How AI Can Improve Software Development

Change is constant and inevitable. While some changes are welcome, others are just a necessary evil.

Continuous calls for digitalization combined with a pandemic-prompted emphasis on transforming business operations, force organizations across industries to modernize their legacy systems. The emergence of cloud computing has shrunk organizations’ on-premise IT footprints. However, with 96 of the world’s largest 100 banks, 90% of the world’s largest insurance companies, 23 of the 25 largest retailers in the U.S. and 71% of the Fortune 500 still using mainframe systems, any modernization efforts—resulting in increased risk of downtime or defects—will frequently lead to disastrous consequences.

So, while making changes can be expensive, uncertain, and lead to customer loss and/or embarrassment in the marketplace, modern business demands that systems change to continue to deliver value. The problem with mainframe modernization, however, is that today’s code search tools, linters and program analysis tools are deficient when it comes to mitigating the risks associated with maintaining and improving systems.

Modernization: A Knowledge Problem

With 10,000 mainframes in active use globally, the average organization annually spends anywhere from 60% to 80% of its IT budget just on maintenance. But, as experienced and proficient programmers retire or move on, many companies are realizing that the specialized domain knowledge of the industry and institution that those developers used to create and update these complex critical systems is walking out the door with them.

As the developers who originally wrote the code have long since moved on, and documentation leaves a lot to be desired, developers new to a system are unable to acquire comprehensive knowledge of the system’s functionality, at a detailed or fast enough level, to become adequately productive and proficient. That lack of speed comes with significant risk and cost to any organization.

Recognizing the full and true intent of the functionality written in code is not simple. Developers today spend roughly 75% of their time searching through source code, to identify the code representing the functionality that needs to be changed. Complicating matters further, understanding the code is not enough. To effectively and efficiently maintain and support any system, software developers must know precisely what an application actually does—and how altering code in one part of the system impacts it as a whole. But since the code representing a behavior that needs to be changed can be littered throughout the system, developers might think they’re making a simple tweak, when really they may be breaking the entire system (and not even knowing it). The risk of change resulting in unintended consequences is real.

All of that being said, how can developers improve their efficiency and effectiveness?

The Shortcomings of Today’s Tools

With code repositories growing unchecked over time to extraordinary sizes, those responsible for maintaining and preserving a system’s functionality say it’s becoming increasingly challenging to find bugs in code without machine assistance. And with debugging being an arduous, time-intensive endeavor, developers are increasingly turning to code search tools, linters, and static and dynamic analysis tools to analyze millions of lines of code, flag errors and suggest fixes.

Whether the tool excels at bug localization, code visualization or “error” detection, entirely too many tools are insufficient when it comes to actually identify the specific lines of code that require attention. Program analysis tools only illustrate code in ways that developers still have to interpret (potentially incorrectly) and draw their own conclusions. While code search tools might speed the rate at which developers can build a mental model of the code, such tools are notorious for false positives, and developers still have to undergo the mentally challenging endeavor of piecing that mental model together to identify the code, and then safely make changes. Worse yet, code completion tools—tools that less-experienced developers disproportionately rely on—can actually propose imprecise changes (one prominent tool offers incorrect code 71% of the time), and that introduces risk.

The cherry on top of this problematic sundae is that today’s tools don’t have any way to effectively validate the scope and accuracy of a proposed change. The influence of a suggested change can be tough to identify, so testing for comprehensive validation is frequently next to impossible.

Ideally, there has to be a better way to mitigate risk while making changes to legacy systems. Enter artificial intelligence. By employing a unique approach to AI, the code repository truly becomes a knowledge repository, enabling many capabilities unrealized before now.

A Developer’s AI-Enabled Accessible Knowledge Repository

By leveraging AI to automate the identification of the specific lines of code that require attention, developers can simply ask this AI-driven knowledge repository where behaviors are coming from—and quickly identify the code associated with that behavior. This puts AI squarely in the position of intelligence augmentation, which is key to leveraging its capabilities.

This novel approach of AI reinterprets what the computation represents and converts it into concepts, therefore “thinking” about the code in the same way humans do. The result is that software developers no longer have to unearth the intent of previous developers encoded in the software to find potential bugs. Even better, developers are able to overcome the inadequacies of automated testing by using AI to validate that they haven’t broken the system before they compile or check in the code. The AI will forward simulate the change and determine whether it’s isolated to the behavior under change. The result is the bounds of the change are confined to the behavior under change so that no unintended consequences arise.

While AI is not yet advanced enough to safely make changes on its own, it can serve as a knowledge repository with the added capability of directing developers precisely where to make any needed changes. With AI, programmers—even on their very first day on the job — can make any necessary fixes with confidence that they won’t break the entire system.

Unlike its code search, code navigation, code completion and program analysis tool counterparts, AI fundamentally empowers developers to safely, efficiently and effectively make any necessary changes. With AI — that is fortified by a knowledge repository that understands source code the same way a human does — developers can truly do more with less.