Here’s why half of developers will soon use AI-augmented software

Generative artificial intelligence (genAI) tools to assist in the creation, testing and operation of software are expected to be adopted by half of all enterprise software engineers by 2027, according to a new study by Gartner Research.

Today, only about 5% of enterprise software engineers use genAI tools to assist in coding. That number is likely to quickly grow because software demand exceeds most organizations’ capacity, existing developers are maxed out, they’re unable to build features fast enough, and they’re less satisfied in their work, the study found.

AI-based code generation products based on large language models (LLMs) such as GitHub Copilot, Replit GhostWriter and Amazon CodeWhisperer, can generate complex suggestions resulting in a significant increase in developer productivity. But, those tools in no way eliminate the need for human software developers and engineers as genAI can still produce errors and is incapable of creating unique code.

That said, within two years, 80% of software engineering organizations are expected to establish platform teams as internal providers of reusable services, components, and tools for application delivery.

According to research firm IDC, the enterprise experimentation with genAI for code creation is second only to its use for text generation.

IDC

Cisco CIO Fletcher Previn has said that one of the places he never expected AI to touch was software development, which he equates to an art form requiring unique creative abilities. ChatGPT, however, has been adept at creating code that addresses corporate data hygiene and security and it can reuse code to build new apps.

A 2022 study by Microsoft showed more than half of all code being checked into GitHub was aided by AI in its development. That number is expected to jump to 80% of all code checked into GitHub within the next five years, according to GitHub CEO Thomas Dohmke.

“…Historically there was no way to compress software development timelines,” Previn said in an earlier interview. “Now, it turns out you can get a significant acceleration in velocity by helping developers with things like Copilot for code readings, code hygiene, security, commenting; it’s really good at those things.”

AI coding assistants are emerging as accelerators, boosting developer productivity and happiness, according to a number of studies. By handling routine tasks, genAI assistants enable developers to focus on higher-value activities, which allows organizations to deliver more features faster with existing teams, according to Gartner.

The AI assistants also boost the ability of “citizen” developers to quickly create apps to meet ever-changing business needs.

Being able to use AI is currently seen as the most important technical skill, according to IDC.

IDC

Gartner Senior Principal Analyst Philip Walsh said there are three software development areas where Gartner is seeing the impact of generative AI tools:

• AI coding assistance.
• AI-augmented testing tools.
• Design-to-development tools.

AI coding assistance tools function as a plug-in to a developer’s integrated development environment and include capabilities such as code completion or suggesting snippets of code to complete what’s already been written.

Developers are also using AI-based coding assistants to help them generate unit tests and software documentation. The tools can also be used to highlight a portion of code and then, using a natural chat interface, developers can ask questions to better understand and explain the functionality of what they’re looking at.

“We know developers are often working on improving or updating code that they didn’t write,” Walsh said. “Or maybe the person who developed that code no longer works for the company. Or it’s a legacy application that not a lot of people have touched in a long time or understand.”

Natural language processing embedded in AI-augmented software development allows humans to talk to the underlying LLMs and try out ideas, brainstorm their approaches to coding, and get reminders about a framework that, for example, hasn’t been used recently. While it’s a value that’s difficult to quantify, from a qualitative metric, natural language processing bolsters the developer user experience.

According to Forrester Research, enterprise AI initiatives are expected to boost productivity and creative problem-solving by 50% in the next few years. “Building on multiple investments over the past decade, generative AI is poised to increase productivity across IT operations. Current projects already cite improvements of up to 40% in software development tasks,” Forrester said in a recent report.

Last year, GitHub published a study showing 88% of developers using its Copilot tool felt more productive, were faster in completing tasks, and spent less time searching the internet (77%) for answers.

“They’ll feel more productive. They’ll see they’re not content switching as much or looking things up on Stack Overflow or Google as much,” Walsh said. “Developer sentiment is relatively high among that suite of capabilities that AI coding assistants bring.”

On average, within the first year in the market, users accept nearly 30% of code suggestions from GitHub Copilot. Over time, the acceptance rate steadily increased as developers became more familiar with the tool.

“That’s an indication that a developer gets used to prompting the tool and gets used to using the tool more efficiently,” Walsh said. “The flipside of that is 60%, 70%, or 80% of suggestions are not being used. So, having a human in the loop is still absolutely essential here.”

While genAI-assisted testing tools, designed to improve an organization’s ability to create test data and help create API tests and regression tests, are not new; genAI is simply adding capabilities to existing products.

Finally, AI-augmented design-to-development tools such as Figma help developers translate designs into code faster and create front-end presentation layers for applications.

But problems with genAI persist across the many places where it’s been deployed. For example, coding errors, hallucinations, and security holes remain ongoing concerns for  organizations eyeing the adoption of such tools.

“We advise all our clients hallucinations are very much real with these things, but our advice hasn’t changed in terms of how to mitigate that risk,” Walsh said.  “You should already have various quality and security scanning tools as part of your overall DevOps workflow, and you should have robust code review practices where a senior engineer reviews anything before it’s merged.”

How good AI-augmentation tools are varies, depending on the complexity and proprietariness of the code. If it’s a boiler-plate task, such as writing code for an HTTP server using JavaScript, acceptance rates tend to be high; that’s because the data used to train the underlying LLM is widely used and available.

Enterprise engineers, however, have found when they’re developing more complex code that relies on proprietary business logic not well represented in publicly available training data, the time savings are not as significant — and the accuracy and performance of the model is not as good, Walsh said.

Even so, in the near to medium term, genAI-enabled software creation tools will increase in accuracy and capabilities, including enabling business users to develop disposable apps for, say, data analysis where enterprise-grade quality isn’t necessarily needed.

“Those cases will be more of a productivity tool to help them with their work,” Walsh said. “That will be like the no-code market today. I do see use cases like that on the horizon. That’s much more closer to becoming a reality than fully automated enterprise grade software created by AI.”