From Obstacles to Strategies: AI in the Software Life Cycle

What is Paladino’s specialist thesis actually about?

Paladino’s thesis is a Spanish-language specialist review of how artificial intelligence is being placed into the software development life cycle, and why that placement still stalls. Marcela Paladino submitted the work in April 2025 for the degree of Especialista en Ingeniería de Software at the National University of La Plata, with Claudia Pons as director. This article is an English reading of that review, not a substitute for the original Spanish text.

The full title is Desde los obstáculos hasta las estrategias: Integración de la inteligencia artificial en el ciclo de vida del desarrollo del software. You can read the research article on Academia.edu, and a deposit is also held on HAL. The institutional copy lives in SEDICI, the university repository, as a trabajo de especialización rather than a journal paper. That status matters. It is a structured narrative review with inclusion rules, a data-extraction matrix, and a quality appraisal of about thirty studies, written to answer one question: what are the main challenges and barriers to integrating AI into the SDLC.

"The tools exist for almost every phase. The stall is not a missing chatbot. It is data, old systems, frightened teams, and models nobody can explain."

Deep learning, natural language processing, and intelligent agents can help planning, design, code generation, testing, and maintenance. Paladino’s finding is that those opportunities sit next to stubborn limits: data quality and structure, compatibility with inherited systems, organizational resistance, and a lack of transparency. The recommended way through is not another vendor list. It is technical cleanup, serious training, and ethics frameworks that people will actually use.

How did she review the literature?

She reviewed the literature as a narrative synthesis with explicit inclusion rules, not as an informal reading list. Studies had to be from 2022 to 2025, in English or Spanish, and they had to address AI inside SDLC phases rather than AI methods in isolation. The search mixed Google Scholar with IEEE Xplore, Scopus, ACM, and Web of Science, then used a quality matrix and a comparative extraction sheet.

That method is a strength and a limit. A narrative review can connect technical, organizational, and ethical threads that a narrow systematic review might split. Paladino herself notes the bias risk: the author chooses what to include, and some relevant work can be missed. She coded findings into technical, organizational, and ethical categories, then mapped tools and obstacles onto planning, requirements, analysis, design, development, testing, deployment, and maintenance. If you only want a catalog of products, you will be disappointed. If you want a picture of why adoption is uneven, the design fits.

The research question is blunt: what are the main challenges and barriers in integrating AI into the SDLC. Specific aims include finding the stages with the worst friction, exploring the three factor groups, examining cases, and proposing strategies. The justification is that the literature is large and scattered. Teams need a joined-up view if they are going to change a real pipeline rather than a slide.

Quality appraisal in the thesis used five criteria: clarity of objectives, methodology, reproducibility, validity of results, and relevance. Paladino finds that many papers state clear aims and sit on the right topics, while reproducibility and empirical depth are weaker. Some studies describe tools without enough detail to repeat the work. Others generalize from thin samples. She also notes a concentration on DevOps, testing, and development, with less attention to planning and maintenance, even though her own matrix later flags those as critical. Generative tools such as ChatGPT, Codex, and GitHub Copilot appear often enough to count as a trend, not a novelty.

Which AI tools show up in each SDLC phase?

AI tools show up in every phase Paladino maps, but they do different jobs: estimate and schedule in planning, structure text in requirements, sketch architecture in design, generate and scan code in implementation, expand coverage in testing, stabilize releases in deployment, and watch debt in maintenance. The catalog is a map of reported practice, not an endorsement of every named product.

In planning she lists OCR for pulling text from images and PDFs, deep models for effort and size estimates, intelligent agents for task and resource assignment in DevOps pipelines, LSTM and RNN models for schedule prediction, and platforms such as Jira Automation and Monday.com for tracking. ChatGPT appears as a way to draft early documentation and shape requirements. In requirements work, NLP structures documents, ChatGPT interprets messy text, and machine learning looks at historical patterns to pin down needs.

Design tools include Sketch2Code for turning sketches into HTML, generative models for architectural patterns and UX mockups, simulators that stress adaptive designs, and environments such as Adobe XD and ArchiMate AI. Implementation is the familiar generative layer: GitHub Copilot, OpenAI Codex, and Facebook TransCoder for generation and language translation, plus machine learning that flags vulnerabilities as you type, and autocomplete assistants for local suggestions.

Testing, in her synthesis, uses deep learning and NLP to generate and classify scenarios, plus frameworks such as Testim, Test.AI, and Applitools. Predictive tools watch runs and try to anticipate critical faults. Deployment brings reinforcement learning for resource adjustment, CI/CD with Jenkins AI and Automation Anywhere, and monitoring with Dynatrace and New Relic. Maintenance uses predictive models of technical debt, automatic refactoring via semantic search, and continuous monitoring that can adjust at runtime. The list is long on purpose. It shows that “AI in the SDLC” is not one product. It is a pile of phase-specific bets, each with its own data hunger and failure mode.

Which life-cycle stages are the hardest to integrate?

The hardest stages, in Paladino’s matrix, are planning, testing and quality assurance, and deployment plus maintenance. Those are the places where missing data, thin traditional methods, and inherited systems pile up. Other phases still have friction. These three show up as the critical path for anyone who wants more than a coding assistant.

