The use of Gen AI in software development is already delivering impressive results, especially in simpler scenarios. But, as the technology advances, it brings new challenges, particularly in more complex environments like those found in large organizations.
In this article, we’ll take a closer look at the role of Gen AI in software development, exploring its impact, how to overcome key challenges, and which strategies can help mitigate risks while maximizing the value of this powerful technological accelerator.
Note: This article draws from structured research, empirical observations, and the author’s extensive expertise.
The Impact of Gen AI on Software Development
The earliest applications of Gen AI in software development involved code assistants. These were tools that answered questions, suggested code snippets or files, and supported autocompletion functions.
These assistants quickly evolved to allow file creation and editing, leading to more advanced AI agents capable of planning work, interfacing with external tools, and independently creating, modifying, and deleting code and files.
As time went on, these agents gained even more capabilities, to the point where they could produce entire websites or applications with no human input.
According to the 2024 DORA Report, most teams are already investing in Gen AI: 89% use it to enhance their products, while 76% rely on it as an everyday productivity tool.
We’ve moved beyond the hype. Companies now seek measurable outcomes from their AI efforts, and software development is no exception. The 2025 State of the CIO survey shows that, for 68% of IT leaders, AI has already reshaped operations and is delivering tangible business value.
Gen AI is helping developers boost productivity, generating quality documentation, writing tests, and speeding up code creation and review.
For instance, the 2024 Stack Overflow Developer Survey featured an entire chapter on Artificial Intelligence. More than 35,000 development professionals worldwide shared how AI has become part of their development workflow:
Still, there are lingering concerns: intellectual property issues, performance and quality in production, and even fears that AI might take away the “fun” parts of development or threaten jobs.
Key Use Cases for Gen AI in Development
Gen AI unlocks powerful accelerators that, when applied well, can boost productivity and code quality. Here are some of the most impactful use cases:
1. Write Documentation
Development teams often lack the time or discipline to keep documentation up to date. Gen AI can generate materials like Mermaid diagrams, Javadoc or OpenAPI documentation, architecture decision records (ADRs), and even C4 architecture diagrams.
2. Write Unit and Integration Tests
Similar to documentation, tests are often neglected. AI agents can create and maintain a wide range of tests, from basic unit tests to mutation and even Test-Driven Development (TDD) or Behavior-Driven Development (BDD) approaches.
Itaú Unibanco, for example, doubled its accessibility test coverage using StackSpot AI.
3. Create Simpler Code
Tasks like writing getters and setters, build scripts, or CRUDs can consume valuable developer time. Gen AI can handle these with quality and speed.
Among other things, you could ask your AI agent to implement CRUD for a new entity using a simple set of business rules.
4. Review Code Changes
According to Harvard Business Review, improving code post-creation ranks as the 8th most frequent Gen AI use case in 2025. AI agents can assist in code reviews, reducing reliance on senior staff and speeding up PR approvals.
Understanding the Risks
Using Gen AI without the right context or at an unsustainable pace can easily backfire. Instead of boosting efficiency, it might actually slow things down, especially when the generated code can’t be used in production. To avoid this, it’s important to understand the key risks that come with relying on general-purpose AI for software development:
1. Failure to Follow Standards and Policies
AI-generated code may not comply with internal policies or development guidelines.
2. Security Flaws
It can introduce vulnerabilities, from simple missteps like storing credentials in files to serious threats like exposing services on the internet.
3. Inefficient Resource Use
The LLM-generated code might consume unnecessary compute, storage, or bandwidth, driving up operational costs, especially in cloud environments.
4. Fragile Solutions
AI might apply short-term fixes that break existing features, introduce bugs, or even blend incompatible architectures and coding paradigms.
5. Oversimplified Code
In complex scenarios, the LLM may fail to meet requirements, miss edge cases, or even hallucinate in its responses.
6. Mediocre Results
Because LLMs are trained on public code, often from personal or low-quality projects, they may suggest solutions that are poorly written, inefficient, or insecure.
7. Overconfident Algorithms
Some models are optimized to always produce an answer, even when the right move is to pause and rethink. This behavior can clash with the experience and judgment of seasoned professionals.
Recommendations for Successful Adoption
1. Use a Development Platform
Large organizations often struggle with governance due to dispersed teams and standards. Engineering and development platforms solve this by centralizing best practices and policies, ensuring visibility and simplifying the process for development squads.
It’s also a growing trend. According to Gartner, by 2026, 80% of companies will have adopted platform engineering strategies. Integrating Gen AI into your development platform ensures that responses are not only more accurate, but also reflective of your internal standards, offering a smoother, more unified experience.
2. Educate Your Teams
AI code agents can dramatically shift a team’s dynamics. They can easily take over most of a squad’s coding duties, for example. While some teams will embrace that, others tend to resist.
