By Ram Ramachandran and Suhani Varadharaj

Generative AI is shaking up the software development sector, prompting engineering teams across industries to reimagine how they work and what they do.

Research by MIT reveals that 94% of business leaders use generative AI in some capacity for application development.¹ More than a third of respondents (38%) anticipate generative AI will “substantially” alter the software development lifecycle in one to three years, and another 31% expect these changes to happen in four to 10 years.

Traditional software development roles are also evolving. Once primarily focused on knowing the algorithms and syntax of programming languages, developers can now use generative AI tools embedded in integrated development environments (IDEs) to understand context, generate blocks of code or functions, and test, document and correct code.

The implications for software development as both a practice and a profession are clear: responsibly adopt and adapt generative AI or risk being left behind. These five critical lessons, learned while implementing GitHub Copilot at Kyndryl, will help guide your way:

Lesson 1
Set clear goals to manage expectations and align priorities

Integrating generative AI into software development requires substantial time and money, so you’ll want to determine what outcomes you plan to achieve before investing.

At Kyndryl, we identified three primary goals when deciding whether to integrate generative AI into our software development:

• Increase the speed of software development
• Enhance planning and resource allocation through better estimations
• Reduce programming errors and improve code quality

These priorities aligned with our company’s overarching strategy of fostering innovation while exploring ways to enhance customer service and operate more efficiently.

Lesson 2
Choose an adaptable generative AI tool to meet current and future requirements

When evaluating generative AI solutions, look for tools that are feature-rich, easy to use and compatible with existing systems. If your company has relationships with cloud and AI providers, try to arrange low- or no-fee trials to test performance in real-world scenarios.

For example, when narrowing our search for a generative AI coding solution, we conducted a pilot rollout to 200 Kyndryl developers. Participants used generative AI to develop code in Python, Javascript and React JS, paying particular attention to the accuracy, relevancy and coherence of AI-generated content to ensure it met our quality standards.

Lesson 3
Refine training and implementation approaches to meet changing needs

Generative AI is of little value if it isn’t deployed properly, so it’s critical to equip software developers with the knowledge and skills they need to use the technology proficiently.

During Kyndryl’s implementation, we hosted two introductory boot camps, a session on prompt engineering, an advanced boot camp, and two garage sessions over two months. We also tapped 100 champions to train colleagues and organized weekly sessions where developers could discuss their experiences with the tool and share best practices.

Using continuous feedback, we refined the curriculum to ensure the training met developers' evolving needs. For instance, to address user experience issues encountered during our initial trial, we expanded the pilot and divided training into two groups: one for hands-on, fundamental education and another for advanced training.

Lesson 4
Develop a robust governance structure to reduce risk

As more software development teams implement generative AI, organizations will be held accountable for using the technology ethically. However, a recent Kyndryl survey found that only 17% of surveyed companies have documented a position on responsible AI.

When structuring a governance framework for our pilot program, we established a human first, human last mandate. This approach reinforced the fact that generative AI is intended to augment the work of software developers, not replace them or their expertise.

To support compliance, we assembled a governance body that performed regular audits and compliance checks and provided continuous monitoring and support to address technical issues during implementation.

Lesson 5
Track and manage technology usage to maximize productivity gains

Since enhanced productivity is a key benefit of deploying generative AI, it’s a good idea to consistently measure how technology usage correlates with work output and quality.

With our pilot program, we used agile methodologies to track improvements in developer velocity and overall code quality. For example, if an developer using generative AI shortened the time to complete a task from five story points to three story points, the reduction indicated faster coding and greater productivity.

We also monitored the number of active generative AI users to ensure developers were using the technology for innovation rather than simply treating it as an autocomplete tool. If a pilot participant went 30 days without using the tool, we reclaimed the license and assigned it to another engineer.

Impactful use cases of generative AI in software development

Although Kyndryl is still in the early stages of deploying generative AI for software development, we’ve already identified three impactful use cases:

• Code development. Generative AI tools like GitHub Copilot, Google Bard and Amazon Q Developer can accelerate software development by autonomously producing code snippets, suggesting completions and, in some cases, writing entire functions for tasks like data manipulation, analysis and visualization. In fact, Google currently uses generative AI to write more than 25% of new code for its products.²
  
• Automated testing and debugging. By analyzing code patterns, logs and system behavior, generative AI can help developers pinpoint the root cause of issues faster and more reliably than manual troubleshooting. Automated testing and debugging also allow software teams to predict and address potential problems before they cause production outages and other costly setbacks.³ 
  
• Documentation and requirements creation. Generative AI uses natural language processing (NLP) to convert requirements and documentation into detailed technical specifications that developers reference when creating or modifying code. Documentation and specifications created directly from the codebase reduce the chance of errors and help maintain code quality and consistency. 

In time, generative AI may be used to automatically refactor and optimize existing codebases, create behavior- and preference-based user interfaces, and support other software-enhancing innovations. Generative AI is a powerful tool that’s revolutionizing software development. Engineering teams that embrace this technology and borrow from Kyndryl’s lessons learned can streamline software development lifecycles while improving the overall quality and performance of programs and applications.

Ram Ramachandran is Head of Software Engineering and Suhani Varadharaj is a Senior Skills Lead in the CTO office of Kyndryl.

¹ Transforming software with generative AI, MIT Technology Review, October 2024
² Google generates 25% of new code using AI, says CEO Sundar Pichai, Outlook Business, October 2024
³ How generative AI is revolutionizing debugging, The New Stack, September 2024