Planning fails when effort estimates rest on incomplete history and when tools do not share standards with the rest of the estate. Testing fails when coverage is still a manual suite with a generative wrapper, and when nobody trusts a case the model invented. Deployment and maintenance fail when new monitors cannot talk to old runtimes, and when defect prediction has no clean feedback loop. Paladino’s point is not that coding is easy. It is that the industry talks about Copilot while the worst bottlenecks sit earlier and later in the cycle.

What technical, organizational, and ethical barriers remain?

Technical, organizational, and ethical barriers remain because a model that writes code still needs clean data, a willing team, and a decision path people can inspect. Paladino groups the literature that way on purpose. A tool that solves only one group will stall on the other two. That is the thesis’s most useful structural claim.

Technical barriers start with data. Fragmented or low-quality records weaken estimates and defect models. Black-box behavior makes results hard to trust in analysis and validation. Compatibility with inherited architectures forces expensive rework. Over-reliance on generators can import vulnerabilities if humans stop reviewing. Paladino cites work that treats interoperable platforms and modular designs as the way through, and she is not naive about the gap between that advice and a fifteen-year-old estate.

Organizational barriers are about people and power. Resistance comes from fear of lost roles and from skills that suddenly look dated. Training is uneven. Management support is often a slogan without budget. Some organizations build mixed teams of AI specialists, developers, and project leads and move. Others stall because nobody owns the change. Paladino treats that split as a real controversy, not a footnote.

Ethical barriers close the triangle. Privacy and security of sensitive project and user data are recurring. Transparency is the other half: if developers cannot see why a model acted, they will not put it on a critical path, and they are right not to. Paladino notes that explainable-AI techniques are still early relative to the regulatory and trust bar. Fairness and bias sit in the same cluster. The literature she read is stronger on naming these problems than on offering a single operational fix. That gap is one of the empty spaces she flags for later research.

What strategies does the thesis put on the table?

The thesis puts a three-layer strategy on the table: improve the data and the architectures, train the people and back them with management, and wrap the whole effort in ethics and governance. Paladino’s closing move is that those layers have to travel together. A new Copilot license with no training plan and no audit trail is not integration. It is a purchase.

Technical strategies include data curation and preprocessing so predictive models are not trained on noise, explainable methods such as attention maps and rule overlays, standardized platforms that can sit next to inherited systems, and modular designs that can absorb AI in slices. Organizational strategies include continuous training, mixed teams, and a leadership story that treats AI as a funded change rather than a side experiment. Ethical strategies include privacy policies, periodic audits for bias and leakage, and mechanisms that make automated decisions inspectable for every stakeholder who has to live with them.

Cases in the review make the same three-layer point in smaller form. Intelligent agents in CI/CD can coordinate delivery and still trip on inherited systems. NLP can extract requirements from documents and user comments, and it still needs people to settle ambiguity. Deep learning can grow test suites, and it still needs a definition of pass. Predictive maintenance can cut downtime, and it still needs data that is structured enough to trust. Paladino’s examples are not trophies. They are illustrations of the same stall: the model helps a phase, then the estate, the culture, or the ethics gap stops the help from spreading.

You can also read the same research article as a snapshot of Spanish-language specialist training in software engineering in 2025. Paladino is synthesizing recent papers, including work she cites from Benitez and Serrano and from Müller, into a barrier-and-strategy frame for practitioners. The institutional record on SEDICI is the citable copy for the degree. If you take one idea from the review into an English-speaking team, take the three-layer insistence. Tools per phase are easy to list. Data quality, legacy compatibility, resistance, and transparency are what decide whether those tools change the cycle or just decorate it. Paladino also records the limit of her own method: a narrative review cannot claim completeness, and later work should pair it with a stricter systematic protocol. That honesty is part of why the thesis is useful. It maps a messy field without pretending the map is the territory.

Frequently Asked Questions

What kind of document is Paladino’s study?

It is a Spanish-language specialist thesis and narrative review, submitted in April 2025 for the degree of Especialista en Ingeniería de Software at the National University of La Plata’s Faculty of Informatics, directed by Claudia Pons. This article is an English explainer of that work, not a translation of the full thesis.

Which SDLC phases does she map AI tools to?

The review classifies tools for planning, requirements, analysis, design, implementation, testing, deployment, and maintenance. Examples include deep models and intelligent agents for planning, NLP for requirements, Copilot and Codex for coding, Testim and Applitools for testing, and predictive models for maintenance.

Which SDLC stages are most critical for AI integration?

Paladino flags planning, testing and quality assurance, and deployment plus maintenance as the most obstacle-heavy stages. Planning lacks clean data for estimates. Testing still depends on thin traditional suites. Deployment and maintenance collide with legacy systems and scale.

What barriers does the thesis group together?

Barriers fall into three groups: technical (data quality, legacy compatibility, black-box models), organizational (resistance, weak training, thin management support), and ethical (privacy, fairness, and lack of transparency). The recommended response is not a single tool. It is training plus ethics frameworks plus technical cleanup.

What strategies does Paladino recommend?

On the technical side: data curation, explainable models, interoperable platforms, and modular architectures. On the organizational side: continuous training, mixed teams, and leadership support. On the ethical side: privacy policies, audits, and mechanisms that make automated decisions inspectable.

Related articles

• Integrating AI into the software development life cycle
• Harnessing AI for advanced software engineering
• Ethical QA practices, bias, and compliance in testing