These agents can function like junior or mid-level developers with three or four times the output, as if you had more “hands on deck” for your projects. Leaders must learn how to integrate them effectively into team workflows.
You should start with training and awareness. Instead of forcing adoption, track usage and results over time. Invest in AI Literacy and upskilling; after all, teams won’t adopt what they don’t understand.
3. Prioritize Contextualization
The more context AI has, the better the output. This reduces inefficiencies and ensures alignment with internal standards.
A development platform is ideal for providing centralized, up-to-date information. Techniques like RAG and Fine-Tuning take this even further.
4. Apply RAG or Fine-Tuning
To enhance the quality and relevance of your AI-generated code, it’s worth exploring techniques like Retrieval Augmented Generation (RAG) and Fine-Tuning.
RAG does not modify the model’s internal parameters (weights). Instead, it enriches the prompt with more specific and up-to-date information pulled from relevant sources. This helps the LLM generate responses that are better aligned with the project’s actual context.
Advantages of RAG:
a. It’s faster and more cost-effective since there’s no need to retrain models—just plug in the right tools.
b. It allows access to recent data, which can be retrieved and vectorized on demand.
c. It’s more transparent, making it easier to trace and debug the data used to build the final prompt.
Challenges of RAG:
a. The output depends heavily on the quality of the retrieval and vectorization (embedding) strategy. A weak setup will directly impact the quality of code generation.
b. There are limitations around prompt size. When combining the original input with RAG-enriched context, the total length may exceed the model’s context window, which can lead to truncated or incomplete results.
Fine-Tuning, on the other hand, directly updates the model’s weights by training it on a curated, domain-specific dataset. This results in a model that’s more specialized for your particular use case.
Advantages of Fine-Tuning:
a. It enables domain-specific expertise, resulting in more relevant and less generic code.
b. It produces faster responses, since the model already “knows” the domain and doesn’t require augmented prompts.
Challenges of Fine-Tuning:
a. It involves higher costs, as it requires high-quality training data and significant computational resources.
b. Updating the data isn’t simple. You’ll need to retrain the model even if only a small portion of the dataset has changed.
5. Develop Prompt Engineering Skills
LLMs are the core of generative AI, and knowing how to “talk” to these models is one of the most effective ways to improve results. It’s also fast and affordable. Better prompts lead to better answers, including more accurate and useful code.
The ability to craft clear and effective instructions is known as prompt engineering. It is a powerful and cost-effective way to boost the quality of what your AI delivers. On our blog, you’ll find recommendations and best practices for getting started on Prompt Engineering.
Use well-known and proven techniques to write clear, production-ready prompts. They can save both time and resources by reducing the number of tokens used during generation. The difference between a bad prompt and a good one might be the difference between an application that’s ready for production and one that doesn’t even compile.
6. Monitor AI-Generated Code
Last but not least, it is important to track what agents produce and how they perform. Identifying this code is straightforward, since agents can commit to repositories using designated user accounts. Reviewing their performance is also simple, thanks to Git metrics such as the number of lines changed, commits created, and PRs approved.
Separating code written by agents from code written by people helps prevent misunderstandings around intellectual property and individual contributions. It also allows teams to measure their performance more clearly, both with and without AI in the loop.
To ensure this clarity, agents should always use specific Git users, and teams should consistently track the following metrics:
- Number of commits.
- Number of lines created, changed, and deleted.
- Number of approved PRs.
- Number of bugs introduced, unresolved security vulnerabilities, and overall test quality.
At the end of the day, it should always be clear which parts of the codebase were created by humans, which by agents, and how the team’s performance has evolved with the help of generative AI.
Gen AI: The New Imperative for Development Teams
The impact of Gen AI on software development is no longer a future trend, it’s already shaping how great software gets built. Teams that embrace these accelerators are gaining speed, quality, and scale, while those that don’t risk falling behind.
The rise of AI-powered agents, whether assistive or fully autonomous, is redefining what development teams look like and how they operate. The question is no longer whether to explore Gen AI, but how to make it work safely and effectively in your context.
Now’s the time to shift from merely “evaluating” use cases to actively reducing risks and unlocking real results.
If you haven’t started yet, hurry while there’s still time. Take control of your AI journey and lead the transformation!
References
- Impact of Generative AI in Software Development | DORA
- Accelerate State of DevOps Report 2024 | DORA
- Is your platform a Fruit Basket or Fruit Salad? – The Architect Elevator
- What is an Internal Developer Platform (IDP)?
- What is platform engineering?
- 2024 Stack Overflow Developer Survey
- How People Are Really Using Gen AI in 2025
- State of the CIO 2025
- Artificial Intelligence: A Modern Approach – Fourth edition – UC Berkeley
- Sequence to Sequence Learning with Neural Networks – Cornell University
- Tree of Thoughts: Deliberate Problem Solving with Large Language Models – Cornell University
