” duties to a senior developer or architect. This en-sures clear ownership of model selection and prompt versioning. 3) Operationalize metrics: Replace vague goals like “be. helpful” with measurable proxies, such as “user edit rate” (for code generation) or “latency < 15s” (for chat inter-faces), as demonstrated in our verification case. This structured approach reduces accumulated evaluation is-sues when teams experiment without defined success criteria. Researchers outline that trust, and workflow com-patibility, are primary drivers of GenAI adoption. Our findings extend this by showing that organizational trust is fragile without structured evaluation. The framework provides the missing “verification layer” that transforms individual de-veloper trust (AI4SE) into organizational confidence (SE4AI). D. Why ISO Quality Characteristics Matter Across the organizations we studied, quality was interpreted through different aspects: efficiency by developers, explainabil-ity by managers, and compliance by legal teams. This fragmen-tation created challenges in aligning expectations, coordinating evaluations, and making trade-offs explicit. What was missing was a common foundation to reason about quality without flattening its complexity. In our study context, the companies did not adopt ISO/IEC 25059 as a formal compliance standard. They still performed quality work through internal policies and routines, such as security and privacy reviews, supplier and licensing evaluations, QA testing, and operational monitoring. However, these activi-ties were not recorded as traceable evidence. We therefore use ISO/IEC 25059 as a shared quality vocabulary to name and group observed quality concerns and to highlight which quality-in-use aspects are seldom evaluated in practice. In the framework verification case, ISO characteristics helped participants translate vague concerns into actionable targets. For instance, instead of “the model should be fast and safe,” teams articulated expectations as “inference time < 45 seconds (efficiency)” and “no confidential leakage from enterprise infor-mation (security).” This shift enabled traceable decisions and role-specific accountability throughout the lifecycle. Moreover, ISO’s separation of characteristics and sub-characteristics (e.g., from reliability to robustness) allowed practitioners to tailor depth without discarding structure. This flexibility appears important in real-world development, where not all quality attributes are equally relevant. The standard helps teams organize what they have already done, while exposing what might be missed in terms of software quality. We acknowledge that ISO 25059 is still under refinement and does not address all GenAI-specific characteristics. However, our evidence suggests that it provides a shared model grounded in both academic rigor and industrial recognition. a) Relation to the EU AI Act: While ISO/IEC quality stan-dards help structure technical quality, they do not by themselves address regulatory duties under the EU AI Act (e.g., risk clas-sification, documentation, conformity assessment, post-market monitoring). When this study started, the operational guidance for the Act was still evolving in the studied companies, and the usage of software development was considered unclear. There-fore, teams mapped existing ISO-based quality characteristics to internal policies for GenAI adoption and usage, conducted legal and privacy reviews, and prepared for later alignment with evolving standards and regulatory requirements, not limited to ISO/IEC 25059. Future studies should investigate how GenAI quality evolves over time and how teams update evaluation criteria accordingly. Moreover, the GenAI Quality Lead role needs further explo-ration, including required competencies and interactions with existing quality assurance activities. VIII. CONCLUSION. The use of GenAI technologies in industrial software prod-ucts introduces new challenges in software quality assurance. Through a case study in two industrial organizations, we inves-tigated the process of GenAI adoption in practice, what quality aspects matter, who is responsible for evaluation, and how these activities are carried out across the lifecycle of GenAI-enabled software. In the study, we contribute a framework: a process-oriented and role-aware view of GenAI quality evaluation. The frame-work is built based on observed practices and practitioners’ feedback. This framework clarifies when and how quality should be assessed between legal, technical, and operational responsibilities. By visualizing where evaluation takes place and what char-acteristics are considered, we help practitioners gain a deeper understanding of GenAI-related quality. The introduction of the GenAI Quality Lead role addresses a missing coordination function, supporting traceability, consistency, and compliance in multi-role environments. The real-world implementation fur-ther confirms the framework’s applicability and exposes oppor-tunities for refinement. Looking forward, GenAI adoption can accelerate across sec-tors and development contexts. It raises the necessity for quality evaluation frameworks that are not only theoretically grounded but also implementable in practice. Our study provides a frame-work for practitioners to follow and offers a lens for rethinking software quality in the era of GenAI. DATA AVAILABILITY STATEMENT The datasets supporting the conclusions of this article are available in the IEEEDataPoint. ACKNOWLEDGMENT. We acknowledge support from the KKS Foundation through the S.E.R.T. Research Profile Project at Blekinge Institute of Technology.